graph_longrange

graph_longrange provides various long-range operatoins in pytorch, for use in charge-aware MLIPs. The repo works mostly with Gaussian type orbitals, and provides the following fundamental operations. All the code is pytorch and for systems up to a few thousand atoms on a single GPU, the code is reasonably fast compared to many modern MLIPs.

• Coulomb Energy from a set of atoms with atomic multipole moments
• Atom Centered Electrostatic Features, similar to the LODE descriptor
• Damped Coulomb Interactions via Gaussian Type Orbitals, instead of point charges
• Consistent Realspace and Periodic Implementations
• Correction terms to energy and features for handling slab geometries, as well as molecules in boxes
• Efficient batching for pytorch training
• Considerable precopmutation for electrostatic features to speed up self-consistent-field loops
• Interpolation functions for probing the potential and density around atomistic systems

Additional functionality will also be coming in the near future, including:

• Confined Jellium slabs of charge, useful when simulating electrochemical interfaces

This repo currently supports several long-range MLIP architectures including the MACE-POLAR-1 foundation model.