What is torch-pme

torch-pme provides an interface in which the positions of the atoms in a structure are stored in torch.Tensor objects, or in a metatensor.System object.

The primary goal is to design a library to compute long-range interactions that can be easily integrated with existing short-range machine learning (ML) architectures, essentially providing an easy-to-use framework to build range separated models for atomistic machine learning. To this end, our reference torch-pme library provides

1. A modular implementation of range-separated potentials working for arbitrary unit cells including triclinic ones, as well as for systems with free (non-prediodic) boundary condistions.

2. Full integration with PyTorch, featuring differentiable and learnable parameters,

3. Efficient particle-mesh-based computation with automatic hyperparameter tuning,

4. Pure long-range descriptors, free of noisy short-range contributions,

5. Support for arbitrary invariant and equivariant features and ML architectures.

torch-pme can be used as an end-to-end library computing the potential from positions and charges and as a modular library to construct complex Fourier-domain architectures. To use torch-pme as an end-to-end a library the main entry points are Calculators, that compute pair Potentials combining real-space and k-space components. They take a description of a structure as input and return the calculated potential at the atomic positions as output. To use torch-pme as a modular library, we provide a set of building blocks that can be combined to build custom range-separated architectures, as shown in the figure below.

A schematic representation of the main building blocks that are contained inside a Calculators of a range-separated architecture, that combines an evaluation of the short-range part of the Potentials \(v_\mathrm{SR}(r)\) based on local interatomic distance information with the evaluation of the long-range part \(v_\mathrm{LR}(k)\) using grids via a Mesh Interpolator and a Kspace filter.