Programme
AI for Nuclear Thermal Hydraulics Summer School
13–17 July 2026 | STFC Daresbury Laboratory
Weekly Timetable
| Day | Morning Block | Afternoon Block |
|---|---|---|
| Day 0 | Self-learning technical onboarding pack with virtual organiser support. Support: Dr Wei Wang, Dr Yu Duan Purpose: remove setup friction and align terminology, tools, and data structure. |
Complete before Day 1. Outputs: run starter notebook, produce one simple data plot, and confirm tools and data work correctly. |
| Day 1 | NTH problem framing, supervised ML, and workflow Lecture 1.1: Programme overview and NTH problem framing Lecturer: Prof. Michael Bluck, Imperial College London Lecture 1.2: Overview of ML methods Lecturer: Prof. Michael Bluck / Dr Yu Duan Lecture 1.3: ML workflow for TH and failure gallery Lecturer: Dr Yu Duan |
Lab 1: Baselines, preprocessing pipelines, and first ML baseline model Focus: supervised tabular ML, preprocessing, evaluation, and Linear / Ridge regression baseline. Lead: Dr Yu Duan; Demonstrators: TAs |
| Day 2 | Tree ensembles for engineering tabular data Lead lecturer: Dr Junwen Yin, STFC Lecture 2.1: Random Forest — decision trees, RF intuition, engineering tabular data, key hyperparameters. Lecture 2.2: Gradient Boosting — relationship to RF, progressive boosting logic, key tuning parameters. Also includes cross-validation, regime-wise diagnostics, and cautious interpretability. |
Lab 2: Tree ensembles with regime-wise diagnostics Focus: use the CHF case study with Random Forest and Gradient Boosting; compare outputs, perform error slicing, and produce an engineering recommendation. Lead: Dr Junwen Yin; Demonstrators: TAs |
| Day 3 | Surrogate modelling and Gaussian Processes Lecture 3.1: Surrogate modelling — what is a surrogate, major surrogate families, how to build one properly, and what can go wrong. Lecturer: Prof. Michael Bluck Lecture 3.2: Gaussian Process regression — GP concept, kernel functions, and sources of GP uncertainty. Lecturer: Dr Tim Rogers |
Lab 3: GP surrogate and extrapolation stress test Focus: train and examine a GP on a controlled subset of the CHF dataset; compare interpolation and extrapolation behaviour; contrast GP with tree-based methods. Lead: Dr Tim Rogers; Demonstrators: TAs |
| Day 4 | Neural networks for tabular engineering data and physics-guided guardrails Lecturer: Dr Alex Skillen, University of Manchester Lecture 4.1: Neural networks — what an NN is, relation to other ML methods, basic training loop, and overfitting. Lecture 4.2: Physics-guided ML — what it is and why physics guidance matters in NTH ML. |
Lab 4: Feed-forward neural network, guardrails, and edge-case testing Focus: train and examine a simple feed-forward NN on a controlled CHF dataset; apply constraints / guardrails and inspect edge cases. Lead: Dr Alex Skillen; Demonstrators: TAs |
| Day 5 | Practical VVUQ and safe optimisation for decision support Lecture 5.1: Introduction to VVUQ — why UQ matters in NTH decisions; aleatoric vs epistemic uncertainty. Lecturer: Dr Saleh Rezaeiravesh, TBC Lecture 5.2: UQ for decisions and safety-critical ML practices — uncertainty in ML predictions, calibration, reliability, and validation for safety-conscious use. Lecturer: Dr Saleh Rezaeiravesh, TBC Lecture 5.3: Optimisation — turning ML models into decisions. Lecturer: TBC |
Guest talks, synthesis, and closing discussion Short invited talks on advanced or emerging topics, followed by synthesis of key lessons and closing remarks. Guest speakers TBC |
Programme Outcomes
By the end of the week, participants should be able to:
- frame an engineering ML problem appropriately
- build and evaluate baseline supervised-learning models
- compare tree-based models for structured engineering data
- explain surrogate modelling and Gaussian Process regression
- train a small neural network for tabular data
- assess uncertainty and domain validity
- communicate model limitations and safe-use conditions