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.
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).
MVE Ensemble with RL simulator param search (WIP!)
- 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
I started by using Hamiltonian-NN and Lagrangian-NNs as templates (and evolved from there).