Explicit, Machine-Learned Two-Body Potentials for Molecular Simulations

Abstract

A new pairwise hybrid machine-learning/molecular mechanics (ML/MM) potential is introduced that is conceived for application to large, heterogeneous condensed-phase systems. The PhysNet ML method describes monomers and short-range dimer interactions, while a classical MM force field describes pairwise interactions beyond a defined switching distance. Models are fitted to MP2 dimer and pairwise cluster energies, and the quality of each model is assessed at different switching distances and using MM approaches with and without detailed distributed charge electrostatics. The applicability of the approach to molecular dynamics simulations is demonstrated for a basic implementation applied to a small model system. Dichloromethane and acetone are used as test systems to demonstrate the accuracy of the approach in describing pairwise reference data, and also to highlight the limitations of the pairwise approach for systems that exhibit significant many-body effects in condensed phase, paving the way for the addition of a general many-body correction in future work.

Figures (16)

• Figure 1: Definitions of "Region 1" and "Region 2" necessary to model the ML/MM boundary.
• Figure 2: Training performance of Dichloromethane and Acetone evaluated on the test set. Panel A: Performance of the PhysNet model for DCM. Panel B: Performance of the PhysNet model for acetone. The green and red symbols are for monomer and dimer energies, respectively. For both models the correlation coefficient is $R^{2} = 0.99$.
• Figure 3: Correlation between reference electronic structure data ($x-$axis) and the MM (panels A and B) or ML (panel C) energies for DCM using "Approach A". In panels A to C, the green traces are results from using CGenFF, CGenFF+MDCM, and PhysNet energies without modification. The orange traces in panels A and B are the results after refitting the LJ-parameters but without ML-contributions. The performance of hybrid ML/MM models are the blue and red symbols and lines in panels A and B, respectively. Light and dark traces are for $r_{\rm cut} = 4$ Å and $r_{\rm cut} = 7$ Å, respectively. Note that for larger values of $r_{\rm cut}$, the mixed ML/MM models progressively approach the PhysNet model. Computationally more advantageous are ML/MM models with shorter $r_{\rm cut}$. For statistical performance measures, see Table \ref{['tab:app-a']}.
• Figure 4: Correlation between reference electronic structure data ($x-$axis) and the MM (panels A and B) or ML (panel C) energies for acetone using "Approach A". In panels A to C, the green traces are results from using CGenFF, CGenFF+MDCM, and PhysNet energies without modification. The orange traces in panels A and B are the results after refitting the LJ-parameters but without ML-contributions. The performance of hybrid ML/MM models are the blue and red symbols and lines in panels A and B, respectively. Light and dark traces are for $r_{\rm cut} = 4$ Å and $r_{\rm cut} = 7$ Å, respectively. Note that for larger values of $r_{\rm cut}$, the mixed ML/MM models progressively approach the PhysNet model. Computationally more advantageous are ML/MM models with shorter $r_{\rm cut}$. For statistical performance measures, see Table \ref{['tab:app-a']}.
• Figure 5: Performance of the unfitted dark (blue and red) and LJ-fitted light (blue and red) models for different values of the cutoff. Top panel for DCM, bottom panel for acetone. The standard deviations of the fits are comparable to the magnitude of the RMSE.