Assessment and Validation of Machine Learning Methods for Predicting Molecular Atomization Energies

Area of Science:

• Computational chemistry
• Quantum chemistry
• Machine learning

Background:

• Accurate prediction of molecular properties often relies on computationally expensive quantum-chemical calculations.
• Machine learning (ML) offers a potential solution for efficient property prediction.

Purpose of the Study:

• To outline established machine learning techniques for molecular property prediction.
• To investigate the impact of molecular representation on ML model performance.

Main Methods:

• Application of various machine learning techniques to ab initio calculations.
• Evaluation of different molecular representations for their influence on prediction accuracy.

Main Results:

• The best performing ML methods achieved prediction errors as low as 3 kcal/mol for molecular atomization energies.
• Significant influence of molecular representation on the performance of ML methods was observed.

Conclusions:

• Machine learning provides an efficient approach to predict quantum-mechanical observables, complementing traditional methods.
• Careful selection of molecular representation is crucial for achieving high accuracy in ML-based property prediction.