In this enlightening episode, we welcome Dr. Cheng, a leading researcher in the field of machine learning (ML). We explore the innovative approach of ML in representing molecular-orbital-based (MOB) features for predicting post-Hartree–Fock correlation energies. Though previous applications of MOB-ML using Gaussian Process Regression (GPR) have shown promise, Dr. Cheng addresses the limitations of this method, particularly its computational constraints, when dealing with large datasets.

Dr. Cheng introduces us to an advanced approach, employing a clustering/regression/classification model of MOB-ML. He elaborates on the three-step process: partitioning the training data using regression clustering (RC), independently regressing each cluster using either linear regression (LR) or GPR, and training a random forest classifier (RFC) for predicting cluster assignments based on MOB feature values. Dr. Cheng's explanation underscores how this method dramatically reduces computational time while maintaining prediction accuracy.

This episode offers fascinating insights into how this pioneering approach can reach chemical accuracy with limited training data, demonstrating significant speed increases compared to traditional MOB-ML techniques. Moreover, we learn how the developed models can predict large-molecule energies accurately, even when trained only on small-molecule data.

Join us as we delve into this captivating topic, shedding light on the intersection of chemistry and ML and the transformative potential these methodologies hold for the future of scientific computation.

Keywords: Machine Learning, ML, Molecular-Orbital-Based Features, MOB-ML, Gaussian Process Regression, GPR, Regression Clustering, RC, Linear Regression, LR, Random Forest Classifier, RFC, Computational Chemistry, Post-Hartree–Fock Correlation Energies.

https://pubs.acs.org/doi/abs/10.1021/acs.jctc.9b00884