Bond Type Restricted Property Weighted Radial Distribution Functions for Accurate Machine Learning Prediction of...

Area of Science:

• Computational Chemistry
• Machine Learning
• Materials Science

Background:

• Understanding machine learning algorithm performance is crucial for developing accurate statistical models.
• Neural network interpretability remains a challenge in many applications.
• Accurate prediction of molecular properties is vital for chemical research and development.

Purpose of the Study:

• To develop a method for calculating machine learning kernels in closed and analytic form.
• To improve the performance of the Bag-of-Bonds descriptor by incorporating bond type information into AP-RDF.
• To enhance the prediction accuracy of molecular atomization energies.

Main Methods:

• Combined atomic property weighted radial distribution function (AP-RDF) descriptor with a Gaussian kernel.
• Incorporated a tensor product into the kernel to capture multidimensional AP-RDF representations.
• Analyzed the performance using the QM7 dataset for molecular atomization energy prediction.

Main Results:

• Achieved improved prediction of molecular atomization energies (MAE = 1.7 kcal/mol on QM7 dataset).
• Demonstrated that the tensor product incorporation captures the multidimensional nature of AP-RDF.
• The numerical AP-RDF descriptor offers computational efficiency compared to Bag-of-Bonds.

Conclusions:

• The proposed method provides accurate and computationally efficient machine learning kernels.
• The incorporation of tensor products significantly enhances predictive performance.
• Established a link between molecular quantum similarity and machine learning kernels using first-principles descriptors.