Accuracy Gains from Fully All-Atom NNPs for Large Systems

Ascertain whether simulating entire biomolecular systems at an all-atom level, including solvent, using neural network potentials would yield a significant accuracy improvement over classical molecular mechanics for large systems, given current impractical runtime compared to classical methods.

Background

While the ultimate goal is end-to-end all-atom simulations using neural network potentials with quantum-chemistry-level accuracy, current implementations cannot match the speed of classical force fields at the scale of entire solvated biomolecular systems. Beyond the computational barrier, it remains unknown whether such all-atom NNP simulations would deliver materially better accuracy for large systems.