Our group seeks to employ interpretable ML to answer some of the most pressing questions in sustainability. To achieve this, we decompose the sustainable supply chain by scale (Fig.1).

Process Scale: Hybrid Real-Time Optimization

Given a partial understanding of the mathematical models that govern new sustainable chemical processes, can their underlying phenomena be learned from data to determine optimal operating policies? We take a real-time approach to solve this problem, where plant inputs are optimized using learned models that adapt as the plant evolves dynamically. Related works:

• Deep-learning-aided modifier adaptation: synergies with process intensification
• Meta-learning for sample-efficient Bayesian optimisation of fed-batch processes

System Scale: End-to-End Demand Response Scheduling

What non-market factors can be used to predict the markets involved in sustainable chemical and energy systems? Can these predictions be efficiently embedded into system models to achieve optimal performance while satisfying consumer demand? We use neural networks to learn market distributions and embed these into optimal system scheduling problems. Related works:

• Integrated Design and Scheduling of Hydrogen Processes under Uncertainty: A Quantile Neural Network Approach
• Risk-aware stochastic scheduling of multi-market energy storage systems

Nation scale: Learned Heuristics for Investment Planning

What are the key decisions that influence the uptake of new sustainable chemical projects over time? Can we take a data-driven approach to isolate these factors? We aim to learn heuristic rules to produce interpretable policies that inform future technology investment.