Multiscale optimization of formic acid dehydrogenation process via linear model decision tree surrogates

Abstract

Multiscale optimization problems require the interconnection of several models of distinct phenomena which occur at different scales in length or time. However, the best model for any particular phenomenon may not be amenable to rigorous optimization techniques. For instance, molecular interactions are often modeled by computational chemistry software packages that cannot be easily converted into optimization constraints. Data-driven surrogate models can overcome this problem. By choosing surrogates with functional forms that are convertible to a mixed-integer linear model, one can connect and optimize these surrogates instead of the underlying models. We demonstrate the interconnection of linear model decision trees to optimize across three scales of a formic acid dehydrogenation process. We show that optimizing across all three scales simultaneously leads to a 40% cost savings compared to optimizing each model independently. Furthermore, the surrogates retain some relevant physical behaviors and provide insights into the optimal design of this process.