Solvation Free Energies from Machine Learning Molecular Dynamics.

Area of Science:

• Computational Chemistry
• Theoretical Chemistry
• Physical Chemistry

Background:

• Calculating ion solvation free energies is crucial for understanding chemical processes.
• First-principles molecular dynamics (FP-MD) offers accurate but computationally expensive methods.
• Existing models require significant computational resources, limiting simulation times and convergence.

Purpose of the Study:

• To extend a hybrid solvation model for calculating ion solvation free energies.
• To improve the efficiency and convergence of FP-MD simulations.
• To achieve chemical accuracy in solvation energy predictions at reduced computational cost.

Main Methods:

• Re-expressing the solvation approach within quasi-chemical theory.
• Training a machine-learned (ML) energy function on FP energies and forces.
• Performing ML-assisted MD simulations with adjusted workflows and durations (approx. 200 ps).

Main Results:

• The extended approach successfully calculates ion solvation free energies.
• ML-assisted MD simulations achieve convergence within chemical accuracy (0.04 eV).
• The method provides accurate results for alkaline and alkaline-earth ions.

Conclusions:

• The proposed extension offers a computationally efficient route to accurate ion solvation free energies.
• ML integration significantly reduces the cost of FP-MD simulations.
• This approach enables reliable predictions for ion solvation relevant to chemical and physical processes.