AI-driven methodology to develop (hybrid) steady-state and dynamic process models

Many chemical engineering applications face the challenge of building accurate and interpretable models from sparse, noisy, and costly data. This need arises especially in early-stage process design or when experimenting with novel unit equipment, where data acquisition is limited and prior knowledge is incomplete. Classical regression methods often fail to provide physically meaningful models that generalize beyond specific datasets or operating conditions. To address this issue, a novel workflow for physics-informed symbolic regression is proposed, with the objective to integrate physical knowledge directly into the learning process, ensuring consistency with first-principles models while retaining flexibility for data-driven discovery. The framework enables model discovery that is both interpretable and accurate, even in the presence of noisy and incomplete information. Two application areas are explored. First, the algorithms reconstruct governing equations, algebraic or differential, purely from data, making them useful where no prior models exist. Second, hybrid workflows embed mechanistic models into the regression process, supporting either the discovery of missing phenomena or refinement of existing equations through physically meaningful corrections. A key feature is the ability to interpret and correct plant-model mismatch, transforming error correction into a tool for scientific discovery. This approach transforms error correction into an opportunity to expand engineering knowledge, allowing the model to evolve in a principled way that enhances understanding of the system rather than just improving predictive performance. The resulting models bridge data-driven and first-principles sources of knowledge, yielding equations that are both accurate and physically sound.