🏆 The Open Challenge: find your OWN way to #1 (beat MAE 0.0114) — no recipe
The dare
We're not going to hand you the recipe.
Here's the bar, and the door's wide open: beat MAE 0.0114 on cell state-of-health, MIT–Stanford–TRI (Severson et al., Nature Energy 2019, 144 cells), 5-fold CV from a 30% observation window (~45 cycles). That's the current #1 — better than a published Attentive NeuralODE.
How you get there is up to you. Pure harmonic descriptor, a hybrid, your own twist on the fold map, a totally different model on top — find your own path. Half the fun is the route.
The rules (light, on purpose)
- Public data only — MIT–Stanford–TRI: https://data.matr.io/1/projects/5c48dd2bc625d700019f3204
- Same task — SOH, 5-fold CV, ~30% window, so numbers are comparable.
- Show your work — post your MAE (and RMSE/PCC/R² if you have them) + a snippet or repo link so others can check it.
- Any method counts — MHM, GNN, transformer, kernel trick, something nobody's tried. Surprise us.
🪜 Where to start (then leave the trail behind)
pip install "git+https://huggingface.co/williamTLmiller/batterymhm"
python demo.py
The whole harmonic method is open in this repo if you want to build on it — or ignore it entirely and bring your own. The math is in docs/METHOD.md; a full training example is in examples/predict_soh.py.
🏅 The board
Reply with your result and we'll keep a running list right here:
| Rank | Builder | MAE ↓ | Approach | Link |
|---|---|---|---|---|
| 🥇 | (the bar to beat) | 0.0114 | harmonic descriptor + tree ensemble | this repo |
| ? | you | ? | your way | ? |
Honest fine print
No trained weights are shipped — you train your own (seconds, CPU). The method is patent pending (CC-BY-NC-4.0 = copyright license only, no patent grant; commercial use may need a separate license). This is a friendly research challenge, not a paid competition.
The number is 0.0114. The road is yours. Post what you find. 🔥
— William T. L. Miller, inventor