Data-driven Optimization for Enterprise-Wide Optimization

Traditionally, optimization of chemical engineering systems relies on gradients of model expressions. Depending on the effort required to develop accurate models, it might be more efficient to jump the model-building step. Instead, we can use derivative-free optimization (DFO) - also known as black-box or 'data-driven' optimization - for optimal controller tuning, reactor design, experiment design, or flowsheet design based purely on input-output evaluations. This would for example involve the use of Bayesian optimization or trust region methods to iteratively construct and optimize surrogates of the system's input-output evaluations. Rather than use black-box optimization to optimize physical systems or simulations, we can solve complex coordination problems where the black-box evaluations involve optimization problems themselves. The holy grail of process systems operations would be to integrate all decision-making units that arise in the interaction between supply chain players and their hierarchical levels of decision-making. To this end, we can use DFO to determine the coordinating 'complicating' variables in value chain and multi-level hierarchical planning-scheduling-control problems, otherwise determined by heuristics, decomposition techniques, or distributed optimization.