Computationally Efficient Optimization of Bio-Jet Fuel Supply Chains Using Machine-Learning-Assisted Mixed-Integer Programming

Abstract

Abstract

Bio-jet fuels produced from biomass-derived feedstocks represent a promising pathway for reducing the carbon intensity of aviation energy systems. However, designing supply chain networks for bio-jet fuel production requires solving large-scale mixed-integer linear programming (MILP) models that integrate facility location, feedstock allocation, material flows, and routing decisions. These models can become computationally expensive, particularly when evaluating multiple network configurations or large candidate sets of production and processing facilities. This study develops a hybrid machine learning and optimization framework to improve the computational efficiency of bio-jet fuel supply chain network design while preserving high-quality decision outcomes. The proposed iterative procedure uses supervised learning to approximate the relationship between facility location decisions and total supply chain cost. First, an initial set of supply chain configurations is generated by solving the optimization model while using randomly selected facility locations. These solutions are then used to train predictive models, including ridge regression, feed-forward neural networks, and ensemble neural networks, with facility-location configurations as inputs and total supply chain cost as the output. The trained learner is subsequently used to identify promising facility-location candidates through Thompson sampling and small-scale linear programming. These candidate solutions are evaluated by the original mixed-integer model, and the resulting observations are fed back into the learning process until convergence. Numerical experiments show that the proposed hybrid approach obtains near-optimal bio-jet fuel supply chain designs while substantially reducing computational time. For the linear case, the method achieves solutions within 0.23–0.29% of the objective function value while reducing computational time by 70.95–81.95%. For nonlinear learning models, the optimality gap decreases further to 0.13–0.15%, with computational time reductions of 45.37–60.36%. For the Texas case study and the modeling assumptions evaluated, the findings demonstrate that machine-learning-assisted optimization can reduce computational effort while preserving high-quality supply chain solutions. The extent of these benefits may vary with network size, candidate-facility structure, facility-capacity assumptions, demand characteristics, and the amount of information available to train the learning models.