kristinlindquist / pinns

Experiments with physics-informed neural networks

PINNs

Various experiments with PINNs for molecular dynamics simulation. More generally, I'm exploring how to give models intuition based on formal systems with a sort of FEP/active learning cycle.

Rough Idea

Attempt to solve a hairy "outer problem" (e.g. protein-protein-interaction) by teaching a model to choose and learn from a large number of physics simulations. In a vaguely FEP sense, have the model attempt to gather more data in areas corresponding to its greatest surprisal. "Gather more data" meaning choosing + parameterizing simulations along with its objective. It will explore parameter space driven by reward based on learning objectives and effort (compute time).

Notebooks

MVE Ensemble with RL simulator param search (WIP!)

Resources

• Steve Brunton's lectures on physics-informed neural networks
• Erik Lindahl's lectures on Molecular Biophysics
• Port-Hamiltonian systems
• Karl Friston (FEP) on Sean Carroll's podcast

Acknowledgements

I started by using Hamiltonian-NN and Lagrangian-NNs as templates (and evolved from there).