No label required: six routine core-logging measurements already carry the geologist's lithology name, and a thousand-fold label weight only drops reconstruction from 0.999 to 0.836

Shipboard lithology labels turn out to be redundant for this archive: an embedding trained on six routine core measurements alone separates lithologies as well as a variant trained with those labels included. The default practice treats the geologist's lithology, assigned by eye from colour, grain size, texture and induration, as the ground truth a model should be taught to reproduce, so downstream analysis starts from reproducing the label rather than from the measurements. Across the same 30 borehole-level folds the label-free model reaches an adjusted Rand index of 0.237 [0.204, 0.283] against 0.247 [0.214, 0.290] for the semi-supervised variant, a difference a fold-by-fold signed-rank test cannot resolve; raising the label weight from zero to 1000 drops reconstruction monotonically from 0.999 to 0.836 while classification stays between 0.961 and 0.968.

The model learns the constraints between the six readings rather than the rock names, compressing each 20 cm interval into a ten-dimensional vector. It is a variational autoencoder, a pair of networks in which an encoder compresses a sample into a vector and a decoder rebuilds the measurements from it; 10% of inputs are hidden at random during training and filled with learnable parameters instead of zeros, which forces the model to infer a missing reading from the others, and the decoder learns a separate noise scale per variable so that magnetic susceptibility, spanning thousands of instrument units, cannot dominate the latent space. Training uses the part of the LILY database where all six properties fall in the same 20 cm bin, 296 boreholes and 236,963 observations carrying 139 lithology classes; clustering the ten-dimensional embedding gives an adjusted Rand index of 0.237, above principal components on the same six inputs (0.203) and the two self-supervised tabular baselines TabNet (0.183) and SCARF (0.167).

The finding rests on soft sediment: only 236,963 intervals in the LILY database of 34 million measurements carry all six properties, a small fraction, and the cores are dominated by intact sediment recovered by advanced piston coring, with fractured low-recovery igneous and basement rock largely excluded. Evaluation splits whole boreholes, 30 folds each holding out 10% of holes and retraining from scratch, and reports medians with interquartile ranges across folds; the intended reader works with sparse, multi-instrument borehole tables that are strongly autocorrelated along depth. Transferable properties are those with a documented physical coupling: on boreholes the model never saw, the embedding reaches 0.53 for porosity and 0.52 for whole-core magnetization, while vane shear strength, which none of the six inputs measures directly, falls from 0.93 to 0.01. The next step the authors propose is coverage rather than a bigger model, since these measurements are cheap and non-destructive: archived cores in the IODP repositories could be rescanned and the rest of the IODP holdings ingested alongside LILY.

Holding out whole boreholes could itself depress the scores, and the authors test this with lithology overlap and an energy distance between folds: overlap has a median of 0.432 and a range of 0.345 to 0.518, so extrapolation to a held-out set drawn like the training set lies outside the data, which rules out a large pessimistic bias but not a small one. In the label-weight sweep a single weight cannot be resolved by the test, and a trend across 13 weights carries the result. Whether a representation trained on marine sediment carries to continental or exploration cores is untested; the authors expect the couplings to carry but the range not to, because cementation and lithification change how porosity, velocity and conductivity relate.