Översikt
- Datum:Startar 16 januari 2026, 14:00Slutar 16 januari 2026, 15:00
- Plats:Kemihuset, Forskarhus 1, vån 10, seminarierum 10:an.
- Opponent:Prof. Henrik Ström, Department of Mechanics and Maritime Sciences, Chalmers University of Technology, Sweden
- AvhandlingLäs avhandlingen (Öppnas i ny flik)
Traditional reactor modeling commonly relies on complete mechanistic knowledge, a requirement that in certain scenarios is deemed impractical due to incomplete kinetic models and complex non-idealities. This thesis addresses these limitations by incorporating mechanistic knowledge with data-driven modeling employing neural networks. Particularly, neural ordinary differential equations (neural ODEs) are adopted as a framework to learn the system dynamics. This framework is integrated with physics-informed constraints based on fundamental conservation laws. Two complementary strategies are developed: (i) identifying component mass balances and (ii) enforcing atom balance constraints. These principles are embedded as soft constraints during training and are shown to improve generalization capabilities and model robustness. These approaches are benchmarked against their data-driven counterparts and when relevant against a kinetic model.
Comprehensive evaluations using synthetic data demonstrate that physics-informed regularization improves generalization capabilities when data is scarce. The performance gap diminishes when the amount of available data increases, however the improved model robustness remains. When considering real-world data, employing physics-informed constraints demonstrates a decrease in epistemic uncertainty and improved predictive accuracy compared to a kinetic model. The incorporation of mechanistic knowledge in the form of soft constraints requires manual tuning of the regularization weights. This can be computationally demanding since a set of different regularization weights need to be tested. It is therefore proposed to integrate the regularization weights into the training process by reformulating the traditional optimization problem for training network weights by performing gradient ascent on the regularization weights. This automatically tunes the regularization weights and circumvents manual tuning. The proposed methods are validated and tested on both synthetic and real-world data, thus demonstrating their potential for real-world applications. By bridging domain knowledge and machine learning, this work establishes physics-informed neural ODEs as a viable tool for chemical engineers facing scarcely available data and limited system knowledge.
Comprehensive evaluations using synthetic data demonstrate that physics-informed regularization improves generalization capabilities when data is scarce. The performance gap diminishes when the amount of available data increases, however the improved model robustness remains. When considering real-world data, employing physics-informed constraints demonstrates a decrease in epistemic uncertainty and improved predictive accuracy compared to a kinetic model. The incorporation of mechanistic knowledge in the form of soft constraints requires manual tuning of the regularization weights. This can be computationally demanding since a set of different regularization weights need to be tested. It is therefore proposed to integrate the regularization weights into the training process by reformulating the traditional optimization problem for training network weights by performing gradient ascent on the regularization weights. This automatically tunes the regularization weights and circumvents manual tuning. The proposed methods are validated and tested on both synthetic and real-world data, thus demonstrating their potential for real-world applications. By bridging domain knowledge and machine learning, this work establishes physics-informed neural ODEs as a viable tool for chemical engineers facing scarcely available data and limited system knowledge.