Machine Learning to Improve the Exchange and Correlation Functional in DFT

Explore machine learning approaches to improve exchange and correlation functionals in density functional theory through this conference talk. Delve into the potential of neural networks as universal approximators for creating more accurate functionals. Examine two novel approaches: injecting prior physical knowledge into training procedures and incorporating physical information directly into optimization algorithms. Discover how these methods lead to data-efficient and reliable models that outperform hand-designed functionals. Consider the cautions and challenges associated with machine-learned models in quantum mechanics. Gain insights into applications for supercritical liquids, water simulations, and energy calculations.

Introduction
Framework
Motivation
Supercritical liquid
Simulations
State of the art
Adult approach
In real space
Parameters
Projections
Regularization
Basin optimization
Covariance matrix
What we learned
Two methods
Double optimization
Results
Results for water
Challenges
Growth and optimization
Gradient optimization
Consistent loop
Loss function
Enhancement Factors
Energy
Hybrid
DeepMind