End-to-End Learning and Auto-Differentiation - Forces, Uncertainties, Etc.

Explore cutting-edge applications of deep learning and auto-differentiation in molecular simulations through this 56-minute lecture by Rafael Gomez-Bombarelli from MIT. Delve into topics such as active learning of machine learning potentials, deep neural network generative models for coarse-grained atomic system representations, and differentiable simulations for reaction path finding. Examine the challenges of overfitting and generalizability in AI for science, discussing the scalability of active learning and sensitivity of deep learning solutions in molecular problems. Gain insights into excited state potentials, uncertainty in interatomic potentials, coarse-graining techniques, and the use of equivariant generative decoders in molecular liquid simulations.

Intro
Virtuous cycle for design
Quick thoughts on Emergence
Excited State potentials
Uncertainty and active learning
Adversarial Attacks on Interatomic potentials
Rev-MD17 uncertainty
Testing the potential - is this emergence?
More on Pair potentials
DiffSim for barrier crossing
Issues and tools needed
Coarse Graining Auto-Encoding Framework
CG of molecular liquids
Equivariant generative decoder
GenZProt
Reconstruction vs Generation