Solvation entropy, enthalpy and free energy prediction using a multi-task deep learning functional in 1D-RISM

Area of Science:

• Computational chemistry
• Molecular modeling
• Physical chemistry

Background:

• Calculating solvation thermodynamics (entropies, enthalpies, free energies) is challenging using single simulation methods.
• Integral Equation Theory, specifically Reference Interaction Site Model (RISM), often yields significant errors in solvation thermodynamics.
• Previous work introduced pyRISM-CNN, combining 1D-RISM with deep learning for improved solvation free energy prediction.

Purpose of the Study:

• To report advancements in the pyRISM-CNN methodology for predicting solvation thermodynamics.
• To introduce solvation free energy predictions for organic ions in methanol and water.
• To expand the training dataset and apply pyRISM-CNN for simultaneous prediction of solvation enthalpy, entropy, and free energy.

Main Methods:

• Utilized the pyRISM solver with a deep learning free energy functional (pyRISM-CNN).
• Expanded the training data to include methanol as a solvent, alongside carbon tetrachloride, water, and chloroform.
• Employed a multi-task learning approach for simultaneous prediction of solvation enthalpy, entropy, and free energy.

Main Results:

• Achieved prediction errors below 4 kcal mol⁻¹ for solvation free energies of organic ions in methanol/water without additional descriptors.
• Obtained prediction errors near or below 1 kcal mol⁻¹ for neutral solutes in organic solvents and water across various temperatures.
• Demonstrated simultaneous prediction of solvation enthalpy, entropy, and free energy with errors of 1.04, 0.98, and 0.47 kcal mol⁻¹, respectively, in water at 298 K.

Conclusions:

• The enhanced pyRISM-CNN methodology offers a significant improvement in the accuracy of solvation thermodynamics predictions.
• The model demonstrates versatility in handling diverse solutes (ions, neutral molecules) and solvents (water, methanol, organic).
• The multi-task learning approach successfully enables simultaneous prediction of key solvation properties, addressing a long-standing challenge in computational chemistry.