Neural network based solvers for partial differential equations and inverse problems :milky_way:. Implementation of physics-informed neural networks in pytorch.
We added Geometries as a new interface for the PDE dataset. This allows a generic inference pipeline and reduces the amount of code to solve a PDE with NeuralSolvers. This includes various sampling methods and brings adaptive sampling into Neural Solvers. Furthermore, we improved the prints and the logging features.
Tensorboard logging is available and can be activated in the fit() function. Logger enables automatic tracking of all loss terms and their weights
Logging of Loss gradients in order to identity gradient pathologies, you can track the gradient updates by adding a logger and setting the
track_gradient flag in the fit() function.
Fix the usage of the to method in MLP and FingerNet
A working example of the heat equation
Learning Rate Annealing is now available and can be activated in the fit()function. This algorithm balances the weights for initial and boundary conditions depending on the stiffness of the underlying PDE
Logger are available and can be activated in the fit()function. Logger enable automatic tracking of all loss terms and theire weights for example with Weights and Biases which is implemented yet. By implementing the Logger_Interface you can build your own loggers.
Pretraining the fit() function allows a pretraining by optimizing only on the initial condition
Callbacks its possible to write custom callbacks in order to modify the behavior in the training loop. At the moment only callbacks at the end of epochs are supported
Fix the bug in the calculation of neumann and robin boundary condition
A working example of burgers equation