Topology optimization with deep learning

SIMP_multi.py code used for generate training dataset.

CNN.py code used for training neural network.

load_model.py code used for load neural network.

SIMP_multi_dist.py code used for generate dataset with a distributed load.

Results folders

• results_1f Folder with the data with just 1 force
• results_2f Folder with the data with just 2 force is divided in third parts, this contain the first part with the first columns.
• results_2f_2 Folder with the data with just 2 force second part
• results_2f_3 Folder with the data with just 2 force third part
• results_merge_# Folder with the data mix of all folders
• results_merge_3 Folder with the data mix of all 2f results 2f_1 and 2f_3
• results_dist Folder with the data of distribution loads
• results_rand Folder with the data random 10000
• results_rand4 Folder with the data random 20000 with just vertical forces
• results_rand5 Folder with the data random 40000 with just vertical forces
• results_rand_6 Folder with the data random 40000 with just vertical forces
• results_rand_7 Folder with the data random 40000 with just vertical forces and no volumen change 0.5 for all

Model folder

• Basic_NN Model more basic, don't work at all
• first_NN Model with more data
• second_NN Model with data of 2 loads
• third_NN Model with data mix of merge results
• U_NN U-Net model with data mix of merge results
• U_NN2 U-Net model with data mix of merge results 50 epoch
• ViT_test ViT model with data mix of merge results 50 epoch wrong data
• ViT2 ViT model with data mix of merge 2 results 50 epoch
• ViT3 ViT model with more parameters 8M approx 50 epoch
• vit_last_100 ViT model with 100 epoch batch size small
• unn_last_100 U-Net model with 100 epoch batch size small
• model_unet_merge U-Net model with earlyStop 5 epochs and train with results_merge_3
• model_rand_# U-Net model with 50 epochs the # correspond to the number of the rand dataset

Models

• Best_Hybrid_# is the best hybrids models
• Best_Hybrid_#_reg is the best hybrids models applying regularization techniques, dropout and dynamic learning rate.