Computing energy changes in chemical reactions more efficiently

Unveiling the underpinnings of a chemical reaction involves knowing how its molecules rearrange, which can alter a system’s energy. Quantum mechanical methods allow researchers to accurately calculate these changes, but their high computational costs limit the timescale of simulations.

Machine learning interatomic potentials (MLIPs) serve as an efficient alternative, learning from these calculations to predict energies. But in training, MLIPs may not be exposed to certain parts of the reactions, like short-lived, high-energy structures important for describing such phenomena.

Verma et al. address this issue by systematically constructing MLIPs that integrate an active learning (AL) framework, where an algorithm requests a quantum mechanical label for the most informative configuration, thus generating training data without an extensive prelabeled dataset.

The authors created an automated, open-source package known as Smart Potential with Atomistic Rare events and Continuous Learning (SPARC) that identifies “the most informative configurations on-the-fly to improve the model,” said author Rahul Verma.

Typical AL approaches generate diverse configurations by running several simulations. However, this process is expensive and often uninformative.

“[SPARC] brings the entire process into a singular workflow, from exploring the configurational space, and identifying informative configurations, to labeling them and improving the model,” said Verma.

It integrates the PLUgin for MolEcular Dynamics library to perform enhanced sampling methods in molecular dynamics, enabling the model to explore and learn a broader configurational space while more effectively utilizing expensive quantum mechanical calculations.

Using SPARC, the authors constructed stable and reactive machine learning models that used graph-based collective variables to explore configurational space rather than sampling along a predefined reaction pathway. However, these models could still be used to study reactions along specific pathways.

The authors hope the open-source framework enables greater accessibility to construct MLIPs. They intend to improve “configuration selection and how the system is explored during early stages of training … [or] combine SPARC with pretrained universal models, which could reduce the amount of new quantum mechanical data required,” Verma said.

Source: “A systematic framework for constructing reactive machine learning interatomic potentials,” by Rahul Verma, Nisarg Joshi, and Jim Pfaendtner, APL Machine Learning (2026). The article can be accessed at https://doi.org/10.1063/5.0344922 .