Chun-Yi Lee (李俊毅) · portfolio
Pick the method the data structure calls for
Three public manufacturing datasets, three different data structures, three deliberately different strategies — not one technique applied three times.
▶ 10-minute deck
12 slides. Arrow keys to navigate,
P to print.📄 Full report
All five write-ups, figures and tables in one page.
🏭 產業技術報告
中文技術版：給製程／良率工程師與製造資料科學團隊，含部署考量與限制。
📖 白話技術文件
中文入門版：用生活比喻解釋整個專案，不需要任何機器學習背景。
⌘ Source on GitHub
Every number here is reproducible from these scripts.
Three layers
| Layer | Dataset | Structure | Headline result |
|---|---|---|---|
| L1 | SECOM | one wafer = a static snapshot | Drift dominates: fail rate falls 23% → 1–5% over three months; the retraining trigger fires in every test segment |
| L2 | PHM 2016 CMP | one wafer = an in-process trajectory | MSE 6.24 / MAE 1.82 on the official competition test set; hand-crafted features beat learned ones 2.3× |
| L3 | Bosch | one product = a path across 52 stations | Station L3_S32 fails at 7.8× the overall rate; the graph helps attribution, not prediction |
What came out of it: drift is not automatic. It is what happens to
the part of the process you did not measure. Two of the three projects fail
out-of-time; the third does not, because its degradation axis (consumable
wear) is an input feature the model already learned.
Five times the obvious reading was wrong
Each sub-project has a place where the numbers said one thing and a single extra diagnostic said another. Those diagnostics are the actual work.
| What looked true | The check | What was true |
|---|---|---|
| Graph features lift Bosch MCC 43% | MCC at the oracle threshold | Plain tabular ranks best; the gain was threshold luck. A GNN, added independently, agrees |
| Network-hub feature selection beats mutual information | Random subsets of the same size | Neither beats random at any k; nothing beats keeping all 444 features |
| CMP virtual metrology reaches R² 0.993 | Fit within the main MRR cluster | R² 0.897 — the bimodal label distribution was inflating it |
| Model error grows as CMP tooling wears | Error stratified by wear quartile | It does not; the most-worn quartile is the most accurate |
| The 1D-CNN is the best CMP model (best validation MSE of all) | Held-out test + seed spread | Worst test MSE of any non-baseline model; 49% swing across seeds |
