Search papers, labs, and topics across Lattice.
Department of Engineering, University of Cambridge, Cambridge, UK
4
0
3
Multihead replay not only preserves out-of-distribution robustness but also consistently outperforms models trained from scratch across diverse chemical tasks.
Current machine learning interatomic potentials fall short in capturing electrostatic interactions in complex systems like metal-water interfaces, but this work provides a roadmap for building more accurate and expressive models.
Widely used short-ranged MLIPs can spuriously predict metallization in water due to missing long-range electrostatics, a flaw that can be fixed by including these interactions explicitly.
Forget generic MLIPs: MACE-POLAR-1 nails long-range electrostatics, slashing errors in protein-ligand binding and crystal energies.
