Synergistic fusion of physical modeling and data-driven approaches for parameter inference to enzymatic biodiesel production system

• Published in: Applied energy 2024-11, Vol.373, p.123874, Article 123874
• Main Authors: Chen, Yu, Yu, Hui, Liu, Chengjie, Xie, Jin, Han, Jun, Dai, Houde
• Format: Article
• Language: English
• Subjects: biodiesel; Biodiesel production; Data-driven; energy; Multiscale; Parameter inference; Physics-informed; Time causality
• ISSN: 0306-2619
• DOI: 10.1016/j.apenergy.2024.123874

Enzymatic biodiesel production systems represent complex chemical dynamical processes that traditional modeling approaches struggle to effectively capture the inherent nonlinearity and dynamics. We propose a novel framework that fuses prior physical knowledge and data-driven methods to enhance parameter inference capabilities for such systems. Grounded in the concept of physics-informed machine learning (PIML), we combine the prior knowledge from traditional physical models with the powerful approximation abilities of data-driven neural networks. This framework enables the physical laws to guide the training of the data-driven model while utilizing data to correct model biases, achieving tight integration of physical constraints and data-driven approaches in the modeling process. To address potential challenges faced by PIML methods in practical applications, such as multiscale features and violations of time causality, we introduce several improved strategies. Evaluating our framework on real experimental data from biodiesel production, we validate its accurate estimation of system parameters. This work provides a promising new approach to enhance the interpretability and generalization of data-driven modeling by incorporating prior physical knowledge, holding potential applications in modeling and control of complex dynamical systems. •A comprehensive framework is proposed that integrates physical constraints into a data-driven approach for parameter inferences of dynamic systems.•Hard constraints and Fourier features layers are adopted to optimize the initial boundary condition and solve the causality caused by spectral bias.•A bootstrap-based uncertainty quantification method is proposed to calculate the uncertainty of the outputs and model parameters.•A lab-scale enzymatic biodiesel production system is investigated where the proposed framework achieves reliable parameter estimation.