π-features for ML

Physics is scale-free. A model built on dimensionless features generalises across scales it never saw in training; a model on raw dimensional features has to relearn the scaling and fails to extrapolate. pifeatures turns dimensional analysis into a one-line, scikit-learn-compatible preprocessing step. It is a standalone, deliberately public library (standard dimensional analysis) that installs on its own.

The value: cross-scale extrapolation

Same model, same split — the raw-feature model never saw those magnitudes and can't rescale; the dimensionless model doesn't need to.

Quickstart

pip install pifeatures

from pifeatures import PiRegressor, U
from sklearn.ensemble import RandomForestRegressor

feature_units = [U(L=1), U(L=1, T=-2), U()]   # length, gravity, dimensionless amplitude
target_units  = U(T=1)                          # period (time)

model = PiRegressor(RandomForestRegressor(), feature_units, target_units).fit(X, y)
model.predict(X_new_scales)                     # extrapolates across scales

Honest scope

• Inputs must be strictly positive (power-law groups use fractional exponents).
• The π reduction is exact and depends only on the units, not the data — it never overfits and adds no train-time cost. It captures scale structure; it does not replace learning the residual form.
• The same machinery is a data-quality check: mislabelling a column's units breaks the dimensionless groups, which surfaces the error.