compliance

values
comparable to
the designs
obtained via
densitybased
title authors abstract
topology
Abstract: The number of studi
optimization. In
for more resilient and sustaina
summary, the
Topology optimization of multi-scale structures: a review durability properties of UHPC.
proposed
a thorough understanding of t
A review of topology optimization design methods foCHEN Xiaoqian et al.CPINNTO
Yu, Yonggyun paper presents a systematic re
Downsizing weight while upsizing efficiency: An ex Gaganelis,
Chen,Taeil
Hur, Georgios;framework
Chun-Teh A three-phase
columns, to
beam–column joints,
Generative Deep Neural Networks for Inverse MaterGu, opens up novel
Jung,Grace
JaehoX. In recentcompressive,
tensile, years, and bon
and interesting
Deep learning for determining a near-optimal topoloSosnovik,
Jang, In Gwun
Ivan different behavior than conven
possibilities for
Neural networks for topology optimization Oseledets, Ivan when
In thisdesign and analysis app
research
structural
A Physics-Informed Neural Network-based TopologyHyogu O Jeong philosophies is
Physics-Informed warranted to ta
topology
analysis and design approache
A complete Physics-Informed Neural Network-basedHyogu f
FernandoJeong
V. optimization.
and limitations of these appro
Machine learning for topology optimization: Physics Senhoraa, The high to
addition computational coststo
normal-strength
UHPC evolution, development, and utilization in construction: a review. components, and structural st
A Review of Developments and Challenges for UHPC in Structural Engine suggests future directions for
 （用途、优 模型会体现出比传
点）Durability 统代理模型更显著
properties, 的拟合优势。
Mechanical 模式 2 构建微结
properties, 构的生成器:
Aesthetic and 模式 3 基于神经
engineering,Eco 网络重参数化的多
-friendly 尺度结构拓扑优化
product. 设计。即利用神经
Topology optimizati
However, apart from 基本方法：Popular methods
Over the
such
past
as few
those
a few exceptions, for instance,12 most year
based
One importa
on density
网络直接得到结构
attempts (理论极限)A
(Sigmund
higher
generally haveresolution
2001), level
in common discre
sets
that
The
工程结构优化问题可被拆解为设计变量、约束条件和目标函数等几部分
and that they are merely built morethe
(传统机器学习的不足)(a)
experimental UHPC的微观结构是一个复杂而高度工程化的系统，通过优化各种成分的比例和配置，以实现材料所需的卓越性能。精细
slender due to the
local minima 的拓扑优化结果
properties
problem. Mostoffered by the problems
optimization new materials, that ii
of interest
and thus adapt either Doing
the so,
reinforcement13–16
the str （轻量化具体表现）However,
or the overall
the structural
HCS truss design.17–19
structure's overall
Doing
(b) the calculation of analytical gradients. For most topology optimization applicat
investigation weight
so, the amo
str
In this paper, a novel Physics-Informed Neural Network-based Topology Optimization (PINNTO) framew
(主动学习的优势)show thatconducted
(c)
thethe
performance
availability
by ofofthe
objective
inversefunctions.
designs canForbe
materials
improveddiscovery
using theandactive
designlearning
prob
Optimization (TO) and Physics-Informed Neural Networks (PINNs). One of the key novelties in the PINN
showed the
ability to attain the optimized topology without the assistance of labelled data or FEA. Based on this in
(加快速度)Theincluding excellent
main novelty
CNN分割图像
of this workthat
is tothe
state the problem as anand
image
domains 3D cases, given boundary conditions the segmentation task.
loading configurations are co
mechanical
avenue for structural design in this ‘data-rich’ age.
properties of
UHPC as a
gh computational cost of topology optimization has prevented its widespread use as a generative design tool. T
construction
material (造价高)This is attributed (降低造价的方法)suggested
to the high cost of materials,
ideas
(还没有相关规范)There
to replace
the lackitsare
of ex
th
n
(UHPC与普通混凝土的不同，因而在设计上有一些不同)Due
（UHPC定义）a fiber-reinforced
(造价高)In
to the ultrahigh
the concrete
past,tensile,
UHPCthat
(在桥梁中的应用)The
carried
compressive,
has a high
minimum
(应力应变曲线)As
material
advanced
andcompressive
bond
Why?
a
and
fiber-reinforced
mechanical
strengths,
construction
Current
strength
R/UHPC
proper
unique
cos
ceof
 s.
c. The
negative
impact of
cement
productio
n on the
environm
ent, since
UHPC
her resolution discretization enables the appearance of fine geometric details, which may bring the performance
productio
roblems of interest in engineering and scientific fields involve many critical points including local minima, saddl
n requires
sptimization
overall weight amounts the
applications, to merely
goal is51 kg, to
often which corresponds
search
about to a reduction
for an optimal of 83% compared
material distribution to thecompliance
to minimize reference
ion (PINNTO) framework has been developed. The proposed PINNTO framework is a combination of Topology
ery and design problems using large-scaledoublematerials data and informatics in existing databases, such as the ma
novelties in the PINNTO framework is the replacement of the FEA component in the traditional TO approach with
the
FEA. Based on this investigation, it can also be deduced that PINNTO can acquire optimal topologies for various
amount of
configurations are correctly imposed for the associated energy-based PINN. In summary, the proposed PINNTO o
cement
used in
conventio
erative design tool. To reduce this computational
nal cost, we propose an artificial intelligence approach to drastical
（优点）UHPC General
has beenUHP
aWith
majorthis
development
t concrete. in concrete technology in the first twenty years of the 21st
Why? Current R/UHPC models require further enhancement in accuracy and reliability, particularly for predicting
 ay bring the performance of optimized structures closer to theoretical limit.
uding local minima, saddle points, and discontinuities. Gradient descent is known to often get stuck at critical p
nmpared to thecompliance
to minimize reference beam.
(maximize stiffness) for a given loading condition and volume fraction constraint. In th
mbination of Topology
atabases, such as the materials project, automatic flow for materials discovery, open quantum materials databa
ditional TO approach with PINNs. Consequently, PINNTO possesses the capability to overcome inherent shortco
mal topologies for various types of complex
y, the proposed PINNTO opens up a new

nce approach to drastically accelerate topology optimization without sacrificing its accuracy. The resulting AI-dr

particularly for predicting the flexural and shear strengths of R/UHPC members. The establishment of a uniform
ten get stuck at critical points where the gradient has vanished. Therefore, the solutions obtained using a gradi
e fraction constraint. In those problems, the analytical gradients could be calculated using adjoint methods.[12]
quantum materials database, and novel materials discovery, even numerical gradients cannot be calculated sinc
ercome inherent shortcomings of FEA-based topology optimization, such as the difficulty of solving nonlinear pr

uracy. The resulting AI-driven topology optimization can fully capture the underlying physics of the problem. As

tablishment of a uniform code is critical to promote the use of R/UHPC on a practical level.
ns obtained using a gradient-based optimization method may have inferior performance due to a bad selection o
ing adjoint methods.[12] Compared to numerical gradients, analytical gradients are exact and the calculation re
cannot be calculated since the
ty of solving nonlinear problems and dealing with complicated materials and domains. To demonstrate the effe

hysics of the problem. As a result, the machine learning model, which consists of a convolutional neural network
e due to a bad selection of initial values in the optimization, and how to select good initial values is unfortunate
act and the calculation requires much less computational cost. However, the calculation of analytical gradients
To demonstrate the effectiveness and the potential of the proposed PINNTO framework, a number of case stud

volutional neural network with residual links, is able to generalize what it learned from the training set to solve
 tial values is unfortunately quite challenging and problematic;
on of analytical gradients is nontrivial and only possible when knowing the explicit form of the optimization prob
k, a number of case studies have been conducted. The numerical examples indicate that PINNTO has the

the training set to solve a wide variety of problems with different geometries, boundary conditions, mesh sizes,
m of the optimization problem as well as the governing equations and the algorithm that is used to solve the equ
hat PINNTO has the

y conditions, mesh sizes, volume fractions and filter radius. We train the machine learning model separately fro
t is used to solve the equations. Therefore, it is impractical to calculate analytical gradients for most materials d

ning model separately from the topology optimization, which allows us to achieve a considerable speedup (up to
ients for most materials discovery and design problems. On the other hand, finite-difference methods (FDM) are

nsiderable speedup (up to 30 times faster than traditional topology optimization). Through several design exam
rence methods (FDM) are commonly used to calculate numerical gradients when the calculation of analytical gr

ugh several design examples, we demonstrate that the proposed AI-driven topology optimization framework is
alculation of analytical gradients is infeasible. However, the calculation of numerical gradients is computational

ptimization framework is effective, scalable and efficient. The speedup enabled by the framework makes topolog
radients is computationally expensive. Consequently, when the number of design variables is large, the calculat

framework makes topology optimization a more attractive tool for engineers in search of lighter and stronger st
ables is large, the calculation of numerical gradients is often the bottleneck in the optimization and the results a

of lighter and stronger structures, with the potential to revolutionize the engineering design process. Although t
mization and the results are subject to inaccuracies;

esign process. Although this work focuses on compliance minimization problems, the proposed framework can b
proposed framework can be generalized to other objective functions, constraints and physics.
PPT

1.研究背景
UHPC介绍，优点 high cost
拓扑优化简介

2.研究现状，研究意义

3. 研究方向
深度强化学习，物理驱动神经网络

4. Reference