In the era of artificial intelligence (AI), machine learning interatomic potentials (MLIPs) have revolutionized materials science and engineering, enabling large-scale and accurate atomistic modeling in various chemical systems. The performance of these potentials hinges on two key components: the model architecture, which defines the mathematical representation of atomic interactions, and the training data set, which supplies foundational knowledge. To scrutinize recent advances and chart the course for future development, this review presents a comprehensive analysis of reactive MLIPs, focusing on these two pillars. We first explore the evolution of model architectures, from descriptor-based models to state-of-the-art equivariant graph neural networks, critically assessing the essential requisites for achieving both physical accuracy and computational efficiency. Concurrently, we examine the sophisticated data acquisition strategies required to capture the complexities of chemical reactions, with a particular focus on uncertainty-driven active learning for sampling high-energy transition states and reaction pathways. Finally, beyond reviewing current methodologies, we provide an outlook on domain-specific challenges and discuss emerging opportunities in next-generation technologies, such as generative AI and cognitive autonomous agents. We believe this review will serve as a valuable guideline for practitioners in selecting and constructing MLIPs tailored for the study of chemical reactivity.
Kim et al. (Wed,) studied this question.