A dynamic nonlinear optimization framework for learning data-driven reduced-order microkinetic models

The use of kinetic models is key for analyzing, designing, controlling, and optimizing the manufacturing of high value chemicals. In particular, microkinetic modeling relies on large-scale networks of elementary reaction steps that map the conversion of feed materials to final products by considering a large number of pathways and intermediate species for a given reactor configuration. Intuitively, several of the (often automatically generated) reaction steps might be redundant or insignificant, and thus determining the true governing reaction network is critical to understanding and modeling the underlying chemistry. This work introduces a nonlinear dynamic optimization framework for discovering governing reaction networks from data, whereby both the model structure (the elementary reaction steps) and the model parameters (reaction rate constants, pre -exponential factors, and activation energies) are simultaneously learned from composition time series data. The proposed framework can also achieve dimensionality reduction to produce accurate reduced-order microkinetic models that are computationally parsimonious and thus better suited for applications involving simulation and optimization. Two numerical examples of different dimensions are presented to illustrate the key properties of our approach.

Automated summary

The use of kinetic models is crucial for the manufacturing of high value chemicals. This research introduces a nonlinear dynamic optimization framework for discovering governing reaction networks and learning model parameters from data. The proposed framework can produce accurate reduced-order microkinetic models that are computationally efficient. Numerical examples are provided to illustrate the key properties of the approach.