Source-linked AI summary
Machine Learning for Quantum Mechanical Properties of Atoms in Molecules
Matthias Rupp, Raghunathan Ramakrishnan, O. Anatole von Lilienfeld
TL;DR
The work addresses the high cost and limited transferability of machine-learning models for quantum-mechanical molecular properties. It models atomic observables using local atom-centered representations and demonstrates DFT-level accuracy with linear scaling and transferability to larger, locally similar systems.
Problem
Quantum-mechanical calculations are prohibitively costly for routine large-system modeling, while existing machine-learning models can lack transferability to larger molecules.
Method
The approach uses local atom-centered coordinate representations within nonlinear regression to model quantum-mechanical properties of atoms in molecules.
Results
Predictions for chemical shifts, core-level ionization energies, and atomic forces achieve accuracy on par with the quantum-mechanical reference, while prediction cost scales linearly with system size.
Takeaways & Limitations
Local atomic environments support transferability to larger molecules with similar building blocks, including polymers much larger than those used for training.
Takeaways & Limitations
The finite-cutoff locality assumption may fail for some systems and properties affected by long-range substituent effects, requiring larger cutoffs or additional measures.
Abstract
from arXiv · showhide
We introduce machine learning models of quantum mechanical observables of atoms in molecules. Instant out-of-sample predictions for proton and carbon nuclear chemical shifts, atomic core level excitations, and forces on atoms reach accuracies on par with density functional theory reference. Locality is exploited within non-linear regression via local atom-centered coordinate systems. The approach is validated on a diverse set of 9k small organic molecules. Linear scaling of computational cost in system size is demonstrated for saturated polymers with up to sub-mesoscale lengths.