Chemical reaction dynamics, critical to advancing technologies in energy, environment, and materials, can in principle be captured via reactive molecular dynamics (RMD). Time-series machine learning algorithms enable prediction of species evolution information hidden in these simulations, but long-horizon autoregressive accuracy often degrades due to error accumulation, and generating sufficiently large and diverse RMD data sets across operating conditions is computationally expensive. Here, we propose DEAL-TCN (deep-ensemble active learning with temporal convolutional networks), an active-learning framework that effectively disentangles the high-dimensional complexity of reaction dynamics spanning species, time, and operating conditions. It adopts the query-by-committee strategy to select informative conditions and leverages one-dimensional temporal convolutions to capture interspecies and long timescale couplings, enabling efficient modeling and long-term prediction of species evolution. In a prototypical case study of Mo-O-S precursors, DEAL-TCN robustly identifies and accurately predicts the concentration evolution of all involved chemical species across orthogonally designed test sets spanning a broad parameter range (1100-1500 K, 2-6 atm, and a feed ratio of 1/25-1/15). Given only the first 50 ps of each trajectory as input, the model attains a mean prediction error of 4.8% at the picosecond level while maintaining a mean error of 18.2% over the subsequent 0.45 ns, representing improvements of ∼29% and 24% over baseline LSTM and Transformer architectures, respectively. Meanwhile, DEAL-TCN outperforms random sampling in 98.8% of active-learning iterations with the same labeling budget. These results underscore DEAL-TCN's potential as a scalable and generalizable approach for mechanistic discovery, reaction design, and optimization.
Dang et al. (Mon,) studied this question.