(753b) Revisiting Hybrid Modeling: Integrating Machine-Learning Models with Physical Constraints

Bradley, W. - Presenter, Georgia Institute of Technology
Data science has demonstrated resounding success in solving complex problems within ChemE disciplines from material discovery to reaction engineering. Process systems engineering (PSE) has likewise seen many uses of machine learning (ML) tools, from modeling unit operations to entire processes. However, practitioners remain hesitant to incorporate machine-learning models for PSE as their inherent black-box nature makes their interpretation difficult and their performance unreliable for predicting system behavior outside the range of data collected for model fitting. Such situations arise frequently in PSE with common examples including process scale-up, changes in feedstock, process deviation, equipment degradation, and design of experiments leading to process optimization.

Hybrid “semi-parametric” modeling offers a promising alternative to purely data-driven process models by merging the flexibility of data-driven models and the reliability of physical knowledge. [1, 2] By constraining the data-driven model to mechanistic constraints, several authors from disparate fields have shown the superior performance of hybrid models over purely data-driven models in terms of interpretability, extrapolation, and data-dependency. [3] Nevertheless, due to its interdisciplinary nature, hybrid modeling as a tool still remains a black-box to most process engineers with examples of incorporating hybrid modeling techniques scarce in industrial practice.

Recent and ongoing advances of ML methods in terms of availability (via open-source, user-friendly software) and power (computational efficiency) merits revisiting the hybrid modeling paradigm. Herein we examine classical approaches to hybrid modeling and employ open-source software environments which enable scalable construction and optimization with hybrid models. A workflow will be presented through which classical hybrid modeling frameworks can be constructed, solved and validated, with an emphasis on serial and parallel approaches as defined by von Stosch et al. [3]. The presentation will further emphasize the difference between approaches that simultaneously solve the data-driven and mechanistic model (“coupled solution”) vs. those approaches wherein the data-driven and mechanistic model are solved separately (“uncoupled solution”), a distinction not readily apparent in hybrid modeling literature.

Finally, we employ case studies in PSE to evaluate the performance of hybrid modeling frameworks with their data-driven and mechanistic counterparts. Specifically, studies will characterize metrics for model performance where data is scarce and noisy and mechanistic knowledge is available at varying levels of sophistication and certainty. While much work has been done on comparing the performance of various data-driven models within hybrid models, results are inconclusive with no candidate demonstrating consistently superior performance. Therefore, recent work in identifying an ‘optimal’ data-driven model for PSE will be presented, considering data-driven models typical in hybrid modeling as well as modern ML models that have thus far received little attention in hybrid modeling literature. We present the potential advantages of incorporating various data-driven models to represent the black-box component of the hybrid model.


  1. Thompson, M.L. and M.A. Kramer, Modeling chemical processes using prior knowledge and neural networks. AIChE Journal, 1994. 40(8): p. 1328-1340.
  2. Psichogios, D.C. and L.H. Ungar, A hybrid neural network-first principles approach to process modeling. AIChE Journal, 1992. 38(10): p. 1499-1511.
  3. von Stosch, M., et al., Hybrid semi-parametric modeling in process systems engineering: Past, present and future. Computers & Chemical Engineering, 2014. 60: p. 86-101.