2025 AIChE Annual Meeting

(681g) Not Every Data-Driven Model Plays Nice: Why Structure Matters in Optimization.

Authors

James Rawlings, University of California, Santa Barbara
Rahul Bindlish, The Dow Chemical Company
Industries often require nonlinear process models to identify profitable steady-states for operation. These models can be developed using first-principles, such as conservation laws. However, a significant limitation is the incomplete process knowledge available to practitioners, often resulting in model mismatches between the true process dynamics and the estimated model. Maintaining these models is expensive, as their performance can degrade over time. Recent advances in machine learning have enabled data-driven models that eliminate the need for detailed process knowledge. These models excel at predictions and are increasingly applied in process control and monitoring (Sansana et al., 2021).

However, data-driven models are typically unsuitable for process optimization because they do not inherently follow conservation laws. While enforcing physical laws using approaches like PINNs can help, these laws are not naturally embedded in the models. Moreover, during optimization, the model takes steps based on its own predictions, with no feedback to correct errors, unlike in process control applications. This limitation can lead to entirely unrealistic optima. A promising approach to address these challenges is to approximate only the unknown components of the first-principles model using neural networks. This strategy naturally enforces physical laws. We refer to the extent of process knowledge embedded in these models as structure (Kumar and Rawlings, 2023).

In this work, we demonstrate the performance of structured models in optimizing the benchmark vinyl acetate monomer plant (Luyben and Tyréus, 1998). In the plant's first-principles model, we approximate the rate laws using neural networks, followed by standard system identification experiments to learn the rate laws. We then use the identified structured model to determine the plant's economic optimum and compare its performance with that of a purely data-driven model. Additionally, we discuss the data requirements and challenges associated with training structured models.

Our results show that both models predict the plant's behavior well. However, despite achieving a high-quality fit to the data, the data-driven model fails to find the economic optimum, while the structured model proves to be more reliable. We conclude that successful optimization requires not more data, but more structure. Our findings highlight the potential of structured models for practical industrial applications.

References:

J. Sansana, M. N. Joswiak, I. Castillo, Z. Wang, R. Rendall, L. H. Chiang, and M. S. Reis. Recent trends
on hybrid modeling for industry 4.0. Comput. Chem. Eng., 151:107365, 2021.

P. Kumar and J. B. Rawlings. Structured nonlinear process modeling using neural networks and application
to economic optimization. Comput. Chem. Eng., 177, 2023.

M. L. Luyben and B. D. Tyréus. An industrial design/control study for the vinyl acetate monomer process.
Comput. Chem. Eng., 22(7-8):867–877, 1998.