Source-linked AI summary
Understanding molecular representations in machine learning: The role of uniqueness and target similarity
Bing Huang, O. Anatole von Lilienfeld
TL;DR
Molecular ML accuracy depends on representations that balance uniqueness with similarity to the target. The paper develops BAML from interatomic many-body energy expansions and finds that higher-order terms improve prediction across molecular properties, while representation-specific uniqueness limits remain.
Problem
Molecular representations are often chosen through ad hoc trial and error, with no general rigorous procedure for systematically optimizing robust ML performance.
Method
The paper uses quantum-mechanical structure to prioritize unique representations and systematically varies target similarity through energy-based bags of bonds, angles, and torsions.
Results
Across nine properties and two molecular datasets, adding higher-order contributions systematically lowers learning-curve offsets; BAML reaches MAE ∼1 kcal/mol for C7H10O2 atomization energies with 5k molecules and MAE ∼2.4 kcal/mol for QM9 with 10k.
Takeaways & Limitations
BAML provides a unified representation that performs well for simple scalar global quantum-mechanical observables and supports fast, accurate out-of-sample prediction after training.
Takeaways & Limitations
Polarizability-based representations can have high target similarity but violate uniqueness because different molecular geometries may share the same atomic-volume representation.
Abstract
from arXiv · showhide
The predictive accuracy of Machine Learning (ML) models of molecular properties depends on the choice of the molecular representation. Based on the postulates of quantum mechanics, we introduce a hierarchy of representations which meet uniqueness and target similarity criteria. To systematically control target similarity, we rely on interatomic many body expansions, as implemented in universal force-fields, including Bonding, Angular, and higher order terms (BA). Addition of higher order contributions systematically increases similarity to the true potential energy and predictive accuracy of the resulting ML models. We report numerical evidence for the performance of BAML models trained on molecular properties pre-calculated at electron-correlated and density functional theory level of theory for thousands of small organic molecules. Properties studied include enthalpies and free energies of atomization, heatcapacity, zero-point vibrational energies, dipole-moment, polarizability, HOMO/LUMO energies and gap, ionization potential, electron affinity, and electronic excitations. After training, BAML predicts energies or electronic properties of out-of-sample molecules with unprecedented accuracy and speed.