A Neural Network Powered Solution Approach for Computationally Expensive Mixed Integer Programs for Bio Jet-fuel Supply Chain Network Design

Bio jet fuels derived from feedstock offer a sustainable alternative to meeting energy needs. Modeling supply chains that produce said fuels can lead to computationally prohibitive mixed integer linear programs (MILP) that consider optimal facility location and materials routing. The present work proposes an iterative machine learning-based hybrid solution procedure that transfers some of the responsibility of facility location to the learner. First, given a random selection of facility locations, a collection of model solutions is generated. Next, a neural network is fit to the collection of solutions, with facility locations being the input and total supply chain (SC) costs as the output. Then, the next set of locations is selected to optimize the predicted output of the neural network. Finally, the MILP optimization model is called to test the selected locations, and the results are fed back into the neural network and the process is repeated. Numerical experimentation demonstrates the proposed solution procedures yield near-optimal solutions with 0.10-0.18% increase in objective function value alongside a 40-65% reduction in computational time.