Physically Inspired Machine Learning for Excited States - IPAM at UCLA

Explore physically inspired machine learning techniques for excited states in this 44-minute lecture by Julia Westermayr from the University of Warwick. Delve into the application of machine learning methods in photochemistry, focusing on overcoming the sparse data problem for excited states. Discover how combining data from various quantum chemistry methods and incorporating physics into machine learning architectures can lead to more accurate and data-efficient models. Follow the application of these techniques in photodynamics simulations of tyrosine, revealing unexpected reaction mechanisms and providing new insights into the photochemistry of biological systems. Gain knowledge on topics such as excited-state surface-hopping dynamics, phase correction in ML algorithms, learning nonadiabatic couplings, and the use of unsupervised ML in molecular design. Recorded at IPAM's Monte Carlo and Machine Learning Approaches in Quantum Mechanics Workshop, this talk offers valuable insights for researchers and students interested in the intersection of machine learning and quantum chemistry.

Intro
Phototherapy
Excited-state surface-hopping dynamics
Problem: Quantum chemistry (QC)
Where can machine learning (ML) help?
Photochemical processes
Proof of concept
Training set generation
Arbitrary phase of the wave function
ML excited-state dynamics
Machine learning for photodynamics
Limitations of existing approach: Phase correction
Phase-free training algorithm
Learning nonadiabatic couplings
Application to tyrosine: Training set
Roaming in tyrosine
Unsupervised ML
Roaming atoms: radicals or protons?
Summary
Learning orbital energies
ML for photoemission spectroscopy
Generative ML for molecular design
Targeted molecular design