Active delta-learning for fast construction of interatomic potentials and stable molecular dynamics simulations

Abstract

Active learning (AL) requires massive time for comprehensive sampling of complex potential energy surfaces to achieve desirable accuracy and stability of machine learning (ML) potentials. Here, we develop an active delta-learning (ADL) protocol for speeding up AL and building delta-learning models yielding stable simulations. ADL converges after a few iterations and needs tenfold fewer sampled points than without delta-learning while leading to models of similar accuracy, as we show on the test simulations of Diels–Alder reactions. The test reactions include one small (ethene + 1,3-butadiene) and one relatively big (C _60 + 2,3-dimethyl-1,3-butadiene) system, treated with a target density functional theory level (U)B3LYP(-D4)/6-31G* and a baseline semi-empirical quantum mechanical method, GFN2-xTB. The crucial advantage of the models built with the delta-learning protocol is their remarkable simulation stability: even models from the initial ADL iterations yield reasonable results. In contrast, the pure ML potentials built without delta-learning often lead to the collapse in simulations, i.e. to unphysical structures.

Keywords

• machine learning interatomic potentials
• delta learning
• active learning
• chemical reactions
• molecular dynamics
• transition state