🔋 What if the best battery-health model isn't a neural net? — a challenge for AI builders
The hook
Everyone reaches for a bigger transformer. BatteryMHM goes the other way.
It reads the first ~15–45 cycles of a cell, runs on a laptop CPU in seconds, ships no GPU and no giant weights — and still lands #1 on the MIT–Stanford–TRI benchmark (MAE 0.0114), beating a published Attentive NeuralODE and crushing Microsoft BatteryML's strongest baseline by 21.6×.
The trick isn't depth. It's representation: fold every measurement into a 9-class harmonic space (HIN(k) = 1 + ((k−1) mod 9)), score the interactions through a 9×9 Chi compatibility matrix, and let a light tree ensemble read a 557-feature harmonic descriptor. That's the whole method — and every line of it is open in this repo (CC-BY-NC-4.0).
Why AI creators should care
The harmonic descriptor is domain-agnostic. It turns any integer-identified sequence or structure into a fixed-length feature vector. Batteries are just the first proof. If you build with sequences, graphs, or tabular data, this is a fresh primitive to put in your toolbox.
🏗️ Build something (open challenges)
- Hybridize it. Concatenate the 557 harmonic features with your favorite GNN/transformer embedding. Does it lift accuracy? (On the materials track, the descriptor beats the classic RF baseline but trails modern GNNs — that gap is yours to close.)
- Port it. Try the fold-map + Chi-matrix descriptor on a non-battery sequence problem (audio, sensor, finance, bio). Show us a notebook.
- Beat the baseline differently. Swap ExtraTrees/XGBoost for your model on the same descriptor. Can you push SOH MAE below 0.0114?
- Explain it. Which harmonic features carry the signal, and why?
model.top_features()is one line away.
🚀 Start in 30 seconds
pip install "git+https://huggingface.co/williamTLmiller/batterymhm"
python demo.py # offline, no data, no weights — see it carry signal
Read the math in docs/METHOD.md, grab the ready-made examples/, and run make test.
Honest fine print
No trained weights are shipped — you train your own (seconds, CPU). The bundled demo is a synthetic signal check; the headline numbers reproduce on the public Severson / Matbench datasets. The method is patent pending (CC-BY-NC-4.0 is a copyright license only — no patent grant; commercial use may need a separate license).
Reply with what you build, what you break, or where it surprises you. Best experiments get featured. Let's see how far a harmonic representation can go. 🔥
— William T. L. Miller, inventor