(334i) Data-Driven Parameter Estimation of Hybrid Models

Data-driven modeling has proven a powerful tool for drawing useful correlations from large datasets, yet the ability to generate physical insights from purely data-driven models remains limited due to their black-box nature. Hybrid models (HMs) have the potential for increased physical interpretability by leveraging mechanistic models for known and hypothesized relationships with data-driven models for unknown relationships in a single framework. HM frameworks often assume mechanistic information is fully known, including parameter values, prior to model regression.[1] However, for many systems of interest, constructing an HM requires estimating the parameters of both data-driven and hypothetical mechanistic submodels on the same limited dataset.[2]

This submission considers novel methods for estimating parameters of dynamic models containing unknown mechanistic and data-driven parameters. Specifically considered are systems that can be formulated as a series of coupled ordinary differential equations. These methods are compared with traditional approaches to parameter estimation using forward sensitivity analysis. The comparison is further extended to the case where experimental data is low-quality and sparse. Results indicate that methods which merge data-driven models with numerical methods provide better estimates of time-evolved data and their derivatives than purely data-driven approaches. Conclusions for this presentation will identify parameter estimation approaches that accelerate the validation of interpretable models for systems when both data and mechanistic knowledge is limited.

References

1. Oliveira, R., Combining first principles modelling and artificial neural networks: a general framework. Computers & Chemical Engineering, 2004. 28(5): p. 755-766.
2. Yang, A., E. Martin, and J. Morris, Identification of semi-parametric hybrid process models. Computers & Chemical Engineering, 2011. 35(1): p. 63-70.

Research Interests

My research has focused on investigating methods for merging data-driven tools and mechanistic knowledge for modeling dynamic process data. The algorithms developed enable modelers to accelerate the systematic validation of hypothesized chemical-physical relationships without system-level information. I have further investigated the application of modern tools in automatic differentiation and numerical methods to support automated estimation of complex differential equations. My research interests lie primarily in developing novel solutions for modeling, control and optimization of manufacturing systems and supply chains.