This project develops algorithmic protocols and computational tools for solving mixed-integer nonlinear optimization problems; we focus on building effective solution strategies for engineering applications.

We’re hybridizing mechanistic, model-based optimization with data-driven optimization by developing mathematical programming formulations of machine learning surrogate models.

We’re developing data-efficient strategies to reduce the number of experiments needed for chemical and pharmaceutical development and manufacturing.

Safety-critical settings sometimes require optimization with respect to the worst-case realization of the problem setting. The adversarial, robust optimization setting is especially useful when errors are very costly.

Optimization solvers are often unexplainable black boxes where the reasoning for a solution is inaccessible to users. We use argumentation to empower the interaction between optimization solvers and users, supported by tractable explanations which certify or refute solutions.

In logistics applications such as our Royal Mail collaboration, a planned schedule is made in advance. But suppose something disruptive happens and we need to adapt quickly. How can we make rescheduling efficient?

Heat exchanger networks reuse excess industrial process heat onsite and thereby improve energy recovery. We consider the resulting mixed-integer nonlinear optimization problem for designing heat recovery networks.

This project considers robust superstructure design and operation of a bioreactor that produces red blood cells. Although we develop an optimization model for a specific bioreactor, the intellectual framework may generally apply to bioprocess optimization.

This project optimizes fuel blending for energy systems on a feed-forward network of inputs, intermediate storage, and outputs; we investigate blending petrochemical feeds in a way that maximizes profit subject to environmental standards.

Our Dimension Scheduling Algorithm reduces the computational burden of Bayesian optimization for many experiments. The key is optimizing the fitness function only along a small set of dimensions at each iteration.