Overview¶
FlashCart provides Cartesian tensor operations and an equivariant interatomic potential built on them. It includes tools for training and evaluating models and running simulations with ASE and LAMMPS.
Install FlashCart and its dependencies.
Train, evaluate, and use a FlashCart model.
Reference for the Python classes and functions.
Main components¶
Cartesian tensor operations: Tensor products, tensor expansions, and equivariant linear maps, with Triton and PyTorch implementations.
FlashCart potential: An equivariant model built from these tensor operations.
Training and evaluation: Configurable losses, checkpointing, and distributed training.
Simulation interfaces: ASE calculators and LAMMPS integration through ML-IAP.
How it works¶
FlashCart expresses Cartesian tensor products in the \(2l + 1\) independent components of each rank-\(l\) irreducible tensor. Symmetry and trace relations determine the remaining components, so they need not be stored during model evaluation. Writing the products in independent components gives polynomial expressions that can be simplified and differentiated symbolically before compilation. FlashCart translates these expressions into fused Triton kernels, including the derivatives required for model evaluation and training.
The FlashCart potential uses these products to describe atomic neighborhoods. Each message-passing block combines features from neighboring atoms with tensors describing their relative directions. Summing these contributions gives an atomic basis, whose repeated tensor products build higher-order correlations. After each product, a learned linear map mixes the resulting features back to a fixed number of channels per tensor rank. Each additional correlation order therefore requires one product and one linear map, without increasing the channel width. The model predicts atomic energy contributions from the resulting features and sums them to obtain the total energy.
The tutorials cover Data preparation, Training configurations, Molecular dynamics, and Inference and training times. See the FlashCart paper for the model configurations and benchmark setup.