Uncertainty-Driven Deep-Ensemble Temporal Convolutional Networks for Predicting Chemical Reaction Dynamics.

Area of Science:

• Scientific computing
• Chemical engineering
• Materials science

Background:

• Chemical reaction dynamics are crucial for energy, environment, and materials technologies.
• Reactive molecular dynamics (RMD) simulations capture these dynamics but face challenges in data generation and prediction accuracy.
• Time-series machine learning models struggle with error accumulation in long-horizon predictions.

Purpose of the Study:

• To introduce DEAL-TCN (deep-ensemble active learning with temporal convolutional networks), an active-learning framework designed to predict chemical species evolution.
• To effectively handle the high-dimensional complexity of reaction dynamics across species, time, and operating conditions.
• To enable efficient modeling and accurate long-term prediction of species evolution in RMD simulations.

Main Methods:

• DEAL-TCN employs a query-by-committee strategy for selecting informative simulation conditions.
• It utilizes one-dimensional temporal convolutions to model interspecies and long-timescale couplings.
• The framework integrates deep ensemble learning with active learning for efficient data selection and model training.

Main Results:

• DEAL-TCN accurately predicts chemical species concentration evolution for Mo-O-S precursors across a wide parameter range.
• The model achieves a mean prediction error of 4.8% at the picosecond level and 18.2% over 0.45 ns, outperforming LSTM and Transformer architectures.
• DEAL-TCN significantly outperforms random sampling in active-learning efficiency, improving results in 98.8% of iterations.

Conclusions:

• DEAL-TCN offers a scalable and generalizable approach for mechanistic discovery in chemical reactions.
• The framework enhances reaction design and optimization by providing accurate long-term predictions.
• This method addresses key limitations in RMD data generation and machine learning prediction accuracy.