Prediction of atomization energy using graph kernel and active learning.

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Predicting molecular properties with data-driven methods faces challenges in data structure, dimensionality, symmetry, and confidence management.
• Accurate prediction of molecular properties is crucial for accelerating materials discovery and chemical process design.

Purpose of the Study:

• To develop a novel kernel-based pipeline for accurate prediction of molecular atomization energy.
• To address challenges in data structure, dimensionality, and confidence management in machine learning for molecular properties.

Main Methods:

• Utilized Gaussian process regression for predictions based on molecular similarity.
• Employed a marginalized graph kernel, converting molecules into graphs labeled by elements and interatomic distances.
• Derived efficient formulas for kernel evaluation and proposed specific functional components.

Main Results:

• The proposed method achieved a high accuracy in predicting atomization energy, with a mean absolute error of 0.62 ± 0.01 kcal/mol on the QM7 dataset.
• Demonstrated the suitability of the graph kernel for predicting extensive molecular properties due to its convolutional structure.
• Achieved high accuracy using only 2000 training samples via an active learning procedure.

Conclusions:

• The kernel-based pipeline, leveraging a marginalized graph kernel and Gaussian process regression, offers an accurate and data-efficient approach for molecular property prediction.
• The method effectively handles molecular data structure and provides reliable predictive confidence, advancing machine learning applications in chemistry.