Stress GAN : A Generative deep learning model for 2d stress distribution prediction

• Abstract

cGAN 을 사용하여 2D von Mises stress 를 예측.

geometries, load, boundary condition 을 condition 으로 사용

기존 CNN 보다 high resolution stress distribution 예측에 성능 우수함

 

• Introduction

기존에 DN 을 사용하여 residual/shear/von mises stress 를 예측하는 문헌은 많았음. 하지만 geometry 와 B.C 가 고정되어있어 범용적인 사용이 어려웠음. 본 문헌에서는 GAN 을 사용하여 그 한계를 극복.

Dataset 은 128x128 fine-mesh 로 38,400 sample 을 사용했으며 10 개 Pattern 의 Load position 과 8개 패턴의 BC 를 사용

 

• Related Work

...

 

• Technical Approach

Architecture of StressGAN

- Input of generator : 3 channel m x m resolution images (geometry, BC, Load ) *m x m meshes

 

1. BC Condition : 0,1,2,3,4 : void / free / affixed horizontaly / affixed vertically / affixed on both

2/3. magnitudes of vertical / horizontal load in the corresponding pixel

 

- Output : von mises stress distribution *m x m meshes

 

 

- Loss function : L2 (MSE) distance loss *Previous work has shown that L2 distance work well for predicting strses distribution

* Gaussian noise vector Z 를 제거 (특정 조건에만 맞는 결과 얻기 위함)

Hyper parameter 사용

 

- Metric : MAE, PMAE, PAE (Peak stress absolute error), PPAE (Percentage peak stress absolute error)

*value 가 stress 랑 그냥 같은거 아닌가..???

yj : ground truth value in each element / yj hat : predicted

Y : set of all ground truth value

 

• Dataset (SolidsPy 사용)

- Fine mesh multiple structure datasset : 38,400 ea (60 geometries, 10 load condition, 8 B.C)

- Coarse mesh cantilever beam dataset : 120,960 ea (6 category geometri, 80 geometrie, load on teh right end of beam, left end of beam is fixed, load orientation 0~355 degree, load magnitude 0~100N)

 

- Train detail : learning rate = 0.001 by Adam optimizer / batch size 64 / lambda 1~10 / 80,20 train,test set

*learning rated and batch size 는 gride search 로 결정됨

 

 

*특이사항 :

 When the network is trained and tested using the coarse-mesh dataset, we remove four blocks close to thebottleneck to keep the bottleneck representation of input conditions as a 512 feature vector. We remove the ReLU layerin the last upsampling block with the consideration that mechanics analysis results other than von Mises Stress mightcontain negative values. The convolutional layers and deconvolutional layers both have kernel sizes as 5×5 and stridesize as 2