ABC-flash: Fast and Scalable D3 Three-Body Dispersion for Large Molecular and Periodic Systems
Year: 2026
Type: preprint
Abstract: Three-body dispersion influences molecular binding and structural stability, yet its computational and memory demands can hinder routine inclusion in large systems. Here, we introduce ABC-flash, a GPU-accelerated implementation of the D3 Axilrod–Teller–Muto (ATM) correction for molecular and periodic systems. Sparse pair tiles enable reuse of shared factors and on-the-fly accumulation of triplet energies and derivatives, retaining the established D3-ATM formulation and its coordination-number dependence without storing global triplet intermediates. Across 2,459 molecular structures from GMTKN55, maximum absolute energy and force-component discrepancies relative to reference calculations are 6.63 × 10−10 hartree and 4.05 × 10−11 hartree bohr−1 in the default GPU mode. On a single GPU, ATM energies and forces for a 105,422-atom molecular assembly require 0.834 s with 4.01 GiB of peak device memory; energies, forces, and stress for 98,304 periodic atoms require 3.05 s. For an approximately 10,000-atom molecular system, ABC-flash achieves a 417-fold speedup over a GPU implementation on the same device. Without parameter refitting, adding D3-ATM reduces protein–ligand interaction-energy RMSEs by 36–44% and improves large-complex and framework energetics, while crystal and ice results remain method-and observable-dependent. ABC-flash combines subsecond molecular evaluation with modest memory requirements, making explicit three-body dispersion practical alongside inexpensive energy models and for repeated evaluation of large structural datasets.
Source: ChemRxiv
Authors: Ilkwon Cho, Bhupalee Kalita, Olexandr Isayev
Institutions: Carnegie Mellon University, University of Pittsburgh
Language: English
Subfield: Atomic and Molecular Physics, and Optics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235868022
