Neural networks applied to the modeling of yield surfaces in metal plasticity.
Repository for code and data supporting the article "On the use of neural networks in the modeling of yield surfaces"
Preprint: ResearchGate
Final:
There are four directories: partA, partB, partC and partD.
Contains all notebooks referenced in Sections 3-6 of the article (direct modeling of yield surfaces via NNs) and also recorded in Appendix-A/Table-3.
Contains all notebooks and data supporting Section-7 of the article (Mapping the convexity domain of a yield function)
The data files in partB are as follows:
- Training data:
aCVXbdry_deg_4_test.txt - Near boundary points in the training data (See Article/Fig.20 for explanations):
aCVXbdry_deg_4.txt - Validation data:
aCVXbdry_deg_4_Valid_test.txt - Near boundary points in the validation data (See Article/Fig.20 for explanations):
aCVXbdry_deg_4_Valid.txt
Other files/notebooks in partB are as follow:
- NN training:
aFitCVXdeg4_12_C_9D4.ipynb - NN validation:
nb24_aValidCVXdeg4_12_C_9D4.ipynb - Poly4 calibration example (used to generate the model in Fig.22):
nb24_aPolyNoptim.ipynb - Supporting python script:
cvxDomainPolyN.py
Contains a UMAT implementation of feed forward (densely connected) NN and some simulation tests featured in Appendix-B and Appendix-C of main article. Also, utility scripts for weights extraction are posted
in sub-dir partC/WeightsExport. Usage instructions are detailed
in /partC/nnYS_Usage.pdf
Contains supporting code for Appendix-D: Curvature and convexity.
The two python scripts Verify_Convexity_2D.py and Verify_Convexity_3D.py can be used to verify the convexity of plane stress and 3D yield functions. They require as input only the function (no derivatives - these are calculated by using the automatic differentiation capabilities of Autograd).
For convenience (quick preview), I added html exports of some of the notebooks. Note, however, that the final version of code is always in the corresponding notebook.
The utility scripts ysnnutil.py and cvxDomainPolyN.py are required imports for partA and partB, respectively. In particular, if Google Colab is used, then variable wdir must be set to the corresponding path, e.g.
wdir = '/content/drive/MyDrive/.../partA'