2024 AIChE Annual Meeting
(373u) Multiscale Optimization Via Linear Model Decision Tree Surrogates
Authors
Kitchin, J., Carnegie Mellon University
Laird, C., NA
Akhade, S., Lawrence Livermore National Laboratory
Large-scale sustainability problems require interconnected models of many distinct phenomena which may occur at different length or time scales. Such models may be built using different techniques or software packages which inhibit their rigorous optimization. Surrogate models can solve this problem. A good surrogate model will maintain good fidelity to the underlying model but have an entirely different functional form. By building a linear model decision tree surrogate of each model, they may all be optimized simultaneously as a mixed-integer linear problem due to their piecewise linearity. We demonstrate this technique by simultaneously optimizing an atomic-scale catalyst model, a reactor model, and an economic model of a formic acid dehydrogenation process. These models can be imported directly into a Pyomo optimization using the Optimization and Machine Learning Toolkit (OMLT). We show that co-optimization results in a 40% reduction in the levelized cost of hydrogen compared to the independent optimization of these three models.