The treasure map, animated
Before the machine learning, here's the physics in plain language — what superconductivity is, why heat is the enemy, and the three unsolved gaps where a fresh idea could actually matter.
The glue is everything
Electrons normally repel and bump into the lattice, losing energy as resistance. In a superconductor they pair into Cooper pairs that glide with zero resistance. Heat is the enemy — it jiggles the atoms and rips the pairs apart. The whole quest is: keep the pairs glued at high temperature. What supplies the glue?
Atoms vibrating in sync. Best result — but only under Earth-core pressure. Great glue, impossible plumbing.
Copper-oxide layers, works at normal pressure. Catch: 40 years on, nobody fully knows why.
An unknown mechanism at ambient pressure. This is the open door — and where an outsider hunch can matter.
A model that predicts critical temperature
We trained a gradient-boosted model on the UCI Superconductivity dataset — 21,263 real materials, each described by 81 features derived from its chemistry. Given a formula it never saw, it predicts the temperature at which superconductivity kicks in.
R² = 0.93 on held-out materials, with 5-fold cross-validation at 0.928 ± 0.002 — the trustworthy number, no fluke. Average prediction lands within ~5 K of the real value. That matches the published state-of-the-art baseline (Hamidieh 2018) on turn one.
One lever rules them all
When we ask the model which features drive its predictions, the answer is startlingly lopsided. The spread in thermal conductivity among a material's atoms carries the bulk of the signal — and thermal conductivity is a phonon story. The machine is quietly shouting the physics: the glue is in the lattice vibrations.
Two findings that actually matter
The dominant lever is a phonon lever. Of all 81 features, the range of thermal conductivity alone accounts for ~70% of predictive power. It's a direct echo of the "phonon glue" from the map — the model rediscovered the physics from raw data.
Where it fails is the interesting part. Nearly every one of the model's biggest misses is a cuprate — the copper-oxide compounds nobody has a theory for. A feature-averaging model literally cannot see their secret. That's not a bug; it's the map pointing straight at the undiscovered physics.
Where we row next
Baseline model
State-of-the-art Tc prediction trained and validated on this box.
Screen for winners
Feed the model untested formulas and let it flag high-Tc candidates — the actual treasure-hunt step.
Crack the cuprates
Build a second model just on the families it fails on, and try to beat it.
Feature injection
Invent a new descriptor — the "faking pressure with chemistry" hunch — and see if it moves the needle.
The brutal honesty
What this is — and isn't
- A prediction is a lead, not a discovery. Someone still has to synthesize and measure a candidate to claim the prize.
- The field is littered with mirages (LK-99, 2023). We stay skeptical by default.
- But screening works — the near-room-temp hydrides were predicted on a computer first and made second. That door is open to us.