Ab initio machine learning of phase space averages.

Area of Science:

• Computational chemistry
• Materials science
• Biochemistry

Background:

• Equilibrium structures are crucial for material properties and biochemical functions.
• Conventional methods like molecular dynamics (MD) and Monte Carlo (MC) simulations are used to determine these structures.
• These simulations can be computationally intensive and time-consuming.

Purpose of the Study:

• To develop a machine learning approach for predicting phase space averages, which represent equilibrium structures.
• To enable a general machine learning pathway for obtaining ensemble properties across diverse chemical compounds.
• To accelerate the exploration of chemical compound space for material and biochemical applications.

Main Methods:

• Developed an ab initio machine learning (AIML) model that does not require bond topologies.
• Trained the AIML model on hundreds of MD trajectories to predict Boltzmann averaged structures.
• Used AIML output to train machine learning models for predicting free energies of solvation.

Main Results:

• AIML successfully predicts Boltzmann averaged structures.
• Machine learning models trained on AIML output achieve competitive prediction errors (MAE ~0.8 kcal/mol) for solvation free energies.
• The AIML approach bypasses the need for traditional MD or MC simulations for phase space sampling.

Conclusions:

• AIML offers a significantly accelerated pathway for exploring ensemble properties and chemical compound space.
• This method enables rapid prediction of material properties and biochemical functions.
• AIML provides a Pareto-efficient approach for solvation free energy predictions in terms of accuracy and computational time.