Nat. Commun. | 物理基线与深度残差学习协同：ResFF兼顾分子力场精度与泛化能力
Ranking
Overall
75
Content
90
Popularity
40
Observed public metrics from 1 member.
Merged summary
TL;DR - A Nature Communications paper introduces ResFF, a molecular force field combining an interpretable molecular-mechanics baseline with an equivariant neural network that learns residual quantum-chemistry corrections. The approach improves accuracy and generalization while remaining stable in molecular dynamics simulations.
- ResFF uses staged training of its physics-based and neural modules, followed by joint fine-tuning.
- On molecule-level held-out tests, it achieved MAEs of 1.16 kcal/mol on Gen2-Opt and 0.90 kcal/mol on DES370K.
- It performed strongly on torsional potentials, intermolecular interactions, and conformer optimization, including chemically dissimilar and highly flexible molecules.
- Demonstrated limitations include local cutoffs and no explicit treatment of long-range electrostatics or polarization.
Sources (1)
Nat. Commun. | 物理基线与深度残差学习协同：ResFF兼顾分子力场精度与泛化能力
Public signals
OpenAlex citations 0
TL;DR - A Nature Communications paper introduces ResFF, a molecular force field combining an interpretable molecular-mechanics baseline with an equivariant neural network that learns residual quantum-chemistry corrections. The approach improves accuracy and generalization while remaining stable in molecular dynamics simulations.
- ResFF uses staged training of its physics-based and neural modules, followed by joint fine-tuning.
- On molecule-level held-out tests, it achieved MAEs of 1.16 kcal/mol on Gen2-Opt and 0.90 kcal/mol on DES370K.
- It performed strongly on torsional potentials, intermolecular interactions, and conformer optimization, including chemically dissimilar and highly flexible molecules.
- Demonstrated limitations include local cutoffs and no explicit treatment of long-range electrostatics or polarization.