Expanding frontiers of complex reaction network exploration through a general reactive machine learning potential

Developing general machine learning potentials (MLPs) for complex reactive systems remains a fundamental challenge, due to insufficient sampling of critical transition states (TSs) and non-equilibrium structures in training datasets. Here, we introduce an MD/CD-AL framework integrating molecular dynamics/coordinate driving (MD/CD) method with active learning to generate the MDCD20 dataset that encompasses diverse reactive configurations with up to 20 heavy atoms. With this dataset, we develop a general reactive MLP with quantum mechanics (QM) accuracy, MDCD-NN, for H-/C-/N-/O-containing gas-phase reactions. Leveraging MD/CD searching, MDCD-NN enables automated and efficient construction of complex reaction networks with 10 ⁴ -fold acceleration compared to the reference QM method, as demonstrated in three real-world reaction systems featuring intricate regio-/stereo-selectivities. In each case, our approach successfully identifies thousands of intermediates and TSs, constructing multistep networks that rationalize experimental observations and extend established mechanisms. Furthermore, MDCD-NN enables nanosecond-scale dynamics simulations for calculating reactive free energy surface, bridging quantum-level accuracy with scalable simulations. Our framework provides a paradigm for developing general reactive MLPs, enabling high-throughput mechanistic insights for complex chemical systems.