MolE8: finding DFT potential energy surface minima values from force-field optimised organic molecules with new...

Area of Science:

• Computational Chemistry
• Machine Learning
• Quantum Mechanics

Background:

• Machine learning (ML) is increasingly used in computational chemistry due to the availability of large molecular databases.
• ML predictions of molecular properties offer a computationally cheaper alternative to traditional quantum mechanics calculations without sacrificing accuracy.

Purpose of the Study:

• To develop a novel, extrapolatable, and explainable molecular representation for training ML models.
• To enable accurate and rapid prediction of electronic and free energies for organic molecules.

Main Methods:

• A new molecular representation based on bonds, angles, and dihedrals was developed.
• Machine learning models were trained using this representation to predict molecular energies.
• The models were tested for accuracy and extrapolation capabilities on small and larger organic molecules.

Main Results:

• Trained models accurately predict electronic and free energies for small organic molecules (C, H, N, O) with a mean absolute error of 1.2 kcal mol⁻¹.
• Extrapolation to larger molecules (≤10 heavy atoms) yielded an average error < 3.7 kcal mol⁻¹.
• Energy predictions were up to 7 times faster than previous ML models, even from non-optimized geometries.

Conclusions:

• The developed molecular representation enables accurate and efficient ML-based energy predictions.
• The approach extends accurate predictions to larger chemical spaces beyond the training set.
• This method significantly accelerates computational chemistry workflows by enabling rapid energy calculations.