Training Neural Network Potentials: Bayesian and Simulation-based Approaches

Explore advanced techniques for training neural network potentials in molecular dynamics simulations through this comprehensive talk. Delve into data-efficient methods like relative entropy minimization and differentiable trajectory reweighting to enhance the accuracy of simulations with limited data. Learn about scalable uncertainty quantification for reliable estimation of credible intervals in molecular dynamics observables. Discover how these approaches can improve the use of neural network potential-based simulations in real-world decision-making for material design and drug discovery. Gain insights into force matching, coarse-grained models, and the importance of uncertainty quantification in molecular dynamics simulations.

- Intro and Overview
- Outline: Training Neural Network Potentials
- Force Matching
- Relative Entropy Minimization
- Prior Potential: Delta Learning for GNN Potentials
- CG Water Model
- CG Alanine Dipeptide
- Bottom-Up/Top-Down Training
- Diferentiable Trajectory Reweighing DiffTRe
- Coarse-Grained Model of Water
- The Need for Uncertainty Quantification
- Lennard Jones Toy Example: Posterior Modes
- Summary and Outlook
- Q+A