Welcome to the Machine Learning for Chemical Engineering (ML4CE) course repository. This repository contains the material for present and past coursework coding assignments.
ML4CE is a module offered at Imperial College London for 4th Year MEng and MSc students in the Chemical Engineering Department. The course introduces students to core machine learning concepts and their applications in solving chemical engineering problems. Assessment is based on mini-projects that apply machine learning and optimisation techniques to chemical engineering case studies.
NOTE: Please refer to the respective coursework's README.md for guidance on setting up your environment for the project.
In this project, you will design a data-driven optimisation algorithm to tune a PID controller for a Continuous Stirred-Tank Reactor (CSTR). The optimization involves a 32-dimensional space (PID controller gains) with the objective to minimize control error.
Key features:
- Freedom to implement any DDO algorithm: direct, model-based, evolutionary search, etc.
- Limited by fixed evaluation budget (100 function evaluations)
- Challenge: balance overall tracking error, control action magnitude, and action smoothness
📖 See README here
📖 Main benchmarking notebook
📄 PDF handout
This project focuses on constrained optimisation of the classic Williams-Otto (WO) chemical process benchmark. The objective is to maximize process profit by adjusting two decision variables while respecting process constraints.
Key features:
- Optimize reactor temperature (TR) and reactant B flow rate (Fb)
- Work within constraints on reactant A usage and waste generation
- Limited evaluation budget makes efficient simulation use critical
📖 See README here
📖 Benchmarking notebook
📄 PDF handout
Design a data-driven model predictive controller to control a multistage extraction column. This coursework emphasizes the challenges of limited data and computational resources in real chemical engineering applications.
Key features:
- Create custom data-gathering routine, data-driven model, and control horizon optimization
- Limited to just 5 system evaluations for exploration
- Time-limited budgets: 2 minutes for data-gathering/model training, 5 minutes for controller simulation
- Submission requires three functions: data-gathering, model training, and controller
📖 See README here
📖 PDF Handout
📖 Quick Start Notebook
This coursework focuses on optimizing inventory management across a three-echelon supply chain using Reinforcement Learning (RL). The challenge is to develop efficient order-placement policies for retailers and distribution centers facing uncertain demand.
Key features:
- Complete freedom to implement any RL algorithm within the provided template
- Supply chain includes supplier, distribution center, and retailers
- Goal: Maximize profit while balancing inventory costs, backlog penalties, and delivery fees
- Benchmark against REINFORCE with baseline and simulated annealing algorithms
📖 See README here
📖 Main notebook
📖 Environment specification
This coursework focuses on optimizing a simulated bioprocess that uses CHO cells to produce proteins. The challenge is to develop a batch Bayesian Optimization algorithm to find optimal bioprocess parameters that maximize titre concentration.
Key features:
- Design and implement a batch Bayesian Optimization algorithm
- Optimize multiple bioprocess parameters simultaneously (temperature, pH, feed amount, cell type, etc.)
- Deal with resource-intensive experimental simulations
- Balance exploration and exploitation in the search for optimal parameters
📖 See README here
📖 Main notebook
📖 PDF Handout