Energy-Geometry Dependency of Molecular Structures: A Multistep Machine Learning Approach

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accelerating in silico materials screening requires efficient methods for estimating quantum observables.
• Traditional computational methods for these calculations are often computationally expensive.

Purpose of the Study:

• To develop and validate a machine learning (ML) based multistep method for estimating total energy.
• To assess the ML method's accuracy for diverse chemical structures.
• To extend the ML approach for estimating experimental thermochemical properties.

Main Methods:

• A multistep machine learning algorithm was employed to estimate total energy using spatial coordinates and atomic charges.
• The method was tested on a database comprising organic molecules, inorganic molecules, and ions.
• Similar molecular representations were used to estimate experimental thermochemical properties, including heat capacity.

Main Results:

• The ML method achieved a root-mean-square error (RMSE) of 0.76 atomic units (au) and a mean absolute percent error (MAPE) of 1.5% for total energy calculations.
• For thermochemical properties, the ML approach yielded a MAPE as low as 6% and an RMSE of 8 cal/mol·K for heat capacity.
• The model demonstrated accuracy on both optimized and unoptimized chemical structures.

Conclusions:

• The proposed ML method offers a computationally efficient alternative for estimating quantum observables.
• This approach facilitates faster and more accessible in silico materials screening.
• The ML models show promise for predicting both electronic properties and thermochemical data.