Lab notebook · 12 / 27Regression

Superconductor critical-temperature regression

Can composition-derived features predict the critical temperature of known superconductors?

Self-directed study

Scope of this result

Test formulas are separate from training formulas. The model predicts temperature for known superconductors, not whether any material superconducts.

The evidence

Fig. 01

Histogram of critical temperature for 21,263 superconductors, concentrated at lower temperatures with a long tail beyond 120 kelvin. — The target distribution across 21,263 known superconductors in the verified UCI archive.

01 · Method and interpretation

A random forest predicts critical temperature for materials already known to be superconductors from 81 composition-derived features in the UCI archive. Exact chemical formulas are kept on only one side of the train/test boundary, correlation pruning is learned from training data, and hyperparameters are selected with grouped cross-validation. On one untouched formula-disjoint fold the model reaches R² 0.933, RMSE 9.08 K, and MAE 5.23 K; it does not determine whether an arbitrary material superconducts.

02 · Supporting figures

Selected outputs from the complete notebook.

Fig. 02

Scatter plot of observed versus predicted critical temperature on the formula-disjoint test fold, with a dashed perfect-prediction diagonal. — Random-forest predictions on a test fold containing no chemical formula seen during training.