Kinematics Modeling of Peroxy Free Radicals: A Deep Reinforcement Learning Approach

TL;DR

The paper tackles the scarcity of kinetic data for reactions of peroxy free radicals with NO, which governs tropospheric ozone dynamics, by predicting rate constants $k$ using a data-efficient deep reinforcement learning framework trained on 51 global descriptors from $G3B3$-optimized geometries. It compares neural regression, representation learning, and DRL-based classification, finding that DRL provides accurate kinetics ranges and enables descriptor-based interpretability via Integrated Gradients. Key contributions include demonstrating data-efficient learning on a small dataset, uncovering descriptor trends (e.g., $C1SP3$, $C2SP3$, PPSA1, MW) that correlate with $k$, and validating results against literature trends. The approach offers a scalable, interpretable tool for atmospheric chemistry that can be extended to related environmental and systems biology kinetics problems.

Abstract

Tropospheric ozone, known as a concerning air pollutant, has been associated with health issues including asthma, bronchitis, and impaired lung function. The rates at which peroxy radicals react with NO play a critical role in the overall formation and depletion of tropospheric ozone. However, obtaining comprehensive kinetic data for these reactions remains challenging. Traditional approaches to determine rate constants are costly and technically intricate. Fortunately, the emergence of machine learning-based models offers a less resource and time-intensive alternative for acquiring kinetics information. In this study, we leveraged deep reinforcement learning to predict ranges of rate constants (\textit{k}) with exceptional accuracy, achieving a testing set accuracy of 100%. To analyze reactivity trends based on the molecular structure of peroxy radicals, we employed 51 global descriptors as input parameters. These descriptors were derived from optimized minimum energy geometries of peroxy radicals using the quantum composite G3B3 method. Through the application of Integrated Gradients (IGs), we gained valuable insights into the significance of the various descriptors in relation to reaction rates. We successfully validated and contextualized our findings by conducting cross-comparisons with established trends in the existing literature. These results establish a solid foundation for pioneering advancements in chemistry, where computer analysis serves as an inspirational source driving innovation.

Figures (12)

• Figure 1: Architecture of the Neural Network used for continuous or discrete predictions.
• Figure 2: Siamese neural network architecture used in the representation learning framework.
• Figure 3: The simplistic Markov Decision Process used, illustrating the state-action interplay.
• Figure 4: The Multi-Input neural network architecture used for the Deep Reinforcement Learning setting. The two components of the state are initially processed in separate branches to ensure compatibility.
• Figure 5: Average confusion matrices (ceiling integer value) obtained for corresponding models on full dataset evaluation, post model training in each proposed setting.