DeePMO: An iterative deep learning framework for high-dimensional kinetic parameter optimization

This study introduces Deep learning-based kinetic model optimization (DeePMO), a novel approach for optimizing parameters in chemical kinetic models. The primary challenge lies in mapping high-dimensional kinetic parameters to comprehensive performance metrics derived from diverse numerical simulations, including ignition delay time, laminar flame speed, heat release rate, and temperature-residence time distributions in perfectly stirred reactors. We propose an iterative sampling-learning-inference strategy to efficiently explore high-dimensional parameter spaces. The approach features a hybrid deep neural network (DNN) architecture that combines a fully connected network for non-sequential data with a multi-grade network for sequential data, enabling effective utilization of performance metrics with varying distribution characteristics. DeePMO’s effectiveness and versatility was validated across multiple fuel models, including methane, ethane, butane, n-heptane, n-pentanol, ammonia, ammonia/hydrogen, and their mixtures, with parameter counts ranging from tens to hundreds. The validation demonstrated successful optimization in all test cases and confirmed the method’s flexibility in incorporating both direct experimental measurements and simulated data from benchmark chemistry models. An ablation study highlighted the critical role of DNN in guiding data sampling and optimization processes, while additional comparative experiments examined hyperparameter effects. This work provides a valuable tool for kinetic parameter optimization and offers insights for applying machine learning algorithms in combustion research. • Boosting optimization performance by iterative sampling-learning-inference strategy. • A novel hybrid deep neural network handles both sequential and non-sequential data. • Extensive validation and ablation studies confirm method’s versatility and robustness.