Same Data, Different Representation

This report examines a narrow modeling question: whether rate-of-penetration predictions change when the model is trained on raw ROP versus log1p(ROP), with transformed predictions converted back to ft/hr.

ROP is nonnegative and strongly right-skewed. A logarithmic target compresses large values and changes the numerical geometry of the learning problem, but it does not add observations, predictors, or physical information. The study is designed to isolate that representational choice rather than search for the best model or feature set.

The comparison uses three Utah FORGE wells and five model families. Raw and transformed treatments share the same rows, predictors, preprocessing, model settings, seeds, training blocks, and test blocks. Performance is evaluated on unseen depth intervals and reported in the original unit of ft/hr.

log1p training lowered fold-balanced macro MAE in 14 of 15 well-model comparisons and macro RMSE in 11 of 15. All reported fold-balanced macro R² values were negative, so the result is a relative improvement between two fixed alternatives, not evidence of successful generalization.