Machine Learning at the Flatiron Institute Seminar: Aditi Krishnapriyan

Title: The role of scaling and training strategies in machine learning for scientific modeling

Abstract: Machine learning in the scientific domain has made rapid progress in recent years. Many models and methods in this field incorporate constraints into the problem setting. This includes constraints such as building in rotational equivariance into neural networks. In this talk, Krishnapriyan will discuss: 1) if such constraints are needed and 2) the role of training strategies, contrasted against model development, in building the next generation of large-scale machine learning models for scientific applications.

For the first point, Krishnapriyan will discuss systematic strategies for scaling neural network interatomic potentials, which are an attractive alternative to ab-initio methods for molecular dynamics simulations. Systematic scaling provides strong results across diverse chemical domains, while being significantly more efficient than models with built-in equivariance constraints. For the second point, Krishnapriyan will discuss better training strategies to improve neural networks for modeling atomistic and continuum phenomena. Here, insights from numerical solvers and physics-based simulation can be used alongside training neural networks through fully differentiable settings.