Search papers, labs, and topics across Lattice.
AccelNet introduces a novel method for accelerating trained aenet and n2p2 neural-network potentials without the need for retraining, leveraging the hidden finite-rank structure of angular terms with separable one-neighbor weights. By substituting explicit neighbor-pair loops with one-neighbor Cartesian moments, the method achieves exact backward compatibility while maintaining precision in reproducing descriptors, energies, and analytic forces. Validation on H$_2$O and TiO$_2$ models demonstrates its effectiveness in LAMMPS molecular-dynamics simulations, with the implementation and tools made available as open-source software.
Achieving exact acceleration of neural-network potentials without retraining could revolutionize molecular dynamics simulations by enhancing efficiency without sacrificing accuracy.
We present AccelNet, an exact, backward-compatible method for accelerating existing trained aenet and n2p2 neural-network potentials without retraining. For angular terms with separable one-neighbor weights and a finite polynomial dependence on $\cos \theta$, the method exploits their hidden finite-rank structure to replace explicit neighbor-pair loops by one-neighbor Cartesian moments. AccelNet reads models trained with either package and reproduces their descriptors, energies, and analytic forces to floating-point roundoff. We verified this equivalence for H$_2$O and TiO$_2$ models and tested the resulting potentials in LAMMPS molecular-dynamics simulations. The implementation, model-conversion tools, and LAMMPS interfaces are released as open-source software.
