Δ-Learning applied to coarse-grained homogeneous liquids

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Coarse-grained molecular dynamics (CGMD) is crucial for simulating large-scale chemical and material systems.
• Current CGMD methods are often bespoke, lacking the accessibility of black-box tools like density functional theory.
• Machine learning (ML) potentials offer a path to simplify CGMD model development but are still nascent.

Purpose of the Study:

• To investigate the efficacy of delta-learning models in improving CGMD simulations.
• To leverage the strengths of both physics-based and ML-based approaches in CGMD.
• To benchmark delta-learning models against ML-only and physics-based models.

Main Methods:

• Implemented delta-learning models that use ML to learn the residual difference between target CGMD variables and physics-based potential predictions.
• Benchmarked delta-models against ML-only and elementary pairwise models.
• Evaluated performance based on atomistic property reproduction across varying CG resolutions, mapping operators, and system topologies.

Main Results:

• Delta-learning models consistently outperformed ML-only CGMD models across various scenarios.
• ML-only models sometimes produced qualitatively incorrect dynamics despite minimizing training errors, an issue corrected by delta-models.
• Unexpectedly, neither delta-learning nor ML-only models significantly outperformed elementary pairwise models in reproducing atomistic properties.

Conclusions:

• Delta-learning models offer significant, low-cost improvements over ML-only CGMD approaches.
• The fundamental limitation in reproducing atomistic properties stems from irreducible force errors inherent in coarse-graining.
• Further research is needed to overcome coarse-graining inaccuracies for more complex potentials to show a clear advantage.