Self-Evolving Machine: A Continuously Improving Model for Molecular Thermochemistry.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Chemical data collection is challenging, limiting data-driven model scope.
• Existing models struggle with broad chemical space coverage.

Purpose of the Study:

• To develop a self-evolving model for predicting formation enthalpies.
• To enhance the accuracy and applicability of computational chemistry models.

Main Methods:

• Integrated active learning with automatic ab initio calculations.
• Employed a molecular graph convolutional neural network with dropout.
• Modeled formation enthalpies of polycyclic species at B3LYP/6-31G(2df,p) level.

Main Results:

• Achieved a root-mean-square error of 2.62 kcal/mol for 2858 hydrocarbons and oxygenates.
• Demonstrated accurate prediction of density functional theory (DFT) enthalpies.
• Showcased model's ability to assess prediction quality.

Conclusions:

• The self-evolving model continuously adapts to new chemical species.
• Active learning expands model's predictive capacity to unseen domains, like nitrogen-containing species.
• Continuous learning improves model accuracy and broadens its applicability.