- One Arc: raw materials data → representations → models → uncertainty and trust.
- Units 1–3 (Data foundation): small data, measurement chains, leakage-safe validation.
- Units 4–5 (Representation): stereology to learned embeddings; unsupervised clustering.
- Units 6–9 (Learning under constraints): transfer learning, time series, inverse problems, spectral signals.
- Units 10–12 (Modern architectures & trust): transformers, physics-informed constraints, uncertainty quantification.
Every unit is recapped through four checkpoints:
- What question did it solve?
- What method & central equation answers it?
- When do you reach for it — and where does it break?
- What must you be able to do in the exam?
- Twelve real lab scenarios mapped to method and unit, e.g.:
- Small tabular data with error bars → Gaussian Process (U12).
- Streaming sensor data, decisions on forecasts → probabilistic LSTM (U7).
- Long-range image correlations, pretraining available → ViT (U10).
- Known PDE, sparse sensors → PINN (U11).
- No labels at all → clustering / autoencoders (U5).
- Two precursor questions: "Where did this data come from?" (U1–3) and "How wrong can I afford to be?" (U12).
- The 14-week / 12-unit arc as one data-to-trust pipeline.
- Each unit's question, method, equation, and failure modes.
- The decision-guide table: task → tool → unit.
- Exam scope and mini-project rubric.
Summary for ML-PC Week 14:
- Recaps Units 1–12 with a four-question framework (question, method & equation, when to use, exam competence).
- Provides a decision guide mapping twelve lab scenarios to the right method and unit.
- Anchors deployment on two questions: data provenance and acceptable error budget.
- Closes with exam scope and the mini-project rubric: ML amplified the materials scientist, it did not replace them.