Beschreibung
Accurate process models are crucial for designing and operating chemical and biochemical processes. Traditional approaches rely solely on mechanistic or data-driven modeling, whereas gray-box models combine physical knowledge with machine learning (ML). This thesis addresses two key challenges: systematic dynamic gray-box modeling with embedded ML models and efficient surrogate modeling of thermodynamic equilibria for simulation and optimization. First, a step-by-step embedded gray-box methodology is introduced. The models are based on standard relationships such as heat and mass balances, while unknown effects are captured by embedded ML submodels. Training data for these submodels are generated by fitting piecewise trajectories to measurements, enabling quick testing of model architectures on static data and avoiding repeated dynamic simulations during identification. The method is demonstrated on three examples, including a real fermentation process, and yields robust models for computing efficient operating strategies. Second, surrogate modeling techniques are developed to replace computationally expensive phase-equilibrium (pT-flash) calculations in thermomorphic solvent systems. The three approaches Sobolev training, tie-line-based data augmentation, and adaptive sampling significantly reduce the required thermodynamic evaluations while maintaining global accuracy. A Bayesian optimization algorithm embedding these surrogates in gray-box formulations further increases efficiency by requiring fewer costly function evaluations. Overall, the presented methodologies offer new strategies for seamlessly integrating physical knowledge with data-driven models, enabling robust, accurate, and computationally efficient process simulation and optimization.
Herstellerkennzeichnung:
Shaker Verlag GmbH
Am Langen Graben 15a
52353 Düren
DE
E-Mail: info@shaker.de
