Micromechanics Generator, or mm-Generator, is an implementation of various DL models to generate stress fields in polycrystalline microstructures. The purpose of the project is to accelerate numerical simulations with the help of trained DL models.
The models were trained on 2D microstructures in a supervised fashion: material property fields including the geometry on the input side and the stress response generated by the numerical solver (DAMASK https://damask.mpie.de/index.html ) on the output side.
As for now, the following DL models are implemented: U-Net, Fourier Neural Operator (FNO), and conditional Generative Adversarial Network (cGAN). The generator for the cGAN can be changed easily to either U-Net or FNO from the config.yml file. Of course, both U-Net and FNO can also be trained independently with standard loss functions like MAE, MSE, etc.
- Refactoring of inference scripts
- Functionality to compute the receptive fields for the discriminator