A data-efficient reactive machine learning potential to accelerate automated exploration of complex reaction networks

Reactive machine learning potentials (MLPs) significantly benefits high-throughput exploration of complex reaction networks. However, the development of such MLPs is currently constrained by the scarcity of datasets generated through efficient sampling methods that adequately capture elusive non-equilibrium configurations. To address this, we introduce an integrated molecular dynamics/coordinate driving-active learning (MD/CD-AL) framework that strategically samples reactive configurations, yielding a compact but comprehensive dataset, MDCD20, with ~1.4 million neutral and radical H/C/N/O structures. The MDCD-NN MLP trained on MDCD20 surpasses existing available MLPs trained on much larger datasets in reconstructing 181 elementary reactions with widespread types. It also accelerates automatic reaction network exploration by 10^4-fold in real-world systems relative to its reference quantum chemistry (QM) calculations, tackling complex scenarios such as multiple reaction centers, enantioselectivity and dynamic free-energy landscapes that are beyond the reach of traditional QM methods. This work establishes an efficient paradigm for constructing reactive datasets, enabling the training of reliable MLPs for complex reactive systems at an affordable cost.