Computational Engineering / module 18
AI and machine learning for structural engineers
What a trained model is, what it cannot know, and where it belongs beside a deterministic method rather than instead of one.
What you will understand
- Say what machine learning is doing, in terms an engineer can check
- Distinguish supervised, unsupervised and reinforcement learning by what each needs
- Examine how class imbalance can conceal poor minority-class recall behind a high overall accuracy.
- Show that class imbalance destroys recall while accuracy stays respectable
- Explain why a model gives no signal that it is extrapolating
- State where AI belongs in structural work, and where it does not
- Apply the rule that a deterministic check, where one exists, must have been run
Lessons
What it is actually doing
Fitting a function to samples — and what that implies about every claim made for it.
Overfitting and generalisation
Why training error falls forever, why test error turns round, and why the gap is invisible from the training data alone.
Bias, and what the model never saw
Class imbalance, distribution shift, and the two numbers that behave completely differently.
Surrogates, and who is accountable
A fast approximation to a slow analysis — what it buys, what it costs off the edge of its data, and the rule that governs its use.
Module summary
- Machine learning fits a function to samples; everything else follows from that one sentence
- Training error falls forever; test error turns round, and the gap is the model learning noise
- Accuracy hides what per-class recall shows — starve a class and it collapses while accuracy stays high
- Nothing in the output distinguishes an interpolation from an extrapolation
- Correlation is what these methods find; causation is what engineers need
- Where a deterministic method exists, use it — a surrogate explores, it does not decide
- The qualified engineer remains accountable, whatever tool informed the decision
Teaching basis and review status
These are educational models and exercises. Check their assumptions before using the results. The full technical content and every diagram have not yet been independently verified.
Concrete and steel calculations retain their stated first-generation Eurocode teaching basis. Use the standards and amendments specified for your project; do not mix generations without an appropriate design basis.
Read the lesson register and outstanding review notesCheck your understanding
Check yourself
A model is proposed to predict bending moments in continuous beams, trained on 50 000 analysis results. What is the strongest objection?
Practice
A balanced classifier achieves 95.0 % recall on a class. After the training set for that class is cut from 80 examples to 3, its recall falls to 26.7 %. By how many percentage points has it fallen?
Practice
A study of 200 combinations at 20 minutes each is replaced by 40 training runs, surrogate evaluation, and 10 confirming runs. What percentage of the original analysis time does that use?
Practice
A classifier is tested on 800 models, of which 40 have a defect. It flags 30 correctly and misses 10. What is its accuracy if it raises no false alarms at all?
Check yourself
Which single figure should be asked for first when a machine-learning tool is offered for structural work?