Dual-Coordinate Relative Free Energy Simulations Using Machine-Learned Interatomic Potentials

Area of Science:

• Computational Chemistry
• Materials Science
• Chemical Physics

Background:

• Machine-learned interatomic potentials (MLIPs) offer near-quantum mechanical accuracy in simulations.
• Alchemical free energy simulations (FES) are computationally expensive and challenging with MLIPs due to unphysical intermediate states.
• Previous work enabled gradual decoupling of molecules in MLIP-described systems.

Purpose of the Study:

• To extend gradual decoupling methods for general alchemical transformations.
• To compute relative solvation free energies (RSFEs) between arbitrary solute pairs using unmodified MLIPs.
• To develop an efficient and adaptable method for MLIP-based FES.

Main Methods:

• A dual-coordinate approach with both solutes present but non-interacting.
• λ-dependent distance offsets in the neighbor list for simultaneous decoupling and coupling.
• Harmonic 'anchor' restraints to enforce spatial overlap without altering intramolecular dynamics.
• Utilized the MACE-OFF23(S) potential for calculations.

Main Results:

• Successfully computed RSFEs for diverse small molecule pairs, including dissimilar species.
• Achieved excellent internal consistency with thermodynamic cycle closure errors ≤ ±0.11 kcal/mol.
• Demonstrated robustness and accuracy of the anchor-based dual-coordinate method.

Conclusions:

• The anchor-based dual-coordinate approach is a highly adaptable and efficient route for relative alchemical FES with modern MLIPs.
• The method requires only single energy evaluations per step and avoids MLIP retraining.
• This technique significantly advances the application of MLIPs in computational chemistry for solvation free energy predictions.