Regression Clustering for Improved Accuracy and Training Costs with Molecular-Orbital-Based Machine Learning

Area of Science:

• Computational Chemistry
• Machine Learning in Quantum Chemistry

Background:

• Machine learning (ML) applied to molecular-orbital-based (MOB) features accurately predicts post-Hartree-Fock correlation energies.
• Gaussian Process Regression (GPR) is effective for small datasets but computationally expensive for large ones due to cubic scaling.

Purpose of the Study:

• To develop a more computationally efficient ML approach for predicting molecular energies.
• To overcome the training time bottleneck of GPR in MOB-ML for large datasets.

Main Methods:

• Introduced a novel regression clustering (RC) implementation for MOB-ML.
• Combined RC with linear regression (LR) or GPR for cluster-specific regression and a random forest classifier (RFC) for cluster assignment.
• Evaluated RC/LR/RFC and RC/GPR/RFC on the QM7b-T dataset of organic molecules.

Main Results:

• RC identified chemically intuitive groupings of frontier molecular orbitals.
• Both RC/LR/RFC and RC/GPR/RFC achieved chemical accuracy (1 kcal/mol error) with significantly reduced training times (35,000x and 4,500x faster, respectively).
• Models demonstrated transferability to larger molecules using small-molecule training data, with further accuracy improvements by capping cluster data points.

Conclusions:

• The RC-based MOB-ML approach offers a substantial speedup in training time without sacrificing prediction accuracy.
• This method provides a scalable and efficient alternative for predicting molecular energies, applicable to large datasets and transferable to different molecular sizes.