Expert lithological labels do not improve learning when used to predict marine sediment properties

• The measurement-only embedding distinguishes rock types about as well as geologists’ labels do.The model also estimates properties that were not measured on most of the core.Source: Plain Language Summary · p. 2
• Adding lithology labels does not improve classification and makes the model worse at describing the measurements.Source: Plain Language Summary · p. 2
• Geologists assign rock names using characteristics such as color and density that are already present in the measurements.Source: Plain Language Summary · p. 2

1 Introduction

• Drilling archives are incomplete because instruments sample at different intervals, core recovery is partial, and destructive analyses prevent later measurement of the same material.Source: §1 · p. 3
• Lithology is commonly treated as ground truth for model training, although geologists assign it from observed rock characteristics.Source: §1 · p. 3
• The study compares a self-supervised model trained on six geologist-used properties with a semi-supervised variant that also uses lithology labels.The inputs are density, magnetic susceptibility, natural gamma radiation, and red, green, and blue reflectance.Source: §1 · p. 4

2.1 The LILY database

• LILY pairs ocean-drilling measurements with lithology descriptions and aggregates records from more than 34 million measurements across 89 km of core and 42 expeditions.The database standardizes lithology descriptions and includes corrected bulk-density and porosity estimates.Source: §2.1 · p. 4

2.2 Building the training set

• After binning readings into 20 cm intervals and retaining bins with all six properties, the modeled set contains 236,963 observations from 296 boreholes and 25 expeditions.Rows with implausible values were removed; the resulting subset is small because the six properties rarely co-occur, and it is dominated by marine sediment.Source: §2.2 · pp. 4–5

2.3 Held-out variables

• Held-out variables include alternate-instrument measurements and independent properties such as porosity, thermal conductivity, P-wave velocity, and whole-core magnetization.Procedural variables and near-linear restatements of other variables were excluded.Source: §2.3 · p. 6

3.1 Representation learning and variational autoencoders

• The variational autoencoder compresses samples into embeddings intended to retain shared measurement structure, which are tested on tasks and boreholes unseen during training.A semi-supervised variant adds lithology labels to test whether they change the embedding.Source: §3.1 · pp. 6–8

3.2 Architecture

• Masking hides input features and requires reconstruction from the remaining measurements, while per-variable decoder noise scales prevent noisier properties from dominating the latent space.Learnable mask values distinguish hidden features from standardized zero values.Source: §3.2 · p. 8

3.3 Objective

• The self-supervised objective combines heteroscedastic reconstruction likelihood with a KL penalty, while the semi-supervised variant adds a weighted lithology-classification loss.Source: §3.3 · pp. 8–9

3.4 Evaluation protocol

• Evaluation holds out whole boreholes because measurements are autocorrelated within cores, and reports medians across 30 borehole-level folds.Each fold holds out 10% of boreholes and retrains the model; transfer prediction trains and scores on separate boreholes.Source: §3.4 · pp. 9–10

4.1 Baselines and ablations

• The embedding exceeded PCA by ΔARI = 0.030 and TabNet and SCARF by 0.065 and 0.059, respectively, with each difference statistically resolved.The paper reports these models on the same 30 borehole-level folds.Source: Table 1 · p. 11, §4.1 · p. 11
• Removing the heteroscedastic decoder reduced reconstruction from 0.999 to 0.976 and classification from 0.958 to 0.939; no-masking left clustering indistinguishable.The authors suggest a single noise scale is set by whichever input has the largest residuals.Source: §4.1 · p. 11
• Every reconstruction-capable model achieved R2 ≥ 0.999, while PCA reached exactly 1.000, so reconstruction accuracy was not informative for this task.PCA's exact reconstruction follows from using six components on a six-feature input.Source: §4.1 · p. 12

4.2 Supervision by lithology label adds nothing

• Self-supervised and label-trained models were indistinguishable; the six inputs already expose much of the information geologists use to assign lithology.The label-trained variant reached ARI 0.247 versus 0.237 for the self-supervised model, with no significant differences in clustering, classification, or reconstruction.Source: §4.2 · pp. 12–13
• As α rose from zero to 1000, reconstruction fell from 0.999 to 0.836 while classification stayed between 0.961 and 0.968 and clustering showed no trend.Across the thirteen weights, reconstruction correlated with log α at ρ = −0.98.Source: §4.2 · p. 13

4.3 Zero-shot prediction

• On unseen boreholes, embeddings predicted porosity (R2 = 0.53), whole-core magnetization (0.52), discrete P-wave velocity (0.39), and thermal conductivity (0.30).Magnetic susceptibility and reflectance are cross-sensor inputs, not independent targets.Source: §4.3 · pp. 14–15
• Vane shear strength fell from 0.93 in-sample to 0.01 held-out, and penetration resistance from 1.00 to −1.23, failing to transfer across boreholes.The authors attribute the near-perfect in-sample scores to evaluating the regressor on the same core used for fitting.Source: §4.3 · p. 16

4.4 Depth does not transfer between boreholes

• Depth alone reached median R2 −0.017 across nine independent targets versus 0.201 for embeddings; embeddings did better on eight, while neither predicted discrete magnetization.Depth also carried some signal for porosity and discrete P-wave velocity, but the embedding did better on both.Source: §4.4 · p. 16

5.1 Expert labels and self-supervision

• Lithology is usually recoverable from measurements, except microscope-dependent fossil distinctions; categorical labels also discard gradational process detail that measurement-trained models retain.The discussion describes lithology as a lossy compression of continuous measurements, not as evidence that geologists are wrong.Source: §5.1 · p. 17

5.2 Mixed-determined problems in Earth science

• Earth science measurements are heterogeneous and uneven, while target quantities are often unmeasured; learned representations over recorded data are one proposed route into these mixed-determined problems.The discussion notes that such problems can be overdetermined in densely sampled locations and underdetermined elsewhere.Source: §5.2 · p. 17

5.3 Depth and compaction

• Depth-dependent models can score highly on rows from familiar holes yet fail on a new hole, motivating evaluation that holds out boreholes.The discussion reports residualized held-out scores of 0.253 for thermal conductivity, 0.432 for whole-core magnetization, and 0.169 for discrete P-wave velocity; these scores are not on the same scale as scores before depth removal.Source: §5.3 · p. 18

5.4 Beyond marine sediment

• Transfer beyond marine sediment remains untested, and the soft-sediment range does not cover crystalline basement, where cementation and lithification change property relations.The discussion identifies testing where the model begins to fail as more informative than simply extending it.Source: §5.4 · pp. 18–19

5.5 Data infrastructure is the bottleneck

• Only 236,963 of 34 million database intervals contain all six properties, and the discussion identifies joint usability of measurements—not a larger model—as the key infrastructure need.The authors note that the six properties are cheap and non-destructive to measure and that archived cores could be rescanned to fill gaps.Source: §5.5 · p. 19

6 Conclusions

• Across 30 borehole folds, labels add nothing to the measurements: label-weight increases degrade reconstruction while leaving classification unchanged.The conclusions state that clustering, classification, and reconstruction are indistinguishable between variants across folds, with no individual lithology benefiting.Source: §6 · p. 19
• On unseen boreholes, the representation predicts untrained porosity at R2 = 0.53, whole-core magnetization at 0.52, discrete P-wave velocity at 0.39, and thermal conductivity at 0.30.The conclusions say each success rests on a coupling described before the model existed, while properties with no such coupling fail completely.Source: §6 · pp. 19–20
• On unseen boreholes, depth alone reaches a predictive median R2 of −0.02 across independent targets, against 0.20 for the embedding.The conclusions emphasize holding out the unit of geological independence and checking that predicted variables are not restatements of inputs.Source: §6 · p. 20

• Energy distance compares training and held-out rows in the six standardized inputs, using subsamples of 2,000 rows per side and a shift-based calibration.Its median value of 0.066 corresponds to a shift of about 0.45 standard deviations.Source: Appendix A · p. 20
• Lithology overlap had a median of 0.432 across folds, while energy distance had a median of 0.066.Source: Appendix A · pp. 20–21
• Transfer R² was uncorrelated with lithology overlap (Spearman ρ = −0.05, p = 0.78) and energy distance (ρ = +0.02, p = 0.90).Source: Appendix A · p. 21
• Extrapolating linear fits to held-out sets indistinguishable from training gives R² of 0.07–0.11, against an observed median of 0.10.Source: Appendix A · p. 21
• Random row splits scored porosity at R² = 0.740 versus 0.531 for borehole splits, reflecting leakage from near-duplicate core rows.Thermal conductivity scored 0.474 versus 0.291; the comparison used 30 folds with the encoder held fixed, and the porosity leak was close to 0.2 R².Source: Appendix A · p. 21
• Lithology overlap ranged from 0.345 to 0.518, so complete overlap is outside the observed range and a small pessimistic bias cannot be ruled out.Source: Appendix A · p. 21

• The LILY database is publicly available on Zenodo.Source: Open Research · p. 23
• The code used for preprocessing, modeling, evaluation, and data curation is available on GitHub.Source: Open Research · p. 23
• The code is also archived on Zenodo.Source: Open Research · p. 23