Autograd vs. elbow grease: Comparing Allegro and FLARE for performance-portable extreme-scale simulations

Allegro and FLARE are two very different packages for constructing machine learning potentials that are fast, accurate, and suitable for extreme-scale molecular dynamics simulations. Allegro uses PyTorch for efficient equivariant potentials with state-of-the-art accuracy, while FLARE is a sparse Gaussian process potential with an optimized C++ training backend leveraging Kokkos, OpenMP, and MPI for state-of-the-art performance, and a user-friendly Python frontend. We will compare and contrast the two methods, discuss lessons learned, and show scientific applications ranging from catalysis to biology.