NVNMD-v2: Scalable and Accurate Deep Learning Molecular Dynamics Model Based on Non-Von Neumann Architectures

XY Yu and G Yang and ZY Zhao and JH Li and XY Xiao and X Zhang and J Liu and PH Mo, JOURNAL OF CHEMICAL THEORY AND COMPUTATION, 21, 8918-8932 (2025).

DOI: 10.1021/acs.jctc.5c01050

Molecular dynamics (MD) simulations have emerged as a transformative computational microscope for probing atomic interactions spanning catalysis, energy storage, biotechnology, and beyond. However, existing machine-learning MD (MLMD) frameworks face a trilemma in balancing accuracy, scalability, and energy efficiency, particularly in compositionally complex systems like high-entropy alloys and multiferroic perovskites. Here, we introduce NVNMD-v2, a co-designed algorithm-hardware architecture that integrates a generalized deep neural-network potential (GDNNP) within a processing-in-memory (PIM) accelerator. Building on the foundation of NVNMD-v1, which was limited to four-element systems, NVNMD-v2 employs optimized type-embedding descriptors to support multielement systems with up to 32 species, eliminating species-dependent parameter scaling. Deployed on a single FPGA, NVNMD-v2 maintains DFT-level accuracy while achieving a flat per- atom computational cost (similar to 10-7 s/step/atom), enabling simulations of system up to 20 million atoms-a 103 x scale-up over DeePMD on an NVIDIA V100 GPU, with similar to 120 x energy reduction. These advances unlock quantum-accurate MD for multielement materials, from semiconductor heterostructures to biomolecular assemblies, bridging the gap between atomic fidelity and industrial-scale simulations.

Return to Publications page