Gradient Based Porosity

Calculation in Casting Simulation

Treball realitzat per:

Daniel Díez Rodríguez

Dirigit per:

Dr. Riccardo Rossi and Jordi Rubio

Màster en:

Métodos Numéricos en la Ingeniería

Barcelona, 15/06/2017

Departamento de Ingeniería Civil y Ambiental

 UNIVERSITAT POLITÈCNICA DE CATALUNYA

Escola Tècnica Superior d’Enginyers de Camins Canals i Ports de Barcelona

MASTER THESIS:

GRADIENT BASED POROSITY

CALCULATION IN CASTING SIMULATION

A Master Thesis submitted by Daniel Dı́ez for the Master in

Numerical Methods in Engineering

Supervised by:
Dr. Riccardo Rossi and Jordi Rubio
 Abstract

Porosity is one of the most common defects in casting production. When

some parts of the casting begin to solidify they contract due to the difference
of densities between the liquid and solid phases. The areas that are still liquid
feed the solidifying regions until at some point feeding is not longer possible
due to the growth of the solid phase and porosity begins to form. If the feeding
cut is due to the growth of dendrites, pores will have a very small size and
will be usually mentioned as microporosity. When a pool of liquid is isolated,
surrounded by solid material, a big void will appear known as macroporosity. If
the liquid pool is connected to the exterior it will form pipe shrinkage.
The aim of this work is the implementation of a microporosity model inside
the commercial casting software Click2Cast, based on the general finite elements
framework Kratos Multiphysics. The model that has been implemented is based
on the Dimensionless Niyama method developed by Carlson and Beckermann
in 2008. This approach requires the recover of the gradient of the temperature.
In order to do this, several methods to compute the gradients were developed
and implemented inside the code.
Finally, another model to calculate macroporosity and pipe shrinkage is being
currently implemented. This method is based on the calculation of the pressure
during the solidification. No results are provided for this model as still further
work is needed to expect a correct under different circumstaces .The objective
is that after validation, this model will be also included in Click2Cast.

2
 CONTENTS

Contents

1 Introduction 14

1.1 Casting Defects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.3 Structure of the work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Porosity Literature Review 18

2.1 Classification and Formation of Porosity . . . . . . . . . . . . . . . . . 18

2.2 Shrinkage Porosity Simulation . . . . . . . . . . . . . . . . . . . . . . . 21

2.3 Microporosity Models : Criterion Functions . . . . . . . . . . . . . . . 22

2.4 Macroporosity Models : Thermal / Volume Calculation Models . . . . 24

2.5 Coupled Microporosity and Macroporosity Models : Thermal/ Fluid

Flow Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 Gradient Reconstruction 30

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.2 L2 Projection Method . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.2.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.3 Newton-Raphson iterative scheme . . . . . . . . . . . . . . . . . . . . . 37

3.3.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.4 Zhang Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.4.1 Description of the method . . . . . . . . . . . . . . . . . . . . . 42

3.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4
 CONTENTS

3.5 Pouliot Gradient Recover Method . . . . . . . . . . . . . . . . . . . . . 46

3.5.1 Description of the Method in 1-D . . . . . . . . . . . . . . . . . 47

3.5.2 The Multidimensional Case . . . . . . . . . . . . . . . . . . . . 49

3.5.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.6 Comparison of results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4 Microporosity Model 58

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.2 Dimensionless Niyama Criterion . . . . . . . . . . . . . . . . . . . . . . 61

4.3 Study Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4.3.1 Lever . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.3.2 Brake Caliper . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.3.3 Oil Pump . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

4.3.4 Comments on the results . . . . . . . . . . . . . . . . . . . . . . 72

4.4 Gradient Recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.5 Sensitivity to the element size . . . . . . . . . . . . . . . . . . . . . . . 74

4.6 Sensitivity to different evaluation temperatures . . . . . . . . . . . . . 76

4.7 Casting parameters that have influence on microporosity . . . . . . . . 77

4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

5 Macroporosity and Pipe Shrinkage Model 84

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.2 Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5
 CONTENTS

5.3 Shrinkage Porosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5.4 Open Shrinkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5.5 Weak Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5.6 Resolution Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

6 Conclusions and Future Work 94

6
 LIST OF FIGURES

List of Figures

1.1 Gas Porosity [10] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.2 Shrinkage Porosity [10] . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.1 Types of Porosity [30] . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2 Microporosity in ductile iron [20] . . . . . . . . . . . . . . . . . . . . . 20

3.1 L2 Projection - L2 Error for Case 1 . . . . . . . . . . . . . . . . . . . . 35

3.2 L2 Projection - L2 Error for Case 2 . . . . . . . . . . . . . . . . . . . . 36

3.3 L2 Projection - Computational Time for Case 1 . . . . . . . . . . . . . 36

3.4 L2 Projection - Computational Time for Case 2 . . . . . . . . . . . . . 37

3.5 L2 Iterative - L2 Error for Case 1 . . . . . . . . . . . . . . . . . . . . . 39

3.6 L2 Iterative - L2 Error for Case 2 . . . . . . . . . . . . . . . . . . . . . 40

3.7 L2 Iterative - Computational Time for Case 1 . . . . . . . . . . . . . . 40

3.8 L2 Iterative - Computational Time for Case 2 . . . . . . . . . . . . . . 41

3.9 Zhang - L2 Error for Case 1 . . . . . . . . . . . . . . . . . . . . . . . . 44

3.10 Zhang - L2 Error for Case 2 . . . . . . . . . . . . . . . . . . . . . . . . 45

3.11 Zhang - Computational Time for Case 1 . . . . . . . . . . . . . . . . . 45

3.12 Zhang - Computational Time for Case 2 . . . . . . . . . . . . . . . . . 46

3.13 Element Edge [28] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.14 Pouliot - L2 Error for Case 1 . . . . . . . . . . . . . . . . . . . . . . . . 50

3.15 Pouliot - L2 Error for Case 2 . . . . . . . . . . . . . . . . . . . . . . . . 51

3.16 Pouliot - Computational Time for Case 1 . . . . . . . . . . . . . . . . . 51

3.17 Pouliot - Computational Time for Case 2 . . . . . . . . . . . . . . . . . 52

8
 LIST OF FIGURES

3.18 Comparison - L2 Error for Case 1 . . . . . . . . . . . . . . . . . . . . . 53

3.19 Comparison - L2 Error for Case 2 . . . . . . . . . . . . . . . . . . . . . 53

3.20 Comparison - Computational Time for Case 1 . . . . . . . . . . . . . . 54

3.21 Comparison - Computational Time for Case 2 . . . . . . . . . . . . . . 54

4.1 Schematic of the end-risered plate [5] . . . . . . . . . . . . . . . . . . . 59

4.2 (d) Averaged through thickness porosity. (e) Predicted Niyama crite-
rion distribution at the mid-plane from “Prediction of Shrinkage Pore
Volume Fraction Using a Dimensionless Niyama Criterion [5] . . . . . 59

4.3 Mushy Zone 1D Analysis from [5] . . . . . . . . . . . . . . . . . . . . . 64

4.4 Lever Part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.5 Lever - Regions over 2% of Microporosity . . . . . . . . . . . . . . . . . 68

4.6 Levers - Midplane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.7 Brake Caliper Part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

4.8 Brake Caliper - Regions with over 0.9% Microporosity . . . . . . . . . . 69

4.9 Brake Caliper - Section Cut . . . . . . . . . . . . . . . . . . . . . . . . 70

4.10 Oil Pump Part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.11 Oil Pump - Regions with over 0.5% Microporosity . . . . . . . . . . . . 71

4.12 Oil Pump - Brake Caliper Section . . . . . . . . . . . . . . . . . . . . . 72

4.13 Relative error for Ny* and Microporosity . . . . . . . . . . . . . . . . 73

4.14 Microporosity Calculation - Total Gradient Recover Computational Time 74

4.15 (a) Medium section with element size h = 0.5 (b) Medium Section with
element size h = 0.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.16 Evolution of the temperature gradient - Red: Exterior node - Blue:

Interior Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

9
 LIST OF FIGURES

4.17 Microporosity - HTC Influence . . . . . . . . . . . . . . . . . . . . . . . 78

4.18 Microporosity - Packing Pressure Influence . . . . . . . . . . . . . . . . 79

4.19 Microporosity - Feeding Temperature Influence . . . . . . . . . . . . . . 79

4.20 Microporosity - Mold Temperature Influence . . . . . . . . . . . . . . . 80

10
 LIST OF TABLES

List of Tables

3.1 Meshes used for testing gradient recover methods . . . . . . . . . . . . 31

3.2 L2 Projection - Convergence Rates . . . . . . . . . . . . . . . . . . . . 37

3.3 L2 Iterative - Convergence Rates . . . . . . . . . . . . . . . . . . . . . 41

3.4 Zhang - Convergence Rates . . . . . . . . . . . . . . . . . . . . . . . . 46

3.5 Pouliot - Convergence Rates . . . . . . . . . . . . . . . . . . . . . . . . 52

3.6 Convergence Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.1 Meshes used to study sensibility to element size . . . . . . . . . . . . . 75

4.2 Influence of mesh size to N y ∗ and microporosity . . . . . . . . . . . . . 75

4.3 Influence of evaluation temperature to N y ∗ and microporosity . . . . . 76

12
 1 INTRODUCTION

1 Introduction

1.1 Casting Defects

Metal casting is one of the oldest manufacturing processes where liquid metal is poured
into a mold which has the shape of the desired final part that will be obtained once
the liquid metal solidifies. The main advantage of this processes with respect to other
manufacturing processes is that very complex geometries can be easily produced. Only
limitation in this aspect is that we have to be able to open the mold and get the final
casting. Besides the mold itself, other components can be used in order the obtain
the final shape such as the cores. Using casting, a part made of practically any
metal and with any size can be produced with relatively no effort compared to others
manufacturing processes.

For these reasons, casting processes have a huge influence in the manufacturing
market and its volume of production has been regularly increasing. However, the
internal processes that occur inside the casting are very complex and the appearance
of defects is very hard to eliminate. According to American Iron and Steel Institute,
it is estimated that between 3% and 7% of the total produced castings have to be
scraped due to the presence of defects. These defects can be classified as follows:

• Filling related defect: Filling defects include misruns, cold shuts and inclusions.
A misrun occurs when a front of liquid metal solidifies before the casting is
completely filled leaving an unfilled portion. Cold shuts are a result of two fronts
of liquid metal which do not fuse properly inside the cavity of the mold due to
their differences, specially in temperature, that will lead to weak areas where the
material is not regular and cracks can form. Both misruns and cold shuts are
caused by a lack of fluidity in the molten metal. An inclusion is any impurity in
the pour metal like oxides which will result in worse material properties of the
final casting.

• Gas Porosity: Gas porosity is the formation of bubbles within the casting after
it has cooled. Liquid metals usually carry dissolved gas in the melt, but when
the metal starts solidifying the amount of gas that the solid phase can contain
is much lower than the amount of gas that was present in the liquid. The
consequence of this is that gas will try to create its own phase nucleating and

14
 1 INTRODUCTION

forming gas porosity (see figure 1.1). Besides, gasses can be present inside the
mold if as the liquid flow fills the mold cavity, the air that was inside the mold
has not enough time to exit ending up as entrapped air.

• Shrinkage Defect: Shrinkage defects due to the differences of densities between

the liquid and solid phases. As the metal solidifies it might happen that there is
no available feed metal to compensate for the shrinkage and voids will appear.
These empty areas are classified as pipes or caved surfaces if they are present on
the surface of the casting or shrinkage porosity if it appears inside the casting.

• Metallurgical defects: When the solid metal is still hot, its mechanical properties
are reduced and residual stresses due to the solidification might end up cracking
the material and forming the so-called hot tears.

Figure 1.1: Gas Porosity [10]

In order to eliminate these defects and achieve optimum operating conditions in-
dustrial experiments are usually conducted. This kind of experiments are complicated
due to the nature of the problem and can be very expensive. Though there is steel a
need of these experimental results, the casting producers understand more and more
that they can improve the product quality through computer simulation. This ap-
proach facilitates the optimization of the process by increasing energy efficiency and
avoiding defects with less cost.

15
 1 INTRODUCTION

In this work we will focus on the prediction of porosity as one of the most persistent
and common complains of casting users.

Figure 1.2: Shrinkage Porosity [10]

1.2 Objectives

The main objective of this work is to implement a gradient based model to compute
microporosity in the finite elements commercial code Click2Cast which will be included
in next versions of the code. The code has been implemented using the multidisci-
plinary framework Kratos Multiphysics. As the main key of this calculation is the
gradient of the temperature, the implementation and comparison of different gradient
recover methods has been also part of the work. At the current time, another model
to compute macroporosity is being implemented with the objective of being included
as well in future versions of Click2Cast.

1.3 Structure of the work

In section 2 a review of the available literature about porosity is exposed, explaining

the different approaches to the problem. One of the models for microporosity presented
in this section has been implemented in Click2Cast and explained in detail in section
4. For the implementation of this method, it is previously needed to compute the
gradient of the temperature. Different ways to compute it are shown and compared
in section 3 . Finally, in section 5, it is presented the current and future work in a
macroporosity and pipe shrinkage model which is being tested and implemented in
these moments.

16
 1 INTRODUCTION

17
 2 POROSITY LITERATURE REVIEW

2 Porosity Literature Review

2.1 Classification and Formation of Porosity

Porosity is one of the major defects in metal castings. Its presence results in a dete-
rioration of the mechanical properties of the final part. Particularly tensile strength
and fatigue resistance decrease significantly when porosity forms, resulting in a failure
in the performance of the material for lower loads or in a lower number of cycles that
we could expect.

Despite all the efforts and studies that have been carried out in order to fully
understand how porosity forms and grows, there is no clear agreement on which are
the mechanisms that rule this process. One of the biggest difficulties in this problem
is the considerable number of variables that play a role in the solidification. We can
classify them in one of the following groups:

• Material Properties

• Composition of the Alloy

• Casting Conditions

• Mold Properties

• Geometry of the part

All these variables will somehow affect the solution and all of them will have to be
taken into account. However, it is very hard to know the relative effect that each of
them will have compared to the rest. For this reason, despite casting has been part
of the human activities for thousands of years, its most complex aspects are not yet
well known and it is almost impossible to consider all of them in one single model.

These difficulties should not discourage us to investigate in this field. More the op-
posite, this means that this problem requires more investigation and efforts as porosity
is known to be one of the most important aspects in casting simulation and a major
cause of casting rejections.

Porosity can be classified according to different criteria, usually origin and size.
The soundness of a casting depends on the flow of liquid metal to the solidifying regions

18
 2 POROSITY LITERATURE REVIEW

Figure 2.1: Types of Porosity [30]

that will contract as they become solid. The primary cause of porosity is therefore the
difference in the densities of liquid and solid phases and the obstruction of the fluid
flow as the solidification advances. When this feeding is cut from a region, shrinkage
porosity will form. Besides the contraction of the metal there are other factors that
can affect to the amount of porosity present. The presence of gas dissolved in the
metal will help the pore nucleation to begin and grow. The solubility of gas in the
solid metal is lower than the solubility in the melt and therefore, as the solidification
advances more gas will be present in the liquid favoring the nucleation of the pores.
By the size of the pores we can distinguish between microporosity or macroporosity.

Macroporosity

As the solidification advances from the exterior to the center of the filling cavity it
might happen that large pools of liquid end up surrounded by solid metal. As there
is no possibility of feeding through the solid walls, when the whole pool becomes solid
there will be a big region of void in the center called macroporosity

19
 2 POROSITY LITERATURE REVIEW

Figure 2.2: Microporosity in ductile iron [20]

Microporosity

On the other hand, microporosity forms in regions where there is a poor feeding
due to the growth of dendrites. If the gas forms in the mushy zone after the dendrite
coherency point has been reached, it will be entrapped in the dendrite network, in
the so called interdendritic spaces and will nucleate forming small pores called micro-
porosity or microshrinkage. The size of this pores ranges between 1-100 µm . In this
case, these pores will grow following the dendritic shape instead of being round.

Open Shrinkage

All the previous forms of porosity are classified as closed shrinkage defect. However,
if the casting part is open somewhere to the exterior it will form open shrinkage defects
or pipe shrinkage. In these regions the free surface is capable of moving, lowering its
level as the rest of the metal solidifies. As long as the solid fraction is smaller than
the coherency point it may be assumed that the solid crystal move along with the
liquid. In this situation the air will enter in the cavity forming characteristic pipe
forms. Once solidification is complete, only limited solid feeding through elastic and
plastic deformation is possible. The latter is responsible in part of the formation of
caved surfaces on the exterior of the part.

20
 2 POROSITY LITERATURE REVIEW

2.2 Shrinkage Porosity Simulation

There are many commercial software packages that give results for shrinkage porosity.
Each of them have to select the physics that will be modeled in order to have an
accurate model. However, users usually have a limited computational capacity and a
trade-off between accuracy and computational time must be satisfied. For this reason
one must be very careful deciding which aspects of the solidification and porosity
modeling will be taken into account. In general, we can number the following aspects
that will have influence in the nucleation and growth of the pores.

• Thermal variables such as the temperature and solid fraction.

• Solid phases growth

• Flow variables such as the pressure or the velocity.

• Interfacial characteristics between the gas and liquid or solid phases.

We will review later porosity models that will have in account some of the phenomena
mentioned above. However, none of them will have all of them into consideration as
such a model would be too costly and inefficient for commercial purposes. Besides, we
have to be aware that many of the parameters that are required to perform a simulation
are unknown and have to be estimated. Even if we do experimental measures, the
related error of the own measurement procedure might be higher than any other
inaccuracy that we might introduce in the model due to simplifications. This fact
only reinforces the idea that it was stated before that the final model that will be
implemented must ensure a not only precise but also efficient calculation.

In general, the existing models to compute the porosity can be classified into
different groups according to Lee et.al [20] and Stefanescu [30].

• Criterion Function Models

• Thermal/Volume Calculation Models

• Thermal/Fluid Flow Models

We will further classify them according to type of porosity that they aim to model.

21
 2 POROSITY LITERATURE REVIEW

• Microporosity Models : Criterion Functions

• Macroporosity Models : Thermal/Volume Calculation Models

• Coupled Microporosity and Macroporosity Models : Thermal/Fluid Flow Mod-

els

For each of the families above it will be shown different implementation methods in
the following sections.

2.3 Microporosity Models : Criterion Functions

Criterion functions are simple rules that relate the local conditions to the propensity
to form micropores. The obtained criterion are functions of thermal parameters such
as the cooling rate, the thermal gradient or the solidification time and material pa-
rameters as the shrinkage factor, the viscosity etc. These models take into account
simplified formulations for the transport problems, however they usually ignored the
contribution of gas rejected by the solidifying melt to microporosity. The application
of these criterion started a long time ago with Pellini in 1953 [26]. Pellini was one of
the first persons to realize the correlation between the presence of porosity and the
gradient of the temperature. He stated that the temperature gradient G should be
greater than a critical value Gcr to avoid centerline shrinkage porosity.

Niyama took this idea from Pellini and developed his own criterion function sim-
ply known as Niyama Criterion [24]. This model has become quite popular specially
between Japanese foundry men and it has been proved to very useful for detecting
microporosity in ferrous alloys. On the other hand, its implementation for non-ferrous
alloys has been frequently questioned[30] . As Pellini did before, Niyama et al. took
measures in a number of commercial castings and confirmed the previous experiments
that had related the formation of shrinkage porosity to the gradient of the tempera-
ture. However, they found that the critical value under which microporosity could be
expected depended largely in other factors as well such as the shape of the casting or
its size so it could not be predicted in advanced [23]. They proposed therefore that
the critical value of the gradient of the temperature depended also on the solidifica-
√
tion time ts . More specifically they stated that it should be proportional to 1/ tS .
Approximating the solidification time as tS = 4Tf /Ṫ , where 4Tf is the difference

22
 2 POROSITY LITERATURE REVIEW

between the solidus and liquidus temperatures and Ṫ is the cooling rate, they arrived
to the well-known Niyama Criterion:

∇T
Ny = p (2.1)
Ṫ

With this criterion, we only need to calculate the local values of the parameters
and compare them with a critical value that will not depend on the size of the casting.
Each material will have a specific critical value under which microporosity should be
expected. Niyama values are evaluated near the end of solidification, when microp-
orosity forms. Different studies and commercial software evaluate Niyama at different
moments, but it is usually calculated at a temperature which is a 10% of the solidifi-
cation range above the solidus temperature, that is TN y = Tsol + 0.1(Tliq − Tsol ).The
choice of Niyama evaluation temperature will significantly influence the results [16].

Finally, Niyama et al. included in [24] the analytical solution of a 1D Darcy flow
to compute the pressure drop within the mushy zone. They used this analysis to
give theoretical support to their choice for the formulation for the Niyama Criterion.
Despite of its wide use inside the industry and its ability to avoid the size problem
that was present in the Pellini Criterion, Niyama fails to solve the shape and material
dependency problem as some authors have stated [11]. For this reason it is should be
carefully used when dealing with complex geometries..

Although Niyama is with no doubt the most important criterion function, quite a
few authors have proposed their own criterion. Lee et al. [21] claimed that the critical
2/3
value under which feeding becomes difficult and porosity was ∇T · tf /Vs where VS
is the solidification velocity. Finally, we can also mention the work of Hansen and
Sahm [11], where they stated that the shape-dependency problem could be overcome
using as a function criterion ∇T Ṫ 1/4 U −1/2 ,where U is the feeding velocity. However,
this claim has only been proved for simple bars and plates and there has not been
further work with this criterion. Besides, in order to compute the feeding velocity, we
have to solve the flow equations, making this method much expensive compared to
the Niyama criterion.

23
 2 POROSITY LITERATURE REVIEW

2.4 Macroporosity Models : Thermal / Volume Calculation

Models

These models only solve the thermal problem and attempt to find porosity areas
calculating volume changes using mass conservation rules. Their point is to avoid
calculating a full flow model that would make the whole process much more expensive
in terms of computational cost. In the model proposed by Imafuku and Chijiiwa
[14] porosity is predicted by evaluating the volume of solidification shrinkage in each
isolated liquid pool that forms as the solidification advances and solid areas appear.
This volume is then removed from the top cells of the liquid pools ensuring that the
mass is conserved. The model assumes that the gravity feeding occurs instantaneously,
considering that the flow velocity due to solidification can be neglected and liquid melt
moves only under the effect of gravity. This approach is supported by the fact that in
many situations fluid flow in the solidification metal can be ignored. A permeability
limit has to be set up from which melt is considered solid and cannot move anymore.
The net change in volume because of shrinkage is calculated with 4v = βv0 4fL ,
where v0 is the initial volume and β = (ρs − ρl )/ρl is the shrinkage ratio being ρs and
ρl the solid and liquid densities respectively. This method can be used to calculate
both macroporosity and open shrinkage. If the volume loss is on the free surface, that
is, on the top of the riser, the change of volume is compensated by lowering the level
of liquid. If it occurs inside the casting, the loss will be compensated by introducing
a void.

A similar model was developed by Beech et al. [1] where the amount of macro-
porosity was also calculated exclusively from the thermal results as the change of
volume 4v = βv0 gs F where gs is the volume fraction of solid in the volume element
and F is the liquid fraction. The main difference with the previously proposed method
is the introduction of a drag coefficient K as part of the feeding criterion. This drag
coefficient K is equal to 0 if the solid fraction is equal or lower than a certain value
fScoh which means that feeding can occur and equal to infinite if the solid fraction is
higher than fscoh . This last situation would correspond to the case where the volume
is considered solid and flow feeding is not allowed anymore.

24
 2 POROSITY LITERATURE REVIEW

2.5 Coupled Microporosity and Macroporosity Models : Ther-

mal/ Fluid Flow Models

In these models, a more or less complex formulation is developed where the fluid flow
is explicitly solved along with the thermal calculation.

Kubo & Pehlke

One of the first models to deal with porosity in such a way was proposed by Kubo
and Pehlke [19] in 1985. In this 2D model, the flow was modeled using Darcy’s Law

k ∂P
u=− (2.2)
µfl ∂x

k ∂P kρg
v=− − (2.3)
µfl ∂y µfl

along with continuity equation

 
ρs ∂fl ∂fl u ∂fl v ∂fp
−1 − − + =0 (2.4)
ρl ∂t ∂x ∂y ∂t

where u and v are the components of the velocity, µ is the dynamic viscosity, ρ
is the density, g is the gravity and fl and fp are the liquid and porosity fractions
respectively. The liquid fraction present in the continuity equation is calculated from
the thermal analysis using the Scheil equation.

 1/(kp −1)
T − Tliq
fs = 1 − (2.5)
mCo

being kp the partition coefficient, m is the liquidus slope and C0 is the initial solute
concentration. When the mushy zone was reached the permeability is computed based
on the Blake-Kozeny model.

fl3 d
k= (2.6)
180(1 − fl )2

25
 2 POROSITY LITERATURE REVIEW

where d is the dendrite cell size. As many other models do now, Kubo and Pehlke
established that porosity formed when the following mechanical equilibrium is fulfilled.

2γ
Pp = P + (2.7)
r

This equation basically says that porosity forms when an equilibrium is reached
between the exterior pressure in the liquid P plus the pressure due to the gas/liquid
interfacial energy 2γ/r is equal to the pressure inside the pore Pp . In this equation r
corresponds to the radius of the pore and γ is the surface tension. In this model, Kubo
and Pehlke defended that the porosity formed at the root of the secondary dendrite
arms, the radius is thus assumed to be half of the Secondary Dendrite Arm Spacing.
The pressure in the pore is calculated with Sievert’s Law.

The algorithm to solve the problem is the following for each time step. When
no porosity has formed, metal pressure P is calculated with the momentum and mass
equations. Then the pore pressure is calculated with equation 2.7 .Subsequently a new
amount of porosity is calculated from Sievert’s Law. If porosity has already formed
and the flux of liquid is negative (fluid leaving the element), porosity is calculated
using the mass and momentum equations using the pressure from the previous step.
With this value of porosity, the pressure in the pore Pp is estimated using Sievert’s
Law and then a new value for the pressure is computed using P = Pp − 2γ/r. If the
flux of fluid is positive, the pressure is calculated from the governing equations using
the porosity obtained from the previous step. Then the porosity is updated again
using Sievert’s Law.

The main idea behind this algorithm has been the base for many the algorithms
that have been developed until this date.

Bounds

In 2000, Bounds et al. [2] presented their work attempting to use a pressure-based
criterion for predicting misruns, shrinkage porosity and pipe shrinkage in a 3D Model.
In their formulation, energy, mass and momentum equations are solved adding terms
to account for the contraction of the metal and the porosity evolution. The continuum
equation is then

26
 2 POROSITY LITERATURE REVIEW

∇ · u = Sm + Sg (2.8)

Where Sm represents the contraction of the metal and Sg is the gas evolution. The
author proposes the following iterative scheme to compute the contraction of metal.

 
i+1 i ∂fl ρl
Sm = Sm − + ∇ · (ρl ul ) (2.9)
∂t

The liquid fraction fl is tracked using a scalar advection equation

∂fl
+ ∇ · (fl ul ) = Sm, (2.10)
∂t

The momentum conservation equation is also modified to account for the presence
of fluid and solid phases by introducing and effective viscosity µef f and a Darcy source
term SD .

∂ρu
+ ∇(ρuu) = ∇ · (µef f ∇u) − ∇P + ρg + SD (2.11)
∂t

The effective viscosity µef f is chosen to account for the solid-liquid mixture as a
function of the solid fraction and the dendrite coherency point. The Darcy source
term SD is simply:

F D µl u
SD = (2.12)
K

where K is the permeability and FD is a switching functions that ensures that

SD ≈ 0 when the solid fraction is lower than the dendrite coherency point. The
permeability is modeled as usual, using the Kozeny-Carmen equation:

d2 (1 − fs )3
K=C (2.13)
fs2

where C is a constant and d is a characteristic dimension, usually related to either

the primary or the secondary dendrite arm spacing.

27
 2 POROSITY LITERATURE REVIEW

Shrinkage porosity results from a pressure drop in the liquid metal when the feed
path to a solidifying region is obstructed by solid or partially solidified material, the
proposed model assumes that below a specified pressure Pp stable gas pores will nu-
cleate within the liquid phase and grow to account for the volume space vacated by
metal contraction during cooling and solidification. This condition is tested for every
element. If it is satisfied feeding is allowed and the element gas evolution term Sg is
set to zero. Otherwise, the element gas evolution term Sg is set Sg = −Sm for the rest
of the solidification.

Pequet

Based on the model previously described created by Kubo and Pehlke [19], Pequet
et al. developed an improved 3D model in 2002 [27]. They coupled a microporos-
ity model based on Darcy flow and gas micro-segregation with a macroporosity and
shrinkage pipe prediction model. In this model the flow equations were only solved
in the mushy zone using appropriate boundary conditions. In order to get accurate
results, they used a refining technique to have a better mesh in these regions. The
governing equations were simply:

∂
(ρs fS + ρl (1 − fs − fp )) + ∇ · (ρl fl vl ) = 0 (2.14)
∂t

K
vl = − (∇P − ρl g) (2.15)
µ

This model has two equations and three unknowns vl , P and fp , so a supplementary
equation of state is require. This additional equation is furnished by the segregation
and precipitation of hydrogen. In order to have a well posed problem, boundary
conditions must be imposed around the mushy zone.

For closed areas of liquid, surrounded by either solid regions or the mushy zone,
the volume of contraction is calculated estimating the flux of liquid through the mushy
zone. This volume is later subtracted from the cells in the top of the closed region.

For the free boundaries where a Dirichlet condition is imposed, the calculated pres-
sure field allows the deduction of the velocity field at such boundaries and thus pipe
shrinkage lowering the free surfaces.

28
 2 POROSITY LITERATURE REVIEW

Carlson

Carlson presented in 2002 [6] a multi-phase model that predicts melt pressure,
feeding flow and porosity formation and growth in steel castings during solidification.
Carlson derived a momentum equation by combining Darcy’s Law and Stokes flow
that is valid everywhere in the mushy zone.

fl fl fl
∇2 v = v + ∇P − ρg (2.16)
K µl µl

The permeability K is computed in the usual way using the Kozeny-Carmen equa-
tion.

(1 − fs − fp )3
K = Kr K0 (2.17)
(fs + fp )2

Where K0 = 6 × 10−4 λ21 , being λ1 the primary dendrite arm spacing, and Kr is
the relative permeability between air and liquid Kr = fl / (fl + fa ). When the solid
fraction is low, the permeability tends to infinite and the momentum equation reduces
to Stokes’ flow equation. When the solid fraction is high at the end of the solidification
the laplacian term in the left hand side is very small and the equation is equivalent to
Darcy’s Law.

Finally the species equation accounts for macro segregation of gas species due to the
flow. Once the total gas pressure is high enough to cause pore nucleation, the amount
of porosity that forms is determined from the continuity equation. This multi-phase
model has been successfully implemented in a general-purpose casting simulation code.
This model has been adapted and will be describe in detail in section 5.

29
 3 GRADIENT RECONSTRUCTION

3 Gradient Reconstruction

As stated in the introduction and in section 2 several methods for computing poros-
ity are present in the literature about casting. In this section and the following we
will focus on microporosity. Due to the nature of this phenomena, where different
processes affect each other in a very small scale, criterion function models are usually
chosen, using only local values of the variables of interest. Between these models, the
Dimensionless Niyama model proposed by Carlson and Beckermann [5] has gained no-
toriety during the last years and it has been implemented in some commercial software
packages specialized in casting simulation. Unlike other criterion function models like
[24, 26] this method accounts for both thermal and material properties. The key as-
pect of this criterion is its dependency on the gradient of the temperature. Therefore,
an accurate and efficient method has to be implemented to recover the gradients from
the known temperature field that we have already calculated. With this aim, several
methods will be explained in the following sections and will be latter compared in
terms of precision and computational cost.

3.1 Introduction

In many problems in the field of engineering the recovery of gradients is needed for
different reasons. For example in the field of the fluid dynamics one might be interested
in calculating the drag coefficient or the vorticity of the field. This quantities can
be determined as a post-process of the previously obtained solution for the pressure
and the velocity. In our case, we are interested in calculating the microporosity of a
casting part during solidification. It has been experimentally proved that the presence
of microporosity is closely related to areas where the gradient of the temperature is
low.

The most simple way to obtain the gradient of a known field is just to take deriva-
tives inside each element. If our finite element solution was of degree k, then the
gradient field that we obtain in such a way is of degree k -1. This means that if we
use linear finite elements then our solution will be discontinuous on the nodes. We
want of course our solution to be continuous so we will focus on methods to obtain a
gradient field such that it belongs to the same space of interpolation as the original

30
 3 GRADIENT RECONSTRUCTION

solution.

There are many authors that have been interested in this problem in the field of
error estimation. In finite elements problems it can be proved that the error of the
solution is a function of the element size h, therefore as h approaches zero so does the
error. Evidently, this will imply also a higher computational cost. The approach taken
in the called adaptive mesh refining methods is just to refine those regions where the
errors are higher. Following the lines firstly proposed by Zienkiewicz and Zhu [25], the
usual way to compute these a posteriori errors is to evaluate them as the difference
between the piecewise discontinuous gradients (if we use linear elements) and the so
called superconvergent post-processed recovered gradients. The reason for this name
is that according to the authors the rate of convergence of these methods is at least
quadratic, that is, order 2 with respect to the element size. With this purpose, some
authors proposed their own method to recover the gradients [32, 28, 25]. In this
chapter we will explore some of these proposed methods.

After each models is explained in detail, we will find out which one is the best
for our needs. We will compare the following methods in terms of the L2 errors with
respect to analytical functions and the computational time.

• L2 Projection with Consistent Mass Matrix

• L2 Projection with Iterative Newton Raphson

• L2 Projection with Lumped Mass Matrix

• Zhang & Naga Method

• Pouliot Recover Method

These methods will be tested for a structured regular mesh with different element sizes.

Element Size Number of Nodes Number of Elements

0.100 1331 6000
0.050 9261 48000
0.025 68921 384000
0.020 132651 750000

Table 3.1: Meshes used for testing gradient recover methods

31
 3 GRADIENT RECONSTRUCTION

Two different analytical functions will be used in order to asses the accuracy of
each method.

Case 1
f1 = sin(πx) sin(πy) sin(πz) (3.1)

Case 2
f2 = e−25x + e−25y + e−25z (3.2)

With the corresponding analytical gradients:

 
πcos(πx) sin(πy) sin(πz)
∇f1 = πsin(πx) cos(πy) sin(πz) (3.3)
 

πsin(πx) sin(πy) cos(πz)

 
−25e−25x
∇f2 = −25e−25y  (3.4)
 

−25e−25z

The error e will be measured using the L2 norm, defined as:

Z
e= (∇Tana − φh )2 dV (3.5)
Ω

This integral is going to be approximated using the obtained nodal values of the
gradient and the nodal volumes V i .

X
i

2
e= ∇Tana − φi Vi (3.6)
nodes

Where ∇Tana is the analytical gradient obtained either from equation 3.3 or 3.4 as
it corresponds and φh is the solution of a specific gradient recovery method.

In Kratos, the nodal volume is defined as the sum of the volumes of the elements
to which that specific node belongs divided by the number of nodes of that element,
four for our case since only tetrahedron are used in Click2Cast.

32
 3 GRADIENT RECONSTRUCTION

3.2 L2 Projection Method

The fist and most simple approach that we can think of is to project the discontinuous
gradient of the temperature field that we have already calculated into the L2 space.
The method and the subsequent implementation follows the same principles explained
in [15].

If we have previously calculated a temperature field Th , we can obtain a function

for the gradient ∇T by simply taking the spatial derivative. If we are using linear
elements, our function Th will belong to the linear finite element space Vh ∈ L2 and
∇T will be a piecewise continuous function.

X
Th = Ni Ti (3.7)

X
∇T = Ni,j Ti (3.8)

The most simple approach is to project ∇T into the finite element space using an
L2 Projection. That is, to find a solution φh ∈ Wh , being Wh a finite element space
Wh ∈ (L2 )3 , such that minimizes the error e a in a Least Squares sense.

Z
e(x) = (ϕh (x) − ∇T (x))2 dΩ (3.9)
Ω

This minimization problem can be solved in the following equivalent way. We first
take derivative with respect to ϕ and multiply by the test function N , arriving to the
weak form of the problem.

Z
N · (ϕh (x) − ∇T (x))dΩ = 0 (3.10)
Ω

The arbitrary test functions N are the shape functions of our finite element space
Wh .

33
 3 GRADIENT RECONSTRUCTION

Z Z
N · ϕh (x)dΩ = N · ∇T (x)dΩ (3.11)
Ω Ω

Knowing that

X
ϕh = Nj ϕj (3.12)

We get

Z Z
Ni Nj ϕj dΩ = Ni ∇T (x)dΩ (3.13)
Ω Ω

Using the following compact notation

Z
MC = Ni Nj dΩ (3.14)
Ω

Z
f= Ni ∇T (x)dΩ (3.15)
Ω

Rewriting the system we obtain

MijC ϕj = f i (3.16)

We can now build the whole system and solve it. This global system is going to
be expensive in terms of computational cost and must be avoided when possible. The
problem can be simplified using a lumped mass matrix ML instead of the consistent
mass matrix MC . For a tetrahedral element:

 
2 1 1 1
 
MC = 1
V olume  2 1 1
 (3.17)
20  1 1 2 1

1 1 1 2

34
 3 GRADIENT RECONSTRUCTION

 
1 0 0 0
 
ML = 0
V olume  1 0 0
 (3.18)
4 0
 0 1 0

0 0 0 1

As we see, the lumped mass matrix is obtained adding the values of the rows in
diagonal terms. This lumped mass matrix is diagonal and therefore the global system
will be diagonal as well. We can avoid then to solve the global system as each node
can be solved independently. The diagonal terms of the left hand side will coincide
with the Nodal Volume Vi , so we can simply calculate the nodal values as the sum of
the element contributions divided by the Nodal Volume.

P
cont elems fi
ϕi = (3.19)
Vi

3.2.1 Results

Figure 3.1: L2 Projection - L2 Error for Case 1

35
 3 GRADIENT RECONSTRUCTION

Figure 3.2: L2 Projection - L2 Error for Case 2

Figure 3.3: L2 Projection - Computational Time for Case 1

36
 3 GRADIENT RECONSTRUCTION

Figure 3.4: L2 Projection - Computational Time for Case 2

Method Case 1 Case 2

L2 Lumped 1.52 1.35
L2 Consistent 1.51 1.45

Table 3.2: L2 Projection - Convergence Rates

3.3 Newton-Raphson iterative scheme

When we substitute the mass matrix Mc by the lumped mass matrix Ml we are doing
an approximation, and therefore we will obtain a different solution from the one we
would have using the consistent mass matrix. This issue can be overcome using a
iterative Newton-Raphson scheme.

From the equation 3.16 we can define the following residual

R(ϕ) = f i −MijC ϕj (3.20)

Our objective is to find the values of ϕ that make R(ϕ) = 0. If we start from
an arbitrary solution φi , we can approximate our solution using a 1st order Fourier

37
 3 GRADIENT RECONSTRUCTION

expansion

∂R(ϕi )
R(ϕi+1 ) = R(ϕi ) + i
(4ϕi+1 ) = 0 (3.21)
∂ϕ

Rearranging terms we get

∂R(ϕi )
4ϕi+1 = −R(ϕi ) (3.22)
∂ϕi

From the equation 3.20 we calculate

∂R(ϕi )
= −MC (3.23)
∂ϕi

Including this result into the previous equation we get the following iterative
scheme

MC 4ϕi+1 = f −MC ϕ (3.24)

Now, we can approximate the mass matrix of the left hand side with the lumped
mass matrix to obtain the following iterative scheme

4ϕi+1 = M−1
L (f −MC ϕ) (3.25)

The solution of this system will converge to ϕ = MC −1 f in a number of steps.

Other methods have been proposed following similar lines, like D.M Hawken et al.
proposed in [13]. In this paper a scheme with optimal convergence was given starting
from the same equation that we formulated in 3.16.

MijC ϕj = f i (3.26)

In this occasion the iterative scheme that Hakwen proposes has an iterative con-
vergence acceleration parameter w such that

38
 3 GRADIENT RECONSTRUCTION

ML 4ϕi+1 = (f − MC ϕ) w (3.27)

If λl and λu are the lower and upper bounds of the eigenvalues of the matrix
−1
ML MC , then the value if w which gives optimal convergence rates wopt is given to
be

2
wopt = (3.28)
λl + λu

Convergence of the scheme is guaranteed provided wλu < 2. Hawken calculated

this parameter w to be wopt = 1.7274.

3.3.1 Results

Figure 3.5: L2 Iterative - L2 Error for Case 1

39
 3 GRADIENT RECONSTRUCTION

Figure 3.6: L2 Iterative - L2 Error for Case 2

Figure 3.7: L2 Iterative - Computational Time for Case 1

40
 3 GRADIENT RECONSTRUCTION

Figure 3.8: L2 Iterative - Computational Time for Case 2

Method Case 1 Case 2

L2 Lumped 1.52 1.35
L2 Iterative 1.51 1.45
L2 Consistent 1.51 1.45

Table 3.3: L2 Iterative - Convergence Rates

3.4 Zhang Method

The methods mentioned above can give us reasonable results with a moderate com-
putational cost. In most cases we can expect them to present linear convergence at
most. Besides these classic approaches, other methods have been developed more re-
cently specially focused on a-posteriori error estimators where the rate of convergence
is higher than one. This family of methods are called superconvergent methods. The
first of these methods was developed by Zienkiewicz and Zhu [25] and it has been ob-
ject of many debates [13], where some of its deficiencies have been exposed. In order
to overcome them, others authors have proposed their own superconvergent gradient
recover methods.

In 2005, a new method for gradient recovery with superoconvergence was proposed

41
 3 GRADIENT RECONSTRUCTION

by Zhang et. al. [32]. The method was first proposed to address some of the problems
detected in the widely used Zienkiewicz-Zhu method. Zhang et. al. claimed that the
ZZ method was not superconvergent for linear elements under rectangular mesh of the
Chevron pattern, nor for quadratic elements at edge centers under uniform triangular
mesh of the regular pattern. The method that will be exposed now is supposed to
overcome these limitations.

Given a finite element space of degree k, the idea is to build a polynomial of order
k + 1 using a cloud of nodes big enough around each node. Once this polynomial is
built, the gradient is recovered by simply taking derivatives. The resulting polynomial
will be of order k, fitting in the finite element space. This method is proved to be
superconvergent and even ultraconvergent for some special cases (order of convergence
greater than 2) and its cost is comparable to the ZZ patch recover method.

3.4.1 Description of the method

The objective is to find a recovery gradient operator Gh : Sh → Sh × Sh , where Sh is a

polynomial finite element space of degree k over a triangulation τ h .Given a solution uh ,
we need to define the operator Gh at the three possible types of points: vertices, edge
nodes and internal nodes. This way, the method is suitable for elements of any degree
of interpolation and not just for linear elements. Once the value of the gradient is
known at the nodes, the solution must be obtained for the whole domain interpolating
with the original shape functions of Sh . The method will be presented for a 2D case,
however its extension to the 3D case can be done in a similar way.

1. Vertices: For a vertex zi , let hi be the length of the longest edge attached to zi
we select all nodes on the ball.

Bhi (zi ) = {z ∈ D : |z − zi | ≤ hi } (3.29)

If the number of nodes n is lower than m = (k + 2)(k + 3)/2, we do not have

enough nodes to build the polynomial and we have to go further and include all nodes
in B2hi (zi ) repeating the process until we have a sufficient number of nodes.These
nodes around the point zi will be defined as zij . Setting the coordinates of zi as the
origin, the fitting polynomial is

42
 3 GRADIENT RECONSTRUCTION

pk+1 (x, y; zi ) = PT a = PˆT a (3.30)

with

PT = (1, x, y, x2 , ..., xk+1 , xk y, ..., y k+1 ) (3.31)

PˆT = (1,ξ,η, ξ 2 , .., ξ k+1 , ξ k η, .., η k+1 ) (3.32)

aT = (a1 , a2 , ..., am ) (3.33)

âT = (a1 , ha2 , ..., hk+1 am ) (3.34)

The coefficient vector â is determined by the linear system

AT Aâ = AT bh (3.35)

where bT
h = (uh (zi1 ), uh (zi2 ), ..., uh (zin )) and

 
1 ξ1 η1 . . . η1k+1
. . . η2k+1 
 
1 ξ2 η2
 
A = 1
 ξ3 η3 . . . η3k+1 
 (3.36)
. .. .. .. .. 
 .. . . . . 
 
1 ξn ηn . . . ηnk+1

Now we define:

Gh uh (zi ) = ∇pk+1 (0, 0; zi ) (3.37)

43
 3 GRADIENT RECONSTRUCTION

2. Edge node: For a node which lies on an edge between two vertices zi1 and zi2 ,
we define:

Gh uh (zi ) = α∇pk+1 (xj , yj ; zij ) + (1 − α)∇pk+1 (x2 , y2 ; zi2 ) (3.38)

The weight α is determined by the ratio of the distances of zi to zi1 and zi2 .

3. Internal node. For an internal node which lies in a triangle formed by zi1 ,zi2
and zi3 . We define :

3
X
Gh uh (zi ) = αj ∇pk+1 (xj, yj ; zij ) (3.39)
j=1

The weight αj is determined by the ratio of the distances of zi to zij .

3.4.2 Results

Figure 3.9: Zhang - L2 Error for Case 1

44
 3 GRADIENT RECONSTRUCTION

Figure 3.10: Zhang - L2 Error for Case 2

Figure 3.11: Zhang - Computational Time for Case 1

45
 3 GRADIENT RECONSTRUCTION

Figure 3.12: Zhang - Computational Time for Case 2

Method Case 1 Case 2

Zhang 2 1.88

Table 3.4: Zhang - Convergence Rates

3.5 Pouliot Gradient Recover Method

As well as Zhang and some others before, Pouliot et al. became interested in the gradi-
ent recovery in the field of a posteriori error estimators. This error is simply measured
as the difference between the recovered gradient and the piecewise discontinuous solu-
tion gradients that we can directly obtain from our finite elements solution. Therefore,
a number a methods have been developed with the aim of approximate the gradient
of the solution called super-convergent methods.

These error estimators are then used to drive isotropic or even anisotropic mesh
adaptation. For a linear FEM solution the error is directly linked to the Hessian
matrix, so not only the gradient but the Hessian matrix (the gradient of the gradient)
needs to be recovered. With a more suitable method to recover the gradients, this can
be achieved more efficiently, with a very practical application in almost all engineering

46
 3 GRADIENT RECONSTRUCTION

problems that involve finite elements.

In the paper published by Pouliot et al. [28], a review of the methods that were
beings used up to that date. As it was pointed out by Zhang and Naga in [32], the
most popular of the superconvergent methods, the Zienkiewicz and Zhu patch method,
lost its superconvergence property when non regular meshes were used.

In the same paper, Pouliot mentioned the Zhang method from [32], recognizing its
capability to maintain its superconvergence in a wide number of cases. Despite the
fact that this method can provide very good results inside the domain, Pouliot et al.
pointed out that as in most of the gradient recovery methods, annoying oscillations
were observed near the boundaries of the domain. Finally, Pouliot also mentioned the
L2 Projection method with a consistent mass matrix that we explained in section 3.2
though this method was disregarded by Pouliot due its linear convergence rate and
its relatively high computational cost.

In the Pouliot Recover Method we have to build a global system that must be
solved using an iterative method. The method constructs linear equations using su-
perconvergent points on each edge of the mesh. This method, as well as the others
presented in this section is superconvergent but besides it should be particularly good
at reducing the oscillations near the boundaries.

3.5.1 Description of the Method in 1-D

To expose the method we will begin with a 1D case where the domain is an interval
[a, b] divided in n + 1 elements and n nodes such that

a = x0 < x1 < ... < xi < ... < xN = b (3.40)

We start from a piecewise linear solution uh , which nodal values are already known.
The derivative of this function is a piecewise constant function. The objective is to
reconstruct the function using the nodal values to obtain dh , a piecewise linear and
continuous approximation of the derivative of our function uh . Pouliot achieves this
by introducing a bubble function ubi defined by ubi (x) = Ci (x − xi )(x − xi+1 ) in the
interval of the element [xi , xi+1 ] and vanishing elsewhere and where Ci are unknown
DOFs. This function is added to our previous solution obtaining a new solution

47
 3 GRADIENT RECONSTRUCTION

up = uh + ub If the coefficients Ci are properly chosen we can expect our solution to

go from second-order to a third-order approximation of u in L2 . The derivative our
function is now simply

0 0 0
up (x) = dh (x) = uh (x) + ub (x) (3.41)

Notice that the bubble function is unknown so we cannot directly compute dh .

However if we want to evaluate the derivative in the middle point of the element
xs = (xi + xi+1 )/2, the derivative of the bubble function is always equal to zero.
These points are called superconvergent points.

0
dh (xs ) = uh (xs ) (3.42)

Knowing that dh and uh are linear functions, we can write this equation just in
term of the nodal values
di + di+1 ui+1 − ui
= (3.43)
2 xi+1 − xi

Which can be rewritten in an elementary form as

" #" # " #

1 1 di −2 iu − ui+1
= (3.44)
1 1 di+1 hi ui − ui+1

Notice that we are imposing one equation for each edge of the mesh. It usually
happens that meshes have more nodes than edges and therefore our system will be
singular and we will not be able to solve it. To overcome this problem, Pouliot et. al.
propose to modify the system adding small terms in order ti stabilize the formulation.
As an example, the propose to use a Laplacian operator to the derive multiplied by a
stabilization parameter h .
(" # " #) " # " #
1 1 1 −1 di −2 ui − ui+1
+ h = (3.45)
1 1 −1 1 di+1 hi ui − ui+1

The parameter h is usually chosen small (h = 10−6 h)

48
 3 GRADIENT RECONSTRUCTION

3.5.2 The Multidimensional Case

The generalization of the method to the 2 and 3D cases is quite straight from the
1D case. In this cases, the superconvergent points are located in the middle of the
edges of the mesh. We have to replace now the derivatives at nodes by the directional
derivatives along the edges. We will show the formulation of the 2D case, being the
3D case very similar.

For each edge we calculate:

Lex = xe2 − xe1 (3.46)

Ley = y2e − y1e (3.47)

q
he = (Lex )2 + (Ley )2 (3.48)

1 e e
le = (lxe , lye ) = (L .L ) (3.49)
he x y

Figure 3.13: Element Edge [28]

Now we write the value of the directional derivative in the center of the edge ass
a function of the nodal values, arriving to:

de1 + de2 e ue2 − ue1

·l = (3.50)
2 he

49
 3 GRADIENT RECONSTRUCTION

This can be written as:

    
(lxe )2 lxe lye (lxe )2 lxe lye dex,1 lxe (ue1 − ue2 )
 e e
 lx ly (lye )2 lxe lye (lye )2  dey,1 
e e
ly (u1 − ue2 )
  
2
(le )2 le le (le )2 le le  de  = − he (3.51)
    
le (ue − ue )
 x x y x x y   x,2  x 1 2 
lx ly (ly ) lx ly (lye )2
e e e 2 e e
dey,2 e e e
ly (u1 − u2 )

This elementary system is computed for each element of the mesh and then as-
sembled in the global system. As a remark, in general, for 2D and 3D meshes, there
are more nodes than edges so the problem will be non-singular. However, this is not
the case for regular meshes, for which stabilization has to be added just as it was
explained in the previous section.

3.5.3 Results

Figure 3.14: Pouliot - L2 Error for Case 1

50
 3 GRADIENT RECONSTRUCTION

Figure 3.15: Pouliot - L2 Error for Case 2

Figure 3.16: Pouliot - Computational Time for Case 1

51
 3 GRADIENT RECONSTRUCTION

Figure 3.17: Pouliot - Computational Time for Case 2

Method Case 1 Case 2

Pouliot 2 1.82

Table 3.5: Pouliot - Convergence Rates

52
 3 GRADIENT RECONSTRUCTION

3.6 Comparison of results

Figure 3.18: Comparison - L2 Error for Case 1

Figure 3.19: Comparison - L2 Error for Case 2

53
 3 GRADIENT RECONSTRUCTION

Figure 3.20: Comparison - Computational Time for Case 1

Figure 3.21: Comparison - Computational Time for Case 2

54
 3 GRADIENT RECONSTRUCTION

In figures from 3.18 to 3.21 we have a reunited all the computed results along with
the convergence rates from table 3.6.

In terms of computational time, the fastest methods are the iterative and lumped
L2 projections. These methods only involve local systems where the solution for each
node can be solved independently. This was achieved by substituting the mass matrix
by the lumped mass matrix. The latter is diagonal and therefore the system of equa-
tions is uncoupled as it was explained in subsection 3.2. Zhang method also involves
only local calculations but it takes larger times to build the necessary polynomials.
Besides, in the implementation that was made in Kratos, some parts of the code were
not parallel though they could be. To some extent, the differences in time with respect
to the L2 projections are definitely due to this factor since in the latter, all the parts
that admitted parallelization were implemented using OpenMP.

The two remaining methods involve the build and solution of a global system,
which explains their higher computational requirements. In the cases tested, Pouliot
method was slightly faster than the Consistent Method. However, we cannot ensure
that this would be the case for any scenario as Pouliot speed usually depends highly
on the stabilization parameter . For all cases, this value was set to  = 10−5 h, being
h the element size.

Regarding the accuracy of each method, Pouliot method is superior to the rest
followed by Zhang method as it was claimed in [28]. In third place we have the consis-
tent method, though we can see that with only a few iterations, the iterative method
gives practically the same error as the Consistent Method with an important reduc-
tion of cost in time. For both functions, using an iterative scheme with 3 iterations
we can have solutions with practically the same errors as the ones obtained with the
consistent mass matrix but with a computational time that is at least one order of
magnitude lower.

The following table sums up the convergence rates for the different cases.

55
 3 GRADIENT RECONSTRUCTION

Method Case 1 Case 2

L2 Lumped 1.52 1.35
L2 Iterative 1.51 1.45
L2 Consistent 1.51 1.45
Pouliot 2 1.82
Zhang 2 1.88

Table 3.6: Convergence Rates

It might be surprising that despite all that we have been saying in this section, the
L2 Projection methods present some degree of superconvergence (rate of convergence
higher than one). This is due to the use of regular meshes where some sources of er-
rors cancel each other. For the first function, both Pouliot and Zhang show quadratic
convergence as stated by their authors. All the methods present degradation in per-
formance for the second function. The main reason for this is the higher complexity
of the function. Besides, we have to take in account that the L2 error is approximated
and therefore some source of error could be induced by this approximation.

Both superconvergent methods show a superior convergence rate compared to the

iterative L2 solution, however the latter is the fastest by far, especially in the second
case. Pouliot mentioned a fair point in his publication about this. Although it is true
that for the same mesh, Pouliot method might be slower than others, if we fix a certain
level of error and compare the time required by the different methods to calculate the
error, in general Pouliot will be the fastest. In the tests that we tested, this was true
for the first case but not for the second one. Anyway, in our problem the element
size is going to be determined by the finite elements calculation of the solidification
and the filling which precise a high number of elements in order to have an accurate
result. Another inconvenient is the need of a stabilization parameter  whose optimal
value cannot be known a priori. With respect to Zhang method, we could have tried
to improve the code by making parallel some of the parts of the code. We decided
nevertheless not to continue with this line as, since the Zhang algorithm requires to
know the neighbors of each node, MPI is not possible, whose implementation is a
mid-long term project in Click2Cast.

For these reasons we decided that the best option to recover the gradients to
compute the microporosity was the Iterative L2 Projection.

56
 3 GRADIENT RECONSTRUCTION

57
 4 MICROPOROSITY MODEL

4 Microporosity Model

4.1 Introduction

Microporosity is a major cause of casting rejection or rework and therefore it has a

great impact on the costs for foundry-men. It is usually only detected in on-work
operations, as it cannot be detected by standard radio graphic testing due to the low
scale at which it appears. The most important failure mode due to microporosity is
the appearance of leaks. The usual way to combat the presence of microporosity is the
use of risers. However, there are not precise rules that can guarantee the necessary size
of the mentioned risers, so the only way to ensure the absence of porosity is to over-size
them resulting in additional costs as well. For this very reason, the implementation
of a model capable to predict microporosity is of crucial importance.

Based on the work made by Pellini [26],which related the presence of porosity with
shallow temperature gradients, Niyama et al.[24] built their own criterion based on the
appearance of porosity as a result of the feeding cut. They found out that they could
remove the dependency of the critical temperature gradient with respect to the size of
the casting by dividing the gradient by the square root of the cooling rate. Besides the
experimental validation provided by Niyama, they also included a theoretical analysis
based on Darcy’s Law from which they arrived to the same criterion function, the
following well-known Niyama criterion:

∇T
Ny = (4.1)
Ṫ

Initially, Niyama et al. defended the existence of a critical value N ycr under
which porosity could be expected for steel alloys. The usual value taken for steel is
N ycr = 1.0 (o Cs)1/2 /mm. However, in different studies, the same procedure has been
applied with more or less success for different metals, finding critical values under
which porosity appears for each one. In [5] it is illustrated the correlation between
Niyama values and shrinkage porosity using a end-risered plate 1x5.5x19 in. as shown
schematically in figure 4.1.

58
 4 MICROPOROSITY MODEL

Figure 4.1: Schematic of the end-risered plate [5]

Fifteen plates were cast from WCB steel in a sand mold by three different foundries.
The plates were then radio-graphed and their values were averaged between the 15
plates giving the results shown in figure 4.2

Figure 4.2: (d) Averaged through thickness porosity. (e) Predicted Niyama criterion
distribution at the mid-plane from “Prediction of Shrinkage Pore Volume Fraction
Using a Dimensionless Niyama Criterion [5]

This image is representative of various points. First, we can see from the results
of the radiographies that the prediction of microporosity has an important level of
dispersion due to the many factors involved in the process. For this reason any un-
controlled variable in the casting process will lead to different results. Moreover, the
measurement process itself has its own source of error as special techniques have to

59
 4 MICROPOROSITY MODEL

be used in order to be able to detect porosity. However, despite these reasons, we can
still find some degree of agreement between the regions were porosity was measured
in the radio-graphed plates and the calculated Niyama values.

Niyama criterion is presently the most widely used criterion function and most
casting simulation software packages show it as an important output to predict poros-
ity. Some works have provided support to the use of this criterion, like in [12], in
order predict the hot spots where macroporosity will be present. Other works have
tried to expand its use to predict not only macroporosity but also microporosity, see
for reference [7, 17, 3]. The main idea behind these works is that we can find a value
N ymicro higher than the critical value N yCr under which we will find microporosity.
The reasoning after this is the assumption that the size of the porosity will depend
on the value of N y. Low values of N y correspond to big pore sizes and high values of
N y will correspond to smaller pores. Therefore, if we expand the values of N y under
which porosity is predicted, we can hope to cover the regions where microporosity
forms.

Though it has been widely used since its development, many authors have claimed
its shortcomings as explained in numerous references [17, 30, 31]. At the end, foundries
use the Niyama criterion in qualitative fashion. This limited use is due to the igno-
rance of the threshold Niyama value below which shrinkage porosity forms and the
impossibility to know the actual amount of shrinkage porosity that will be present
in our casting part. One of the most important attempts to overcome some of its
difficulties is the work presented by C. Beckermann and K. Carlson in [5], where a
dimensionless form of the Niyama criterion was presented that aimed to solve the
material dependency. Besides, they provided a model to calculate microporosity in
a quantitative fashion as a function of this Dimensionless Niyama only knowing the
solid fraction - temperature curve and the shrinkage factor of he material. Sigworth
pointed out some limitations for this method in [29], to some extent solved by the own
authors in [4].

In this section, the Dimensionless Niyama method will be explained in detail along
with some study cases. In the last part, we will study the sensibility of the method to
the element size and the point at which temperature is evaluated and how different
casting parameters will affect the microporosity.

60
 4 MICROPOROSITY MODEL

4.2 Dimensionless Niyama Criterion

The dimensionless Niyama criterion is derived from a solidifying 1-D system. The flow
of a liquid inside a porous media is described with Darcy’s Law:

K ∂P
fl vl = − (4.2)
µl ∂x

where l is the liquid fraction, vl is the velocity of the liquid phase, µl is the
dynamic viscosity and P is the melt pressure.

The permeability of the mushy zone K is calculated from the Kozeny-Carman

relation:

fl3
K = K0 (4.3)
(1 − fl )2

K0 = λ22 /180 (4.4)

being λ2 is the secondary dendrite arm spacing (SDAS). Using the shrinkage factor
β = (ρs − ρl )/ρl to simplify the mass conservation equation and then integrating
the result, we can arrive to the conclusion that the velocity of the liquid phase or
shrinkage velocity is constant and can be expressed as vl = −βR, where R is the
isotherm velocity which can be calculated at the same time as the relation between
the gradient of the temperature ∇T and the cooling rate Ṫ . The whole formula will
be thus:

∇T
vl = β (4.5)
Ṫ

Inserting this equation into equation 4.2:

∇T K ∂P
fl β =− (4.6)
Ṫ µl ∂x

reordering terms

61
 4 MICROPOROSITY MODEL

∂P ∇T µl
= −fl β (4.7)
∂x Ṫ K

As the solidification advances, the melt pressure starts falling from an initial value
Pliq down to a critical value Pcr at which porosity begins to form. The initial value of
the pressure Pliq is simply determined by the hydro-static pressure due to the head of
liquid metal h. In gravity casting:

Pliq = Patm + ρgh (4.8)

In this term we can include the packing pressure for high pressure simulations.
This pressure is made by the piston when the casting is already solidifying in order
to make the void areas collapse and reduce the porosity. We can simply include it by
using:

Pliq = Ppacking + ρgh (4.9)

Where ρ is the density of the liquid, h is the head of liquid and g is the gravity.
This pressure can be determined using a mechanical equilibrium. In the inside we
have the pore pressure Pp and in the outside the critical pressure Pp . We can also take
into account the capillary pressure Pσ which also tries to close the pore. Therefore,
the equilibrium can be written as:

Pcr + Pσ = Pp (4.10)

The capillary pressure is usually described as:

2σ
Pσ = (4.11)
r0

Where σ is the surface tension between the liquid and the pore and r0 is the initial
radius of curvature of the pore. If the presence of gasses in the melt is negligible the
pressure inside the pore will be null and the critical pressure will simply be:

62
 4 MICROPOROSITY MODEL

2σ
Pcr = − (4.12)
r0

Notice that the value that we assign to the capillary pressure is rather arbitrary
as the initial radius of the pore will depend on many factors and cannot be known
in general. It is sometimes evaluated as r0 = λ2 /4. In more rough estimations, the
whole term is neglected and they simply use Pcr = 0
With this idea in mind we can now integrate the equation 4.7 from the beginning of
the solidification when fl = 0 and P = Pliq until porosity forms when fl = fl,cr and
P = Pcr . We define as well 4Pcr = Pliq − Pcr

Z Pliq Z 0 Z T liq Z 1
µl β Ṫ fl µl β Ṫ fl µl β Ṫ f l dṪ
dP = 4Pcr = dx = dT = dfl
Pcr xcr K∇T Tcr K∇T (∇T )2 fl,cr K dfl
(4.13)

In order to have a non-dimensional formula, a dimensionless temperature is intro-

duced θ = (T − Tsol ) /4Tf , where Tsol is the solidus temperature and 4Tf is the alloy
freezing range. We can define the integral value now as

1
(1 − fl )2 dθ
Z
I(f l, cr) = 180 dfl (4.14)
f l,cr fl2 dfl

63
 4 MICROPOROSITY MODEL

Figure 4.3: Mushy Zone 1D Analysis from [5]

Putting altogether we arrive to

µl β4Tf Ṫ
4Pcr = I(f l, cr) (4.15)
λ22 (∇T )2

We are going to define here the dimensionless Niyama

√
∗ ∇T λ2 4Pcr p
Ny = q = I(f l, cr) (4.16)
µl β4Tf Ṫ

The dimensionless Niyama criterion given by 4.16 accounts not only for the local
thermal conditions, but also for the properties and solidification characteristics of the
alloy and the critical pressure drop across the mushy zone. If we determine the SDAS
as a function of the cooling rate from the relation λ2 = Cλ T −1/3˙ the dimensionless
Niyama can be expressed in the following alternate form

64
 4 MICROPOROSITY MODEL

s
∇T 4Pcr
Ny∗ = Cλ (4.17)
Ṫ 5/6 µl β4Tf

Notice that in this equation, unlike the usual Niyama Criterion which is propor-
tional to ∇T /Ṫ 1/2 , the dimensionless Niyama criterion is proportional to ∇T /Ṫ 5/6
which according to the authors will result in a more generally applicable criterion for
varying section thicknesses and mold materials.

The approach here is to determine the liquid fraction that is still present in a point
when the pressure reaches the critical value. From this point, feeding is assumed to
be impossible and the amount of porosity will be calculated applying the conservation
of mass.

∂ρ
+ ∇ · (ρv) = 0 (4.18)
∂t

Where ρ is the mixture density, defined as ρ = fl ρl + fp ρp + fs ρs and each subscript

s,l and p accounts for the solid, liquid and porosity contribution respectively. Making
use of the fact that 1 = fl + fs + fp and neglecting the density of the porosity we can
define the density as:

ρ = ρs − fp ρs + fl (ρl − ρs ) (4.19)

If there is no feeding ∇ · (ρv) = 0 and therefore the equation 4.18 is simplified to:

∂ (ρs − fp ρs + fl (ρl − ρs ))
=0 (4.20)
∂t

∂fp (ρl − ρs ) ∂fl

= (4.21)
∂t ρs ∂t

Integrating from the point where porosity begins to form, that is when the pressure
reaches the critical value Pcr , fl = fl,cr and fp = 0 until the end of the solidification
when fp = fpend and fl = 0.

65
 4 MICROPOROSITY MODEL

(ρs − ρl )
fpend = fl,cr (4.22)
ρs

Recalling that β = (ρs − ρl ) /ρl we end with:

β
fpend = fl,cr (4.23)
β+1

The algorithm to solve the problem is the following for each point:

1. Compute the local thermal conditions and obtain the properties of the material.

2. Calculate the value of the Dimensionless Niyama using equation 4.16 or 4.17

3. Compute the critical fluid fraction from equation 4.14 knowing that from equa-
p
tion 4.16. Ny∗ = I(f l, cr)

4. Compute the porosity using the mass continuity equation.

There is another aspect that we have not mentioned yet and might have a great
relevance. We refer to the moment of the solidification at which we compute the
gradient of the temperature and the cooling rate. In the paper, Beckermann and
Carlson suggest to compute these values when the temperature is a ten per cent of
the freezing range above the solidus temperature, this is

TN y = Tsol + 0.14Tf (4.24)

They based this decision on the assumption that porosity forms late on the so-
lidification, however there are other authors that claim that porosity could nucleate
and grow for values of the temperature much higher than this one. In the following
sections we will check the importance of this factor in different simulations.

4.3 Study Cases

In this section we are going to show some results for microporosity using the method
previously exposed using different parts and materials. All the simulations were made

66
 4 MICROPOROSITY MODEL

using Click2Cast. The results are shown as void fraction volume using Hyperview,
a post-processing software from Altair Engineering. For all the results, the variables
were evaluated at TN y = Tsol + 0.34Tf , the packing pressure was set to Ppacking = 0
and the secondary dendrite arm spacing was SDAS = 40µm. To recover the gradients,
the iterative L2 Projection method proposed in section 3.3 was used with 5 iterations.
The filling was not computed for any simulation.

4.3.1 Lever

In the first case, the simulation corresponds to the lever of a motorbike. The material
used for the casting was Aluminum AlSi7Mg with a temperature at the beginning
of the simulation Tinitial = 700o C. The mold material was Steel-H13 with an initial
temperature of 150 C.

The solidification was simulated using 1,185,840 elements.

Figure 4.4: Lever Part

67
 4 MICROPOROSITY MODEL

Figure 4.5: Lever - Regions over 2% of Microporosity

Figure 4.6: Levers - Midplane

As we can see in the figures above most of the porosity is located at the center of
the casting part with a maximum value of 3.29% of porosity. It is also present with
lower intensity in some areas that have small thicknesses in both the extremes of the
casting.

4.3.2 Brake Caliper

The second study case correspond to a brake caliper made of Magnesium AE42 with
a temperature at the beginning of the simulation Tinitial = 727o C. The mold material
was Steel-H13 with an initial temperature of 150 C.

68
 4 MICROPOROSITY MODEL

The solidification was simulated using 1,425,004 elements.

Figure 4.7: Brake Caliper Part

Figure 4.8: Brake Caliper - Regions with over 0.9% Microporosity

69
 4 MICROPOROSITY MODEL

Figure 4.9: Brake Caliper - Section Cut

In this casting the microporosity is distributed on the more massive areas at the
top and the bottom as well as in the thin walls of the center cylinder. In this case,
the maximum porosity value is 3.33%.

4.3.3 Oil Pump

The last case that we will show in this section corresponds to a oil pump made of
Carbon-Steel ASTM-A216. The initial temperature of the liquid was set to Tinitial =
1592C. Same as in the other cases, the mold material was Steel-H13 with an initial
temperature of 150 C.

For this casting, a mesh of 1,183,087 elements was created.

70
 4 MICROPOROSITY MODEL

Figure 4.10: Oil Pump Part

Figure 4.11: Oil Pump - Regions with over 0.5% Microporosity

71
 4 MICROPOROSITY MODEL

Figure 4.12: Oil Pump - Brake Caliper Section

Microporosity in this part is mainly situated around the two cylinders at a medium
height that coincides with the last solidifying regions, where the maximum microp-
orosity is present with a value of 2.017%. As well as in the other cases, microporosity
also appear in narrow areas of the casting.

4.3.4 Comments on the results

As we have seen through all the cases, using the Dimensionless Niyama Method pro-
posed in [5], microporosity appears in two types of areas.

• Center of massive areas: We can expect microporosity to appear at the center of

massive areas, or in general, wherever two solidifying fronts meet each other (See
for example figure 4.6) As these areas present some degree of symmetry in the
temperatures, the gradient tends to be very low. This implies a low N y ∗ value
(see equation 4.16) and therefore a high value of microporosity. These results
require some interpretation, as we have to realize that in some of these areas
large voids of macroporosity might be present and therefore no microporosity
will grow there.

• Thin walls: In general we can also porosity to appear in thin wall as for example
the front part of the cylinder in figure 4.9. Here two effects are combined: first,

72
 4 MICROPOROSITY MODEL

the gradient is low due to similar reasons as explained above. Secondly and
most important, these zones will solidify very fast due to their small thickness
and therefore the cooling rate will be low, which will lead to a low N y ∗ and high
microporosity values.

4.4 Gradient Recovery

We have already mentioned several times that the key aspect of the Dimensionless
Niyama method is the gradient of the temperature. For this reason, different methods
to recover the gradient were implemented and tested in section 3. As it was justified
in subsection 3.3, the Iterative L2 Projection method was chosen to be implemented
in Click2Cast. In this section, we are going study the influence of the number of
iterations to the minimum value of the Dimensionless Niyama calculated in a casting
part and the corresponding value of the microporosity.

The chosen casting to do this test is one of the example parts offered by Click2Cast
which corresponds to the lever of a motorbike previously shown in figure 4.4. The
casting material is aluminum AlSi7Mg and the mold is made by Steel H-13.

Figure 4.13: Relative error for Ny* and Microporosity

73
 4 MICROPOROSITY MODEL

Figure 4.14: Microporosity Calculation - Total Gradient Recover Computational Time

In figure 4.13 we show the relative error committed by the iterative scheme with re-
spect to the consistent solution. We can see that in a relative low number of iterations,
the relative error falls quickly. If we use 5 iterations we can see that the corresponding
relative error will be around 2% with a low computational cost in terms of time as
shown in figure 4.14

4.5 Sensitivity to the element size

In this first sensitivity study, we are going to see how the element size can affect the
values of N y ∗ and the microporosity. Again, we used the Lever Part shown in 4.4 with
aluminum AlSi7Mg. The mold material was Steel H-13. Other simulation parameters
were:

• Packing Pressure = 0

• SDAS = 40µm

• Temperature of Evaluation = Tsol + 0.34Tf

74
 4 MICROPOROSITY MODEL

Element size Number of elements

1.5 53841
1 164375
0.8 310145
0.5 1185840
0.4 2250762

Table 4.1: Meshes used to study sensibility to element size

Element size N y ∗ Microporosity %

1.5 6.332 2.023
1 1.664 3.034
0.8 2.31 2.855
0.5 0.804 3.260
0.4 0.689 3.293

Table 4.2: Influence of mesh size to N y ∗ and microporosity

(a)

(b)

Figure 4.15: (a) Medium section with element size h = 0.5 (b) Medium Section with
element size h = 0.8

75
 4 MICROPOROSITY MODEL

The tables above and figure 4.15 show us that although the values of N y ∗ and
microporosity converge, a pretty fine mesh is required in order to have an accurate
result. This is specially true for complex geometries where the gradient reconstruction
might be more problematic. Notice that in usual finite elements the H1 norm of the
error, linked to the error of the gradients, is only of first order with respect to the
element size (eH1 = C1 h) whereas the L2 norm, which corresponds to the error of the
problem variable itself is of second order (eL2 = C2 h2 ) .

4.6 Sensitivity to different evaluation temperatures

One of the most crucial aspects in the evaluation of Niyama or like in our case,
Dimensionless Niyama, is the point at which the variables are evaluated. As shown in
[16], the election of this parameter has a great impact in the final results and therefore
each commercial software suggests to use a specific value. In this sensibility study we
are going to check how much this parameter affects to the final results.

We are going to use the same conditions for the simulation that we stated before
for the Lever Part shown in 4.4 made of aluminum AlSi7Mg. The mold material was
Steel H-13.

• Packing Pressure = 0

• SDAS = 40µm

• Mesh size = 1185840 elements

Evaluation Temperature N y ∗ Microporosity %

Tsol + 0.14Tf 4.32 2.38
Tsol + 0.24Tf 11.42 0.75
Tsol + 0.34Tf 0.80 3.26
Tsol + 0.44Tf 1.24 3.15

Table 4.3: Influence of evaluation temperature to N y ∗ and microporosity

As it can be seen in the table above and as stated in [16], the value of the evaluation
temperature will have a great influence on the final results. This is caused by the
evolution of the temperature gradient as shown in the graphic below.

76
 4 MICROPOROSITY MODEL

Figure 4.16: Evolution of the temperature gradient - Red: Exterior node - Blue:
Interior Node

In this graphic we see the evolution of the gradient of the temperature for an
interior and an exterior node. It is clear that though the values in the interior are
always lower, the point at which we take the value will have a big influence. Peaks are
present due to the shape of the temperature-solid fraction curves, which are usually
not linear, and the effect of the latent heat during the solidification.

Although Beckermann et al. suggested in [5] to use Teval = Tsol + 0.14Tf as

evaluation point, we have obtained more realistic results using Teval = Tsol + 0.34Tf .
This is no surprise as each commercial code uses different formulations and parameters
and therefore different results will be obtained.

4.7 Casting parameters that have influence on microporosity

In this subsection we are going to see how different casting parameters affect the final
amount of microporosity. This value is going to be calculated using the nodal values
in the following way:

P
nodes Vi fp,i
M icroporosity(%) = (4.25)
VT

Where Vi is the nodal volume, fp,i is the nodal value of the microporosity and VT

77
 4 MICROPOROSITY MODEL

is the total volume of the part.

Heat Transfer Coefficient (HTC)

Figure 4.17: Microporosity - HTC Influence

The heat transfer coefficient, usually known as HTC is the relation between the
heat flux and the temperature difference between two points:

q
HT C W/m2o C =


(4.26)
4T

It basically gives us an idea of how fast heat is extracted. It depends on the

material of both the casting and the mold and the gap that might form as solidification
advances. This graphic clearly shows us than low HTC values will lead to a higher
amount of microporosity. This is caused by lower temperature gradients that will
make the flow of liquid metal through the mushy zone more difficult. For this reason,
we can expect in general to have lower levels of microporosity in steel molds than in
sand molds.

78
 4 MICROPOROSITY MODEL

Packing Pressure

Figure 4.18: Microporosity - Packing Pressure Influence

In many occasions, the metal is forced to be introduced in the cavity of the mold
by a piston instead if letting it flow simply by gravity. When the filling has concluded,
the piston keeps pressuring the metal, in a process known as packing pressure, in
order to ensure that metal does not flow back outside, but also to reduce the level of
porosity as the pressured metal will try to close the empty spaces. The effect of this
pressure can be introduced in our model and as we expected, it helps to reduce the
amount of microporosity.

Feeding Temperature

Figure 4.19: Microporosity - Feeding Temperature Influence

79
 4 MICROPOROSITY MODEL

The feeding temperature is an important factor in the casting since it must be en-
sure that the metal does not solidify anywhere during the filling, which would provoke
defects like misruns as explained in section 1. In counterpart, as we see in the graphic,
high feeding temperatures will lead to slightly higher levels of microporosty, though
its effect is very low.

Mold Temperature

Figure 4.20: Microporosity - Mold Temperature Influence

In this last graphic we can see how higher mold temperatures have a consider-
able negative effect on microporosity. Again, this is due to the presence of lower
temperature gradients that as we stated is the key aspect in microporosity.

All these results agree with [22], where it was experimentally shown an increase
of the mold temperature would be followed by an increment in the total amount of
porosity in the casting part. Moreover, in the same work, Y.M. Lee proved the little
influence of the temperature of the liquid feeding. This conclusion also matches our
results.

4.8 Conclusions

The Dimensionless Niyama and the original Niyama Criterion have been object of
many debates within the academic world. It cannot be denied its importance as
practically all casting simulation packages offer at least the Niyama values and some

80
 4 MICROPOROSITY MODEL

of them based their microporosity prediction on it. However one must be aware of the
important limitations that this approach has and the need of some interpretation by
the user. In this last section we will try to put together some of the conclusions that
can be found along the literature about Niyama and Dimensionless Niyama together
with our own conclusions from the results that we have obtained.

First thing that we have to take into account is that the nucleation and growth
of pores is a very complex problem and many factors have to be taken into account.
As stated in section 2.2 and [30] the following aspects should be included in an ideal
model.

• Thermal Field

• Solid phases growth

• Flow Field

• Interfacial characteristics between the gas and liquid or solid phases.

The Niyama Criterion and in general most of the criterion functions only computes
the thermal field and neglect the importance of the rest of the fields. The Darcy’s
Law based assumption for the flow does not cover the whole physics behind the defects
formation. For example in clean melts, porosity only appear because of the segregation
of gases which is completely neglected in Niyama.

In spite of being true that the velocities in the solidification phase are small,
they are going to have a key role since the liquid will flow from the interior areas to
those where the solidification has already begun. Besides, though some researchers
have approved its application for simple steel casting tests [7, 3] in general its use is
not advised if working with complex geometries as it does not take into account the
geometry of the part or other types of alloys which might have very different freezing
ranges [31, 30].

One could wander how this criterion has acquired so much notoriety with the
limitations mentioned above in sight. We have to realize first that it is very difficult to
validate microporosity results as most of the radio graphic techniques that are usually
used to measure porosity fail to detect microporosity as it belongs to a very small size
scale. Secondary, as its simplicity might make it questionable for general applications,
it is also a reason why it became so popular since results can be obtained very quickly

81
 4 MICROPOROSITY MODEL

and with little implementation effort. This was specially important at the moment
when it was developed by Niyama et al. in the early 80’s where computers were much
less powerful and complex simulations could not be run as easily as nowadays.

Beckermann and Carlson created the Dimensionless Niyama Criterion with the
aim to address some of the doubts that have been formulated by other authors. They
tried to come with a solution where other factors could be taken into account. Besides
the local thermal variables (the cooling rate and the gradient of the temperature),
their formulation took into account also material parameters in order to be able to
use it with materials with different freezing ranges. Other of the factors that allows
us to have a better control of the simulation is the critical pressure, which defines
a threshold under which porosity will appear. Although this pressure is sometimes
simply taken as the atmospheric pressure, a more precise approximation can be taken
including the effect of the gravity and the capillary pressure. As we have shown in
subsection 4.7 this model is able to deal with different parameters that will affect to
the problem in a realistic way.

We have seen in 4.3 that results require some degree of interpretation. Dimension-
less Niyama and general Niyama criterion also mark areas where macroporosity will
be present. These results do not have to be considered as stated in [18] and for that
reason, another model to compute macroporosity is required.

In this study we have shown some calculations for some example castings were
microporosity seemed to give reasonable qualitative results. Regarding to the quan-
titative aspect, many variables such as SDAS or the evaluation temperature might
require some tuning due to its high influence in the final result. Beckermann and
Carlson show some experimentally agreement in their paper, but only for very simple
geometries. It is actually very hard to quantify even experimentally the actual val-
ues of microporosity, so it can be quite difficult to validate the model in this aspect.
Future work must be focused on this direction.

82
 4 MICROPOROSITY MODEL

83
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

5 Macroporosity and Pipe Shrinkage Model

5.1 Introduction

As some parts of the casting begin to solidify, a mass deficit is created as a consequence
of the difference densities between the solid and liquid metal. When this mass feeding is
cut, shrinkage porosity begins to form. It has been previously mentioned that when the
cause of this cut is the growth of dendrites, this kind of porosity is called microporosity.
When this effect happens in a larger scale, with isolated pools of liquid, porosity can
be seen the naked eye and is referred as macroporosity. When the isolated pool of
liquid is open to the air, porosity is usually called open shrinkage or pipe shrinkage
because of its characteristic shape. Due to its impact on mechanical properties such
as ductility, dynamic properties or fatigue life, macroporosity is considered a major
cause of casting rejection. In many occasions, if the macro-shrinkage is not limited to
the riser the casting is not accepted.

In this section, a pressure-based model will be presented capable to compute macro-

porosity and pipe shrinkage. Following other approaches as the ones introduced in
[8, 6, 27, 2], pressure is calculated using Darcy’s Law model along the whole casting
and shrinkage is assumed to form in those points where pressure falls below a certain
value.

As this model is still being implemented, no numerical results or simulations are

shown, which will be reserved for future work.

5.2 Governing Equations

The present model states that each element is composed of a combination of air(a),
porosity(p), liquid metal (l) and solid metal(s), such that the sum of the volume
fractions is equal to 1. The air fraction a corresponds to the air that enters through
the top of the riser and forms the open shrinkage.

a + p + l + s = 1 (5.1)

The mixture properties are obtained as a function of the properties of each phase

84
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

multiplied by the corresponding volume fraction.

ρ = a ρa + p ρp + l ρl + s ρs (5.2)

Energy Conservation

The energy conservation equation is written as:

 
ds ∂T
ρc − ρL = ∇ · (λ∇T ) (5.3)
dT ∂t

Where ρ is the mixture density, c is the specific heat capacity, T is the temperature
and λ is the conductivity. With this equation we can compute the temperature and
the solid fraction s , as we suppose that the solid fraction is a known function of the
temperature.

Mass Conservation

For the mass conservation we are only going to calculate the liquid phase, assuming
that the porosity and the solid phase are stationary. Subtracting the air phase equation
(∂(a ρa )/∂t + ∇ · (a ρa va ) = 0) to the continuity equation we get

∂
(p ρp + l ρl + s ρs ) + ∇ · (ρl v) = 0 (5.4)
∂t

where v is the superficial velocity v = l vl

The last term can be rewritten as

∇ · (ρl vs ) = ρl ∇ · v + v · ∇ρl (5.5)

where the second term is very small and can be neglected. Introducing this modi-
fication into the previous equation and reordering terms

85
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

 
1 ∂
∇·v = [p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl ] (5.6)
ρl ∂t

Momentum Conservation

The momentum equation corresponds to the Stokes’ equation where an additional

Darcy term has been added. As the velocity of the fluid and therefore the Re number
are very small, the inertial term can be neglected as well as the convective term.

l l l
∇2 v = v + ∇P − ρref g (5.7)
K µl µl

Where g is the gravity, ρref is the density of the liquid at the liquidus temperature,
µl is the dynamic viscosity and K is the permeability given by:

3l
K = Kr Ko (5.8)
(1 − l )2

with

l
Kr = (5.9)
l + a

Ko = 6.10−4 λ21 (5.10)

being λ1 the primary dendrite arm.

Notice that the momentum equation reduces to the Stokes equation when the
permeability tends to infinite, that is, when liquid is the only present phase. Likewise,
when the liquid fraction is very small, the left hand side term becomes very small and
equation reduces to Darcy’s Law.

86
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

In order to have a well posed problem we have to include the following boundary
condition

n·v =0 on ∂Ω (5.11)

where ∂Ω is the boundary of the domain Ω and n is the normal to the boundary.

Pressure Equation

Taking the divergence of the momentum equation and introducing the mass equa-
tion we get

     
K 1 ∂ K
∇· ∇P = [p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl ] + ∇ · ρref g
µl ρl ∂t µl
(5.12)
 
K 2
+∇· ∇ vs
l

The last term is very small and can be neglected arriving to

     
K 1 ∂ K
∇· ∇P = [p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl ] +∇· ρref g (5.13)
µl ρl ∂t µl

When the solidification begins we suppose p = a = 0. It will be explained

later how to calculate these variables as the solidification advances. When the liquid
fraction is either 0 or 1, the permeability K goes to 0 and infinite respectively. In
order to avoid this problem, when we calculate the permeability using equation 5.8
the liquid fraction can only take values from 0.01 to 0.99.

With these assumptions we have only one equation with one unknown, P , which
can be solved using proper boundary conditions.

n · ∇P = n · ρref g on ∂Ω (5.14)

87
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

The last equation can be obtained from substituting the boundary condition for
the velocity n · v = 0 on ∂Ω in the momentum equation multiplied by the normal n,
arriving to n · ∇P = n · ρref g if the left hand side term is neglected.

5.3 Shrinkage Porosity

As we mentioned before, we can compute the pressure on the whole domain using
equation 5.13. As the volume contracts due to the difference of densities between the
molten and the solid metal, pressure will begin to drop. When it reaches a certain
critical value Pp , the pressure in the melt will be the same that the pressure in the
pore so it will be able to stabilize and grow. When this happens, the pressure at
this point will be fixed for the rest of the solidification according to P = Pp . From
this point all the volume changes will contribute toward the development of shrinkage
porosity p from equation 5.13.

One of the keys of this model is the value that we assign to the critical value Pp .
As a first approximation we can state as Pp = 0 if there is no gas porosity. A more
complete model will be explained below where the presence of gas in the melt is taken
into account.

Gas Porosity

Gas porosity can be included into our model considering the following species
conservation equation. We consider that only one gas is present on the melt.

s ρs Cs + l ρl Cl + p ρp Cp = ρl C0 (5.15)

Where C0 is the initial concentration of gas in the melt. Cl ,Cs and Cp are the gas
concentrations in the liquid, solid and porosity in weight fraction respectively . We
can state that Cs = KCl , where K is the partition coefficient and Cp = 1.

Once we have calculated the value of Cl at a certain time, we can calculate the
pressure in the pore Pp , using Sievert’s Law.

88
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

 2
100Cl f
Pp = Patm (5.16)
Ke

In this equation, f = f (T ) is the activity, which will be considered constant f = 1,

and Ke = Ke (T ) is the equilibrium coefficient which has to be derived for each gas.
For nitrogen we have for example:

a
log(Ke ) = − −b (5.17)
T

Being a = 188.1K and b = 1.246.

5.4 Open Shrinkage

In those nodes at the top of the riser that are emptying of liquid, the pressure is
forced to atmospheric pressure Patm . For these nodes we assume that p = 0, that is,
no porosity forms. Instead, as the pressure is fixed, we can solve for the air fraction
a present on equation 5.13. Notice that fixing the pressure at the top of the riser
acts like a Dirichlet Boundary Condition that does not allow the pressure to float
arbitrarily. As the node empties of liquid, the pressure will have to be fixed in the
node which is directly below as it begins to form open shrinkage as well.

5.5 Weak Form

In order to be able to implement the system of equations proposed above into the
finite elements code Kratos, the weak form of the problem must be obtained first.
The line here is based on Cornthwaite’s PhD Thesis [9].

We start with the pressure equation with its boundary conditions.

89
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

     
K 1 ∂ K
∇· ∇P = [p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl ] +∇· ρref g (5.18)
µl ρl ∂t µl

n · ∇P = n · ρref g on ∂Ω (5.19)

Multiplying the equations by a smooth test function φ and integrating over the
domain Ω we get the weak form of the problem
       
K 1 ∂
∇· ∇P ,φ Ω = [p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl ] , φ Ω
µl ρl ∂t
(5.20)
   
K
+ ∇· ρref g , φ Ω
µl

(n · ∇P, φ) ∂Ω = (n · ρref g, φ) ∂Ω (5.21)

We integrate by parts the left hand side term

          
K K K
∇· ∇P , φ Ω = n · ∇P , φ ∂Ω − ∇P , ∇φ Ω (5.22)
µl µl µl

and the last term on the right hand side

          
K K K
∇· ρref g , φ Ω = n · ρref g , φ ∂Ω − ρref g , ∇φ Ω (5.23)
µl µl µl

Substituting into equation 5.20

90
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

            
K K 1 ∂ρ K
n· ∇P , φ ∂Ω − ∇P , ∇φ Ω = ,φ Ω + n · ρref g , φ ∂Ω
µl µl ρl ∂t µl
(5.24)
  
K
− ρref g , ∇φ Ω
µl

Where p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl has been replaced by ρ.

Now we are going to take the equation 5.21 and multiply by K/µl

   
K K
n · ∇P, φ ∂Ω = n · ρref g, φ ∂Ω (5.25)
µl µl

If this Boundary Equation is introduced inside 5.24 , and reintroducing ρ = p (ρp −

ρl ) + s (ρs − ρl ) − a ρl + ρl we arrive to

      
K 1 ∂
∇P , ∇φ Ω + [p (ρp − ρl ) + s (ρs − ρl ) − a ρl + ρl ] , φ Ω =
µl ρl ∂t
(5.26)
  
K
ρref g , ∇φ Ω
µl

5.6 Resolution Algorithm

Equation 5.26 can be used to write an element system after discretizing the middle
term in time.

 
h i P
L Kp Ka p  = RHS (5.27)
 

a

For each element we have 1 equation and 3 unknown fields: P ,p and a . The
key of this method is that for each node we are going to calculate only one of these
variables:

91
 5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

• If the pressure in the previous step was higher than the critical pressure Pp ,
there is no porosity and therefore p = a = 0 and we calculate P.

• If the pressure in the previous step was lower or equal than Pp , we fix the pressure
to P = Pp and a = 0 and compute p .

• If the node is at the top of the riser we fix p = 0 and P = Patm and calculate
a .

92
5 MACROPOROSITY AND PIPE SHRINKAGE MODEL

93
 6 CONCLUSIONS AND FUTURE WORK

6 Conclusions and Future Work

The objective of this work was the implementation of a model to compute porosity in
order to implement it in the commercial software Click2Cast. In this last section we
are going to summarize the different tasks that were carried out and the conclusions
that we can extract from each of them as well as the current and future work that is
being done in each specific area.

In section 2 we did a review of the literature available on porosity and particularly,

the different models that exist to compute each specific type of porosity. We classified
them as follows

• Microporosity Models : Criterion Functions

• Macroporosity Models : Thermal/Volume Calculation Models

• Coupled microporosity and macroporosity Models : Thermal/Fluid Flow Models

and Stochastic Models

Following this classification, we developed a Microporosity model based on Becker-

mann and Carlson’s paper from 2008 [5]. In order to do this we first had to implement
a method to recover the gradient of the temperature in an accurate and efficient way.
In section 3 we explained and tested 5 different methods to recover gradients:

• L2 Lumped Projection

• L2 Consistent Projection

• Iterative L2 Projection

• Zhang Method

• Pouliot Method

For each method, we obtained results for the error and the calculation time in order
to be able to compare them. Although Pouliot method was proved to be the most
accurate, we finally decided to go along with the iterative method which demonstrated
its capacity to calculate gradients much faster than Pouliot with a reasonable accuracy.

94
 6 CONCLUSIONS AND FUTURE WORK

The microporosity model was explained in detail in section 4 along with results
for different parts and materials. By doing a sensibility analysis of the mesh size,
we showed that in order to have precise results a fine mesh is needed, in general
finer than the mesh that we would require if only the solidification simulation was
taken into account. Another sensibility analysis of the evaluation temperature also
warned us about the importance of this parameter in the simulation. The tuning of
this and other parameters such as the Secondary Dendrite Arm Spacing will be an
important part of the future work. It would be desirable to have the chance to validate
this model with some experimental results, although we know that due to the small
size of these defects they cannot be detected by usual radio-graphic procedures and
especial techniques are required. Once this validation work is finished, the current
implementation will be included in future versions of Click2Cast. Finally, in the last
part we studied how different parameters such as the mold temperature or the heat
transfer coefficient affected the final results with satisfactory results that matched
experimental findings. [22]

In section 5, we presented a macroporosity model based on a simplified flow model,

following the same approach as other methods like [6, 8, 19]. In this model, only the
pressure was calculated throughout the solidification. When pressure falls under a
certain value, pressure is no longer calculated and porosity is assumed to grow. This
model is currently implemented but some work is still required to correct some wrong
behavior under certain circumstances. We expect to solve this problems and validate
the results in the next months in order to implement it along with the microporosity
model in Click2Cast.

95
 REFERENCES

References

[1] J. Beech, M. Barkhudarov, K. Chang, and s. B. Chin. Modeling of casting welding

and advanced solidification processes. B.G Thomas and C. Beckermann, eds, The
Minerals, Metals and Materials Soc. p.1071, 1998.

[2] S. Bounds, G. Moran, K. Pericleous, M. Cross, and T.N. Croft. A computa-

tional model for defect prediction in shape castings based on the interaction of
free surface flow, heat transfer, and solidification phenomena. Metallurgical and
Materials Transactions B January 2000, 2000.

[3] Kent D. Carlson and Christoph Beckermann. Use of the niyama criterion to
predict shrinkage-related leaks in high-nickel steel and nickel-based alloy castings.
Proceedings of the 62nd SFSA Technical and Operating Conference Paper No. 5.6
SFSA, Chicago, IL, 2008.

[4] Kent D Carlson and Christoph Beckermann. Authors reply to discussion of pre-
diction of shrinkage pore volume fraction using a dimensionless niyama criterion.
Metallurgical and Materials TransactionsA, 40(13):3054 3055, 2009.

[5] Kent D. Carlson and Christoph Beckermann. Prediction of shrinkage pore volume
fraction using a dimensionless niyama criterion. Metall. Mater. Trans. A, Vol.
40A, pp. 163 175, 2009.

[6] Kent D. Carlson, Zhiping Lin, Richard A. Hardin, and Christoph Beckermann.
Modelling of porosity formation and feeding flow in steel casting. Proceedings of
the 56th SFSA Technical and Operating Conference Paper No. 4.4, Chicago IL,
2002.

[7] Kent D. Carlson, Shouzhu Ou, Richard Hardin, and Christoph Beckermann. De-
velopment of a methodology to predict and prevent leaks caused by microporosity
in steel castings. Proceeding of the 55th technical and operating conference, SFSA,
Chicago, 2001.

[8] A.V. Catalina and C.A. Monroe. Simplified pressure model for quantitative
shrinkage porosity prediction in steel castings. Materials Science and Engineering
33 (2012) 012067, 2012.

96
 REFERENCES

[9] John. P. Cornthwaite. Pressure Poisson Method for the Incompressible Navier-
Stokes Equations Using Galerkin Finite Elements. PhD thesis, Georgia Southern
University, 2003.

[10] Pengfei Du. Numerical Modeling of Porosity and Macrosegregation in Continuous

Casting of Steel. PhD thesis, University of Iowa, 2013.

[11] P.N. Hansen and P.R. Sahm. How to model and simulate the feeding process
in casting to predict shrinkage and porosity formation. Modeling of Casting and
Welding Process IV, pages 33 42. TMS AIME, 1988.

[12] R.A. Hardin, S. Ou, K. Carlson, and C. Beckermann. Relationship between

casting simulation and radiographic testing: Results from the sfsa plate casting
trials. 1999 SFSA Technical and Operating Conference, 1999.

[13] D. M. Hawken, P. Townsend, and M. F. Webster. A comparison of gradient recov-

ery methods in finite-element calculations. Communications in Applied Numerical
Methods, VOl. 7, 195 204, 1991.

[14] I. Imafuku and K. Chijiiwa. A mathematical model for shrinkage cavity prediction
in steel castings. AFS Transactions 91: 527 540, 1983.

[15] Ute Israel. Implementation of an algorithm for data transfer on the fluid-structure
interface between non-matching meshes in kratos and an algorithm for the gen-
eration of an interface between gid and kratos. Master’s thesis, Technische Uni-
versität München, 2006.

[16] Neelesh Jain, Kent D. Carlson, and Christoph Beckermann. Round robin study to
assess variations in casting simulation niyama criterion predictions. Proceedings of
the 61st SFSA Technical and Operating Conference Paper No. 4.4 SFSA, Chicago,
IL, 2007.

[17] J Jakumeit, S Jana, B Bötger, R Laqua, M Y Jouani, and A BÃŒhrig Po-

laczek. Simulation-based prediction of micro-shrinkage porosity in aluminum
casting: Fully-coupled numerical calculation vs. criteria functions. IOP Confer-
ence Series Materials Science and Engineering 27(1):2066, 2012.

[18] Maodong Kang, Haiyan Gao, Jun Wang, Lishibao Ling, and Baode Sun. Predic-
tion of microporosity in complex thin-wall castings with the dimensionless niyama
criterion. Proceedings of the 8th Pacific Rim International Congress on Advanced
Materials and Processing., 2013.

97
 REFERENCES

[19] Kimio Kubo and Robert D.Pehlke. Mathematical modeling of porosity formation
in solidification. Metallurgical Transactions B, June 1985, Volume 16, Issue 2,
pp 359 366, 1985.

[20] P. D. Lee, A. Chirazi, and D. See. Modeling microporosity in aluminum-silicon

alloys: a review. Journal of Light Metals 1 (2001) 15 30, 2001.

[21] Y.W. Lee, E. Chang, and C.F. Chieu. Met. trans. b,. 1990, 21b, 715, 1990.

[22] Y.M. Li and R.D. Li. Effect of the casting process variables on microporosity
and mechanical properties in an investment cast aluminium alloy. Science and
Technology of Advanced Materials 2, 2001.

[23] E. Niyama, T. Uchida, M. Morikawa, and S. Saito. Predicting shrinkage in large

steel castings from temperature gradient calculations. AFS Int. Cast Metals Jour-
nal, 6(2):16 22,, 1981.

[24] E. Niyama, T. Uchida, M. Morikawa, and S. Saito. A method of shrinkage pre-

diction and its application to steel casting practice. AFS Cast Metals Research
J., 1982, 7, 52, 1982.

[25] Zienkiewicz OC and Zhu JZ. A simple error estimator and adaptive procedure
for practical engineering analysis. International Journal for Numerical Methods
in Engineering 1987; 24:337357., 1987.

[26] W.S. Pellini. Factors which determine riser adequacy and feeding range. AFS
Transactions, 61: 61 81, 1953.

[27] Ch. Pequet, M.Gremaud, and M. Rappaz. Modeling of microporosity, macrop-

orosity, and pipe-shrinkage formation during the solidification of alloys using a
mushy-zone refinement method: Applications to aluminum alloys. Metallurgical
and Materials Transactions A, July 2002, Volume 33, Issue 7, pp 2095 2106,
2002.

[28] B. Pouliot, M. Fortin, A. Fortin, and É. Chamberland. On a new edge-based

gradient recovery technique. Int. J. Numer. Meth. Engng (2012), 2012.

[29] G. K. Sigworth. Discussion of prediction of shrinkage pore volume fraction using

a dimensionless niyama criterion. Metallurgical and Materials Transactions A,
40(13):3051 3053, 2009.

98
 REFERENCES

[30] Doru M. Stefanescu. Computer simulation of shrinkage-related defects in castings

- a review. Int. J. Cast. Met. Res. 18 129 143, 2005.

[31] R. Tavakoli. On the prediction of shrinkage defects by thermal criterion functions.

nt J Adv Manuf Technol (2014) 74: 569, 2009.

[32] Zhimin Zhang and Ahmed Naga. A new finite element gradient recovery method:
Superconvergence property. SIAM Journal on Scientific Computing 26(4):1192
1213, 2005.

99