Reciprocal-Space Attention (RSA)
RSA is a next-generation architecture that captures long-range interactions directly in reciprocal (Fourier) space using Fourier positional encoding and kernelized attention. It provides linear-scaling, charge-agnostic modeling of electrostatics and dispersion within machine-learned interatomic potentials. This framework can be combined with any short range MLIP backbone in an end-to-end differential fashion and allows access to long-range physics seamlessly, enabling density functional theory accurate modeling at scale.
Repository: reciprocal_space_attention