Theoretical and computational chemistry

Machine-learned potentials, virtual screening

Explore machine-learned interatomic potentials that approximate quantum-level energies and forces, active-learning workflows, and staged virtual screening of large molecular libraries.

1 Lower secondary 2 Upper secondary 3 Undergraduate 4 Advanced 5 Research

IntuitionFast approximations need a trust region

A learned interatomic potential replaces repeated expensive electronic-structure calculations with a model of energy and forces. Its predictions are useful only for configurations represented well by its training data; unfamiliar chemistry calls for new reference calculations.

The simulation contrasts uncertainty-driven sampling for a learned potential with a multistage funnel that reduces a library to a few costly, physics-refined candidates.

SchoolEnergy and force are linked

Definition: Interatomic potential

A potential-energy model maps atomic positions and elements to a scalar energy $E(\mathbf R)$ . Forces are its negative coordinate gradient, so energy and force predictions should be consistent.

\mathbf F_i=-\nabla_{\mathbf R_i}E(\mathbf R)

For molecular dynamics, small force errors can accumulate into incorrect structures or kinetics. Evaluate energies, forces, and relevant observables on configurations outside training—not only random samples drawn from the training trajectory.

UndergraduateSymmetry, locality, and active learning

Many atomistic models build invariance to translation, rotation, and permutation of identical atoms into their architecture. Active learning iterates: train on reference structures, find configurations where models disagree or uncertainty is high, compute new reference labels, and retrain.

\mathcal D_{n+1}=\mathcal D_n\cup\{(\mathbf R^*,E_{\mathrm{ref}},\mathbf F_{\mathrm{ref}}):\mathbf R^*\in\arg\max_{\mathbf R}u(\mathbf R)\}

Example: Select a query for labeling

An ensemble of models predicts nearly identical energies in equilibrium configurations but diverges strongly near a bond-breaking geometry. Which region should receive the next expensive reference calculation?

Solution

Query the high-disagreement bond-breaking region, provided it is physically relevant and the reference method is trustworthy there. Add the result and retrain, then test dynamics and observables independently.

AdvancedVirtual screening is a cascade, not a verdict

A screening cascade uses cheap models to triage many compounds, then applies more accurate scoring, conformational sampling, and ultimately experiments to a smaller set. Ranking is conditional on the target structure, scoring assumptions, protonation states, and sampling coverage.

ResearchFrontier: transferability and trustworthy screening

References

E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials · Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky, 2022
Machine learning for molecular and materials science · Keith T. Butler, Daniel W. Davies, Hugh Cartwright, Olexandr Isayev, and Aron Walsh, 2018