Deep learning-based chemical kinetic modeling of SF₆ dissociation with trace water and oxygen

Sulfur hexafluoride (SF<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</inf>), widely used as an insulating gas in the power industry, exhibits complex decomposition behavior under trace H<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>O/O<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> conditions, necessitating chemical kinetic modeling for accurate analysis. However, the inherent stiffness of these models presents significant computational challenges. To address this, a Physics-Informed Neural Network (PINN)-based framework is proposed, where the input layer weights and hidden layer biases are randomly initialized. Only the output layer weights are optimized, thereby mitigating gradient imbalance issues in traditional neural networks. Additionally, a temporal domain decomposition strategy is introduced to handle multi-timescale concentration variations. After a Box-Cox transformation, the time domain is partitioned into multiple subdomains, with independent solutions computed and terminal values propagated as initial conditions to subsequent segments. Experimental results confirm that the proposed PINN architecture with domain decomposition effectively mitigates stiffness in SF<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</inf> decomposition systems under trace H<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf>O/O<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</inf> conditions. It accurately captures the temporal evolution of initial reactants, intermediates, and final products across varying timescales, achieving prediction accuracy comparable to reference solutions from the Backward Differentiation Formula (BDF) solver for stiff ordinary differential equations (ODEs).