Week 01 · Page 3 of 4 · Practical Work (Travaux Pratiques)
No code this week. The practical skill being trained is harder than syntax: correctly diagnosing what kind of problem you actually have.
Writing scikit-learn code before you can correctly identify a feature, a label, and the supervised/unsupervised distinction produces students who can run cells without understanding what the cells compute. Coding begins in Week 2, once the vocabulary from the Lesson page is solid. This week's practical is a structured written task, completed individually, and used directly in the Week 1 in-class debate.
Select one materials physics problem from your own background or current research interest — Heusler alloys, MAX phases, transition-metal chalcogenides, or any compound family you know well. The problem should be specific enough to name an actual property (formation energy, band gap, magnetic moment, Curie temperature, elastic modulus — not "properties" in general).
Answer each line below for your chosen problem. Write actual sentences, not single words — the debate format from Week 1's seminar (see the course presentation page) depends on you being able to defend each answer out loud.
| Prompt | What to write |
|---|---|
| The property I want to predict | Name it precisely (e.g. "zero-temperature magnetic moment per formula unit") |
| Do I currently have labels for this property, for a reasonable number of compounds? | State how many, roughly, and where they would come from (your own DFT runs, a public database, literature values) |
| Is this supervised or unsupervised, given your honest answer above? | State which, and why, using this week's definitions |
| What would the features be? | List 3–5 candidate features you could compute or measure for each compound |
| One reason ML could plausibly help here | A concrete, specific reason — not "ML is good at finding patterns" |
| One reason ML might fail or struggle here | A concrete, specific limitation — e.g. too few labeled examples, property not well-defined numerically, dominant physics not captured by your candidate features |
Before the seminar, try to argue against your own "reason ML could help" line as if you were a skeptical colleague. If you cannot find a single weakness in your own argument, you have probably stated it too vaguely — go back and sharpen it.
Bring your completed template (handwritten or typed, one page) to the Week 1 seminar. You will be asked to state your two reasons (for and against) in under two minutes as part of the in-class debate exercise. No submission is collected separately this week; the deliverable is your preparedness to speak.
What this practical is actually training
The skill being practiced here — correctly classifying a real, messy research problem as supervised or unsupervised, and identifying plausible features before any data has been collected — is a skill you will use in every subsequent week of this course, and arguably more often in real research than any specific algorithm. Algorithms can be looked up. Correctly diagnosing the problem in front of you cannot.
My name is Dr Abderrahmane Reggad. I'm a 54 year old and I am a researcher in materials' sciences.
0 Comments