DESIGN OF FRICTIONAL MACHINE ELEMENTS

(ME3202)

Module 7: Optimization in Design

Lesson 1: Introduction

B. Tech., Mech. Engg., 6th Sem, 2022-23

Introduction – Analysis & Design
• Analysis: Determination of the output of a system or component subject to specified
inputs (geometries, configurations, and loading) and the characteristics (properties,
constraints, etc.) of the system.
• E.g. – Estimation of deflection and induced stresses of a prismatic ‘simply supported
beam’ of known cross-section during application under known loading conditions.
(Strength of Materials)
• Synthesis is the determination of the characteristics of the system or component when specific output is
wanted given the input conditions. It is sometimes called concept design.

• Design: Estimation of the sizes and shapes of parts of a system to meet performance and
durability requirements.
• E.g. – Estimation of the unknown geometry (length and cross-section) of a ‘power
transmission shaft’ for specific loading conditions (Design of Machine Elements)
Difficulties with conventional design
• Again, for a component, when designed prioritizing its durability, by avoiding hidden flaws
and uncertainties in working conditions, it is called a ‘conservative design’. Conservative
designs are bulky yet again may not prevent the catastrophe.
• E.g. – Incorporation of ‘Factor of Safety’ and other uncertainty factors in design.
• To avoid bulkiness, entire design may be changed, and ‘new design’ is adopted. The term
‘new’ may signify different geometry or configuration of the part and/or different
manufacturing technique.
• E.g. – The failures of the Liberty Ships during the World War II may be considered. The
Liberty ship was designed adopting of completely new design, all welded hull for a ship.
• Innovation and commercial rivalry sometimes result in ‘over-engineering’.
• E.g. – 4K display in smartphones makes no visible difference than HD display.
• In conventional design, the designer assumes a particular mechanism, a particular material
and mode of failure for the component. With the help of this information, he determines
the dimensions of the product.
Iterative Design – Trial & Error approach
• The estimated design is analyzed, called ‘Design Analysis’, to see if the corresponding component
performs according to given specifications. If it does, we have an acceptable (feasible) design.
• Although, the design may still be changed for an existing acceptable design to further improve the
component’s performance and/or durability (but may increase cost or complexity).
• It is realized, by experience, that there are many ways to accomplish the same task. Different
‘choices’ of features lead to different designs of the same component, called ‘workable designs’.
• E.g. – Transmission shaft under torsional loading may be 1) long with large dia. or, 2) short with
small dia. Resulting in similar functionality and life (as per continuum consideration)
• The choices, though, are not arbitrary, there are certain ‘constrains’
• E.g. – Space restrictions, budget constrains, the part may be belonging to a pre-designed assembly,
Fastener and bearing are available in specific ratings, etc.
• Design cannot be explicit, is an iterative process improved by feedbacks
Optimization in Design
• To this point, we have many alternative (workable) designs but it cannot be said which
one is the ‘best design’ (or optimum design) among those many possibilities.
• Optimization: It is a process of finding the ‘best possible’ or ‘most efficient’ solution.
Mathematically, it is the process of maximizing or minimizing the ‘objective function’ by
systematically choosing or altering (collectively known as ‘optimizing’) the design
variables subject to constraints (or not).
• Design Synthesis is the process where the designer (creates or,) selects optimum
configurations, materials, shapes and dimensions for a product from a number of
alternative solutions.
(a) Conventional Design
vs.
(b) Optimum Design

• The optimization in design

process is more formal.
• An objective function that
measures a design’s merit
is identified.
• Trend is calculated and
used in block-6 to make
decisions for improving the
design.
Feasibility or Need for Optimization
• Optimization is a costly ‘add-on’ in terms of time and effort starting from identifying and
quantifying objective functions itself. So, design optimization should be worthwhile. And
whether optimization is at all required should be decided before investing.
• E.g. – One of the objective functions in optimizing the project of making a flyover may be
estimated cost, another may be – what will be the objective function? ‘How easy
travelling by the road, now and then – rate the convenience subjectively out of 5’.
The second one may be quantified based on ‘user survey & rating’ which needs
dedicated human resources (engaging and costly) and time consuming (may be after a
year or so, a valid database will be generated).
• Tedious repetitive calculations (during the iterative solution search) can be carried out in
today’s fast computers, making optimization feasible, though a considerable amount of
time still has to be devoted.
Typical flow-chart for Formulating Optimization Problem
• Design variables: A design problem usually involves many design parameters, of which
some are highly influential to the proper working of the design. These parameters are
referred to as design variables (𝑥𝑗 ). Other non-influential parameters should not vary in
relation to the design variables. The formulation of the optimization problem begins with
the identification of design variables.
• Constraints: The constraints represent some functional relationships among the design
variables and other design parameters satisfying equality type constraints like, certain
physical phenomena or law of nature (like, equilibrium laws, conservation laws) and
inequality type resource limitations (like, availability of material, 𝜎𝑤𝑜𝑟𝑘𝑖𝑛𝑔 ≤ 𝜎𝑎𝑙𝑙𝑜𝑤𝑎𝑏𝑙𝑒 ).
Identifying constraints comes second after identifying design variables.
• Constraints bind the feasible domain where the optimum solution lies within but
themselves do not decide the final solution (optimum values of all design variables).
 Optimization problem formulation …contd.
The equality constraints are difficult to handle and therefore, those are, sometimes, broken
in two inequality constraints (mostly done in the optimization in machine design). It can only
be done in case of discrete optimization (design variables vary discretely, like integer values
only).
Inequality constraints are easy and most commonly used and in general, have no restrictions
in terms of values
There are also non-negative constraints (like, 𝑥𝑗 ≥ 0), which are different from variable
bounds (discussed later)
• Objective functions: are functions of design variables (𝑥𝑗 ) and other design parameters
(𝑎𝑘 ), as 𝑓 𝑥𝑗 , 𝑎𝑘 like, cost, weight, Durability of the product, etc. Objective functions are
maximized (or, minimized) for the optimum values of the design variables (𝑥𝑗 ). It is the third
step of formulation of the optimization problem. Some objective functions are difficult to
define, like aesthetics, reliability, etc.
Optimization problem formulation …contd.
Properties of the objective Function: Let, for example, the objective function is the total
cost 𝐶 = 𝑎 + 𝑓 𝑥1 , 𝑥2 , … , 𝑥𝑛 , where 𝑎 is a fixed cost (constant) and 𝑓 is variable cost
which is a function of 𝑛 design variables, 𝑥1 to 𝑥𝑛 . In this case, maximisation of the
objective function 𝐶 may be written as:
𝑀𝑎𝑥 𝐶 = 𝑀𝑎𝑥 𝑎 + 𝑓 𝑥1 , 𝑥2 , … , 𝑥𝑛 = 𝑎 + 𝑀𝑎𝑥 𝑓 𝑥1 , 𝑥2 , … , 𝑥𝑛
Again, 𝑀𝑎𝑥 𝑓 𝑥1 , 𝑥2 , … , 𝑥𝑛 = 𝑀𝑖𝑛 −𝑓 𝑥1 , 𝑥2 , … , 𝑥𝑛 , Which is called the ‘Duality
principle’. Sometimes, in some special optimization techniques like ‘genetic algorithm’,
1
𝑀𝑖𝑛 𝑓 𝑥1 , 𝑥2 , … , 𝑥𝑛 = 𝑀𝑎𝑥 is also used.
𝑓 𝑥1 ,𝑥2 ,…,𝑥𝑛
 Any constant can be added to the objective function or the objective function can be scaled by multiplying
by a positive constant, without affecting the optimum design. However, the extremum value of objective
function changes due to scaling.
• Final task of formulation of optimization problem is to set minimum and maximum bounds
on each design variables (like, 𝑝 ≤ 𝑥𝑗 ≤ 𝑞, etc.). Though, some optimization problems do
not require these information.
Solution to the optimization problem
• We know the merits and demerits of optimization in design and can decide if at all the
optimization be worthwhile.
• If required, we can formulate an optimization problem by identifying the 1) design
variables, 2) constraints, 3) objective function, and 4) variable bounds.
• Every point in the solution space, bounded by the constraints, is a candidate solution
(combination or set of design variables) which satisfies the given constraints and
represents a design. But not every solution is equally desirable or optimum.
• There is only one solution, which is optimum, that maximizes or minimizes the objective
function.
• So, we need to now solve the optimization problem for optimum design variables.
The Helter-Skelter Approach to the Optimization Problem
• If there are only handful of combinations of the design variables are available (like,
specific bearing sizes or, limited manufacturing techniques available, etc.), then objective
function is evaluated for each combination
• And simply comparing the values one can pick the solution or optimum set of design
variables for which the objective function attains an extremum (maximum or minimum).
• This is based on limited and discrete choices of design variables, so the global optima
may not be ensured. (probably the crudest and most casual approach)
Graphical Solution to the Optimization Problem
• Optimization problems having only two design variables can be solved by observing how
they are graphically represented.
• All constraint functions are plotted, and a set of feasible designs for the problem is
identified.
• Objective function contours are then drawn, and the optimum design is determined by
visual inspection.
• Any computation software like MATLAB or MATHEMATICA etc. can be used for assistance.
• A suitable grid display throughout the feasible region would be helpful.
Analytical Solution to the Optimization Problem
• The analytical solution techniques are somewhat different for unconstrained and
constrained optimization problems
• For unconstrained single-variable optimization problems, the first derivative should be zero
for stationarity (necessary condition) and second or higher even non-zero derivative should
be positive for minima or, negative for maxima (sufficient condition)
• Solving the necessary condition itself, the optimal point is obtained.
• The objective function evaluated at the optimal point gives the minimum or maximum
function value depending on the sufficient condition.
• Only difference in unconstrained multi-variable optimization problem is, instead of having
scalar derivatives, we work with gradient vector (vector of first partial derivatives) and
Hessian matrix (array of second and mixed partial derivatives).
Analytical Solution …contd.
• The analytical solution techniques for constrained optimization problems are further more
different because the objective function has to be modified incorporating the constrains
• Then the necessary and sufficient conditions are formulated for the new function and the
conditions not only contains the gradient or Hessian information, also the constraints
information, variable bounds as well
• The design variables are not the only variables in constrained optimization, other variables
and multipliers are introduced to tackle different types of constraints.
• Analytical solution techniques for constrained optimization problems will be discussed in a
later section with the optimality conditions specific to the techniques.
 Constrained Classification of Solution Techniques
for Constrained Multi-Variable
Multi Variable Optimization
(will be discussed later)
Analytical Numerical

Equality Inequality or Transformation Direct Linearized Gradient

Constraints General method Search Search methods
Constraints
Lagrange’s Penalty Variable Frank-Wolfe Feasible Direction
Multiplier Kuhn-Tucker Function Elimination method search
technique Condition method
Random Cutting Plane Sequential Quadratic
Method of Search method Programming
Multipliers
Generalized Reduced
Gradient method

Gradient Projection
method
Extremum and Optimal Point
Local optima: A point or solution 𝑥𝑖∗ is said to
be a local optimal point, if there exists no other
point (𝑥𝑖 ) in the neighbourhood of 𝑥𝑖∗ , which is
better than 𝑥𝑖∗ ; or in other words, if the
objective function y(𝑥𝑖 ) evaluated at the every
neighbouring points 𝑥𝑖 of the solution 𝑥𝑖∗ do not
attain the value as extremum as y 𝑥𝑖∗ .
Global optima: A point or solution 𝑥𝑖∗∗ is said
to be a global optimal point, if there exists no
other point (𝑥𝑖 ) in the entire search space,
which is better than 𝑥𝑖∗∗ ; or in other words, if
the objective function y(𝑥𝑖 ) evaluated at any
point 𝑥𝑖 in the entire search space, do not attain
the value extremum than y 𝑥𝑖∗∗ .
Inflection Point
An inflection point 𝑥𝑖∗ is a point on a curve at which the sign of the curvature (i.e., the
concavity) changes. Inflection points may be stationary points, but are not local maxima or
local minima.
First Derivative and the Gradient Vector
• The optimality criteria for multi-variable functions are different as the derivatives are not
𝑑𝑓 𝑑2 𝑓
scalar quantities or, , in case of single variable function anymore.
𝑑𝑥 𝑑𝑥 2

• The gradient vector at any point 𝑥𝑗∗ is represented by 𝛻𝑓(𝑥𝑗∗ ) which is an 𝑁-dimensional
vector given as:
𝜕𝑓 𝜕𝑓 𝜕𝑓 𝑇
𝛻𝑓 𝑥𝑗∗ = , ,…, ฬ , here, 𝑇 is for ‘transpose’.
𝜕𝑥1 𝜕𝑥2 𝜕𝑥𝑁
𝑥𝑗 =𝑥𝑗∗

• The first-order partial derivatives can be calculated numerically by central difference

𝜕𝑓 𝑓 𝑥𝑖 +∆𝑥𝑖 −𝑓 𝑥𝑖 −∆𝑥𝑖
method, as, =
𝜕𝑥𝑖 2 ∆𝑥𝑖
Second Derivative and the Hessian Matrix
• However, the second order derivatives in multi-variable functions form a matrix, known as
the Hessian matrix, 𝛻 2 𝑓 𝑥𝑗∗ given as follows:
𝜕2 𝑓 𝜕2 𝑓 𝜕2 𝑓
𝜕𝑥12 𝜕𝑥1 𝜕𝑥2
⋯ 𝜕𝑥1 𝜕𝑥𝑁
𝜕2 𝑓 𝜕2 𝑓 𝜕2 𝑓
⋯
𝛻 2 𝑓 𝑥𝑗∗ = 𝜕𝑥2 𝜕𝑥1 𝜕𝑥22 𝜕𝑥2 𝜕𝑥𝑁
⋮ ⋮ ⋱ ⋮
𝜕2 𝑓 𝜕2 𝑓 𝜕2 𝑓
⋯ 2
𝜕𝑥𝑁 𝜕𝑥1 𝜕𝑥𝑁 𝜕𝑥2 𝜕𝑥𝑁
𝑥𝑗 =𝑥𝑗∗

• The second-order partial derivatives can also be calculated numerically, as

𝜕2 𝑓 𝑓 𝑥𝑖 +∆𝑥𝑖 −2 𝑓 𝑥𝑖 +𝑓 𝑥𝑖 −∆𝑥𝑖
=
𝜕𝑥𝑖2 ∆𝑥𝑖 2

𝜕2 𝑓 𝑓 𝑥𝑖 +∆𝑥𝑖 , 𝑥𝑗 +∆𝑥𝑗 −𝑓 𝑥𝑖 +∆𝑥𝑖 , 𝑥𝑗 −∆𝑥𝑗 −𝑓 𝑥𝑖 −∆𝑥𝑖 , 𝑥𝑗 +∆𝑥𝑗 +𝑓 𝑥𝑖 −∆𝑥𝑖 , 𝑥𝑗 −∆𝑥𝑗

=
𝜕𝑥𝑖 𝜕𝑥𝑗 4 ∆𝑥𝑖 ∆𝑥𝑖
Optimality Conditions for Unconstrained Problems
• The definition of a local, global, or inflection point (a change of curvature from convex to
concave at a particular point on a curve) is the same in case of single-variable and multi-
variable functions

• The necessary conditions for optimality:

1) The objective function should be 𝑛-times differentiable
2) a) A point 𝑥 ∗ (one dimensional) is a stationary point (where a maxima, or a minima, or
an inflection point exists) for a single-variable objective function, 𝑓(𝑥), if the first
𝑑𝑓
derivative at 𝑥 ∗ 𝑑𝑥ቚ =0
𝑥∗

b) A point 𝑥𝑗∗ is a stationary point for a multi-variable objective function, 𝑓(𝑥𝑗 ), if the
gradient vector at 𝑥𝑗∗ , 𝛻𝑓 𝑥𝑗∗ = 𝟎. (Remember, 𝟎 is a null vector of dimension 𝑗)
Optimality Conditions for Unconstrained Problems
• The sufficient condition for optimality:
1) a) For a single-variable objective function, 𝑓 𝑥 , there exists a local maxima, or a local
minima, if first non-zero 𝑛𝑡ℎ derivative of the objective function at 𝑥 ∗ be negative
𝑑𝑛𝑓 𝑑𝑛𝑓
i.e. 𝑑𝑥𝑛 ቚ ∗ < 0 and 𝑛 is even, or, positive i.e. ቚ > 0 and 𝑛 is even, respectively.
𝑥 𝑑𝑥 𝑛 𝑥 ∗

b) For a multi-variable objective function, 𝑓 𝑥𝑗 , The Hessian matrix at 𝑥𝑗∗ , i.e.

𝛻 2 𝑓 𝑥𝑗∗ , may be positive semi-definite, if a local or relative minima exists at 𝑥𝑗∗ ;
or, it may be negative semi-definite if there exists a local or relative maxima at 𝑥𝑗∗ .
Furthermore, the local minima or maxima point 𝑥𝑗∗ is a global minimum or
maximum, if the Hessian matrix there, is positive-definite or negative-definite
respectively.
• If 𝑓 𝑥 or 𝑓 𝑥𝑗 is a convex function, there will be a minimum point and if the function is
concave, there will be a maximum point.
Optimality Conditions for Unconstrained Problems
• The sufficient condition for optimality:
2) a) For a single-variable objective function, 𝑓 𝑥 , 𝑥 ∗ be an inflection point if the 𝑛𝑡ℎ
derivative of the objective function at 𝑥 ∗ gives the first non-zero value and 𝑛 is
odd.
b) For a multi-variable objective function, 𝑓 𝑥𝑗 , if the Hessian matrix at 𝑥𝑗∗ , i.e.
𝛻 2 𝑓 𝑥𝑗∗ is equal to zero, it represents 𝑥𝑗∗ is an inflection (saddle) point.
• This conditions cannot say what happens if the derivative fails to exists
• And, this condition cannot say what happens when the extrema occurs at the end point
Checks for Positive Definiteness of Matrices
• A square matrix 𝑀 is positive definite if 𝑀 is symmetric and any one of the following is
true:
1) All the Eigen values are strictly positive (all strictly negative for negative definite)
2) Sylvester’s criteria (probably fastest) – All the leading determinants from the upper
𝑎11 𝑎12 𝑎13
𝑎11 𝑎12
left, i.e., for a 3 × 3 matrix, 𝐷1 = 𝑎11 , 𝐷2 = 𝑎 and 𝐷3 = 𝑎21 𝑎22 𝑎23 ,
21 𝑎22
𝑎31 𝑎32 𝑎33
of the matrix are to be strictly positive. (Must start with a strictly negative 𝐷1 < 0,
then alternating in sign 𝐷2 > 0, 𝐷3 < 0, 𝐷4 > 0 …, but not zero gives negative definite)
3) All its pivots (the diagonal values obtained after reducing the matrix into the row
echelon form without exchanging rows, i.e. Gauss elimination without pivoting) are
strictly positive
Checks for Positive Definiteness of Matrices
• If any of Eigen values is zero but all values are non-negative ≥ 0 , the matrix is called
positive semi-definite. (If any of Eigen values is zero but all values are non-positive ≤ 0 ,
the matrix is called negative semi-definite)
• Or, starting with a strictly positive 𝐷1 , (> 0) if at least one of the 𝐷2 , 𝐷3 , … is zero, the
matrix is called positive semi-definite. (For negative semi definiteness, Must start with a
strictly negative 𝐷1 < 0 , then alternating in sign 𝐷2 ≥ 0, 𝐷3 ≤ 0, 𝐷4 ≥ 0 …, and at least
one has to be zero)
• If the Eigen values are mixed with both positive and negative values, the matrix is indefinite
• Or, if the principal minors or 𝐷𝑖 for 𝑖 = 1,2, … are none of the above, the matrix is
indefinite. It includes 𝐷1 = 0, randomly signed 𝐷𝑖 , alternating 𝐷𝑖 starting with a positive,
etc.
Numerical Solution to the optimization problem
• The optimum solution is obtained by searching the feasible solution space bounded by the
constraints.
• Remember, constraints alone cannot yield an optimum solution, unless number of
constraints are equal to number of unknown design variables
• The search in the solution space has to be exhaustive to ensure global optima. The
method can be tedious and repetitive if not done in a methodical way.
• Therefore, there is a need to develop appropriate numerical search algorithms which are
generally iterative in approach and, may or may not depend on the derivative information
of the objective function.
• A few numerical search algorithms for both unconstrained single- and multi-variable
optimization problems will be discussed in the subsequent sections.
Classification of Numerical Solution Techniques for
Unconstrained Optimization (will be discussed in subsequent sections)
Unconstrained Optimization Problems

Single Variable Multi Variable

Bracketing Region Direct Search Gradient Based

Method Gradient Based Method Method
Elimination
Method
Method
Evolutionary Directional Cauchy’s
Exhaustive
Algorithm Search Steepest
Search Bisection
Golden Descent
technique Method
Section Unidirectional Method
Search Search
Newton – Newton’s
Raphson Simplex Search Method
Method
Order of Convergence in Numerical Analysis
• Let, the sequence 𝑥 ∞ 𝑖=0 is constituted of evolved values of a single design variable 𝑥, in many
iterations 𝑖. The sequence 𝑥 ∞𝑖=0 can be called as sequence of approximate solutions.
𝑖
• If the sequence of approximate solutions converges to the actual solution 𝜉 , where 𝑥 ≠ 𝜉, ∀𝑖 (∀
means ‘for all’), there remains an absolute errors, 𝜀𝑎𝑖 = 𝜉 − 𝑥 𝑖 , ∀𝑖 = 0, 1, 2, … , ∞.
𝑖 𝑖
• It can be said that, if 𝑥 converges to 𝜉, then the absolute error 𝜀𝑎 converges to 0.
𝑖+1
𝜉−𝑥 𝑖+1 𝜀𝑎
• If there exist two constants, 𝐶 ≥ 1 and 0 < 𝑅 ≤ 1, such that, lim 𝐶 = lim 𝑖 𝐶 =𝑅
𝑖→∞ 𝜉−𝑥 𝑖 𝑖→∞ 𝜀𝑎
𝐶
Then, it can be expected that, 𝜀𝑎𝑖+1 ≈𝑅 𝜀𝑎𝑖 , for a large enough 𝑖.
𝑖 𝑖+1
So, 𝑥𝑖 ∞𝑖=0 converges to 𝜉 means, absolute error 𝜀𝑎 converges to 𝜀𝑎 and both converges to 0,
with the order of convergence 𝐶 and rate of convergence 𝑅, also known as asymptotic error constant.
• The order of convergence signifies how fast the convergence occurs.
If 𝐶 = 1 (and 𝑅 ≤ 1), the sequence is linearly convergent or has first order convergence (it is called
sublinear, when 𝑅 = 1 and superlinear, when 𝑅 < 1. Later is comparatively faster)
If 𝐶 = 2, the sequence is quadratically convergent or has second order convergence