Molecular force fields with gradient-domain machine learning (GDML): Comparison and synergies with classical force...

Area of Science:

• Computational chemistry
• Molecular modeling
• Materials science

Background:

• Modern machine learning force fields (ML-FF) achieve high accuracy comparable to ab initio methods at reduced computational cost.
• Classical molecular mechanics force fields (MM-FF) are faster and transferable but less accurate due to fixed functional forms.

Purpose of the Study:

• To investigate the complementary nature of ML-FF and MM-FF.
• To understand the differences between ML-FF and MM-FF by analyzing their ability to reconstruct dynamic and thermodynamic observables.
• To enhance the accuracy of MM-FFs by reparametrizing interactions.

Main Methods:

• Contrasting ML-FF and MM-FF performance in reconstructing dynamic and thermodynamic observables.
• Qualitative analysis of the differences between ML-FF and MM-FF approaches.
• Reparametrization of short-range and bonded interactions in the generalized AMBER force field (GAFF).

Main Results:

• ML-FFs provide accurate predictions of energy and forces, similar to high-level ab initio methods.
• MM-FFs demonstrate faster computation and better transferability within molecular classes.
• The comparative analysis provided insights for improving MM-FF accuracy.

Conclusions:

• Combining insights from ML-FF and MM-FF can lead to more accurate classical force fields.
• Reparametrizing GAFF with more expressive terms enhances accuracy without compromising MM-FF advantages.
• This work paves the way for developing hybrid approaches in molecular simulations.