PaperlayerView exact revision on arXiv

Source-linked AI summary

Autoregressive Drillhole Modelling Under Distribution Shift

Yihao Ding, Daniel Yitian Su, Yiran Zhang, Christopher M. Gonzalez, Wei Liu

arXiv:2610.01204v1cs.LG

TL;DR

Drillhole modelling has lacked a public evaluation of autoregressive prediction and generation that follows the one-directional drilling process. DrillBench benchmarks both tasks across spatial and geological shifts and combines sequence modelling with pretraining and spatial retrieval. Sequence models transfer more robustly than direct spatial or geochemical conditioning, while retrieval and pretraining contribute in different geological regimes.

  • Problem

    Existing drillhole modelling is dominated by interpolation and reconstruction, while well-log methods largely use masked modelling or point-wise classification, leaving autoregressive roll-forward evaluation without a public benchmark or shared protocol.

    Source: §1 · p. 2, §2 · p. 3

  • Method

    DrillBench evaluates next-layer prediction and free-running generation across graded shifts and develops a recipe combining layer-level autoregression, historical-drillhole pretraining, and retrieval of neighbouring lithology.

    Source: §1 · p. 2, §5.3 · p. 6

  • Results

    Sequence-only autoregressive models balance local accuracy and transfer better than direct spatial and geochemical conditioning, while retrieval helps most in weathered cover and pretraining more in bedrock and broader transfer.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The results support evaluating subsurface models across geological shifts and combining sequence structure with retrieval where local continuity is informative and pretraining for broader transfer.

    Source: §1 · p. 2, §7 · p. 10

    The study uses company logs without correcting depths and thicknesses for hole orientation, so reported thicknesses are apparent; lithology is harmonized from free-text descriptions.

    Source: A SCOPE AND LIMITATIONS · p. 13

Abstract

from arXiv · show

Autoregressive modelling has achieved remarkable success in language and sequence tasks by learning to predict future states from previous observation. Mineral-exploration drillholes provide a natural but largely unexplored setting for this paradigm: as drilling proceeds, lithology is revealed sequentially from shallow to deep, making prediction of deeper strata inherently autoregressive. Existing drillhole modelling, however, is dominated by spatial interpolation and reconstruction, or largely rely on masked modelling, leaving strictly autoregressive prediction largely underexplored. We introduce DrillBench, a benchmark of 49,671 Western Australian drillholes for next-layer prediction and autoregressive stratigraphic generation across a graded transfer spectrum, from local prediction through spatial shift to cross geological province transfer. Benchmarking classical, geostatistical, and neural models reveals a clear \emph{transfer boundary}: spatial and geochemical conditioning provides large local gains but deteriorates sharply under stronger shift, whereas lithology-sequence autoregressive models transfer more robustly. Guided by this finding, we develop a backbone-agnostic recipe combining large-scale pretraining on historical drillholes with spatial retrieval of neighbouring lithology. Retrieval is most effective in weathered cover, when local spatial continuity remains informative, whereas pretraining contributes more strongly in bedrock and under broader geological shift. Together, they retain strong local performance while improving generalisation under spatial and cross-province shift, most markedly on the most distant splits. The benchmark and code are available at https://github.com/yihaoding/drillbench.

1 INTRODUCTION

Drillhole sequences naturally support autoregressive prediction, but existing modelling does not evaluate rolling a partially drilled hole forward from its own predictions. DrillBench establishes that evaluation and reports stronger transfer from sequence-only models, then adds pretraining and retrieval with complementary roles across geological settings.

1 INTRODUCTION

  • Drilling reveals lithology sequentially, making prediction of the next stratum from the observed sequence a natural autoregressive task.Existing drillhole modelling emphasizes interpolation, simulation, and reconstruction; related well-log work largely uses masked representation learning or point-wise classification.Source: §1 · pp. 1–2
  • Sequence-only autoregressive models balance in-distribution accuracy and transfer better, while direct spatial and geochemical conditioning can become harmful under stronger shift.The benchmark evaluates next-layer prediction and free-running generation across progressively larger spatial and geological shifts.Source: §1 · p. 2
  • DrillBench provides 49,671 drillholes and a roughly 112k-hole pretraining archive for next-layer prediction and free-running generation across spatial and cross-province shifts.Next-layer prediction uses ground-truth history; generation recursively predicts the remaining lithology from an observed prefix.Source: §1 · p. 2
  • Retrieval helps most in weathered cover, whereas pretraining contributes more in bedrock and under broader geological transfer.The recipe adds neighbouring lithology at inference without directly exposing absolute coordinates, alongside pretraining on historical drillholes.Source: §1 · p. 2

2 RELATED WORK

Existing subsurface modelling either relies on spatial relationships or uses sequence data without matching the one-directional drilling process. Public benchmarks have also been geographically restricted, motivating autoregressive evaluation across graded shifts.

2 RELATED WORK

  • Drillhole modelling is dominated by coordinate- and neighbour-based interpolation and simulation, while well-log methods largely use masked modelling or point-wise classification.These approaches do not directly evaluate rolling a partially drilled hole forward from its own predictions.Source: §2 · pp. 2–3
  • Existing public benchmarks are generally small and geographically restricted, limiting study of graded transfer, large-scale pretraining, and retrieval-based use of neighbouring drillholes.The paper addresses these gaps with autoregressive prediction and generation across spatial and geological shifts.Source: §2 · p. 3

3 BENCHMARK PROTOCOL

DrillBench represents holes at layer and cell granularities and evaluates both next-layer prediction and free-running generation. The tasks distinguish predicting at known boundaries from recursively generating a hole without known future boundaries.

3.1 DRILLHOLE REPRESENTATION

  • Each drillhole has a compact layer-level sequence and a cell-level sequence over depth, accommodating autoregressive and depth-specific modelling approaches.The figure introduces the two granularities and task definitions.Source: §3.1 · p. 3, Figure 1 · p. 3

3.2.1 TASK 1: NEXT-LAYER PREDICTION

  • At true stratum boundaries, next-layer prediction uses only information available up to that boundary and predicts the next lithology.Ground-truth boundaries isolate lithology prediction from boundary or thickness prediction; macro-F1 is reported over classes with at least 20 events per split.Source: §3.2.1 · p. 3

3.2.2 TASK 2: FREE-RUNNING GENERATION

  • Free-running generation starts from the true first p% of a hole and recursively predicts the remaining layers until the known hole depth.Unlike next-layer prediction, it supplies no ground-truth boundaries, so errors can accumulate; joint KL compares lithology and layer-thickness distributions across five rollouts per hole, with lower values better.Source: §3.2.2 · p. 4

4 DRILLBENCH DATASET

DrillBench is curated from Western Australian exploration logs and represents holes through lithology intervals alongside collar metadata and geochemistry. Its benchmark is paired with a larger archive used for pretraining and retrieval, with evaluation holes and nearby twins excluded from that archive.

4.1 SOURCE CORPUS

  • The source corpus covers the Eastern Goldfields and Pilbara and contains 49,671 drillholes and 1,499,181 depth intervals collected over 30 years.The paper describes it as the first public drillhole corpus formatted explicitly for one-dimensional autoregressive sequence modelling.Source: §4.1 · p. 4

4.4 DRILLBENCH SPLITS

  • Figure 2 presents the spatial layout of DrillBench's six splits, and Table 1 summarizes those splits.The benchmark evaluates from in-distribution prediction through increasing spatial shift to cross-province transfer.Source: Figure 2 · p. 4, Table 1 · p. 5

4.2 DATASET CURATION

  • Curation removes invalid and shallow or poorly logged holes, normalizes lithology to 37 codes, harmonizes coordinates, and removes evaluation twins.Twin-hole de-duplication removes 1,051 duplicates from validation and test, leaving 49,671 drillholes and 1,499,181 intervals.Source: §4.2 · p. 4

4.3 DRILLHOLE DATA MODALITIES

  • Hole records pair collar metadata with depth intervals containing lithology labels and geochemical measurements for 25 elements.Each geochemical vector has a binary mask indicating which elements are available.Source: §4.3 · pp. 4–5

4.4 DRILLBENCH SPLITS

  • WA-DASC-full contains 498,713 drillholes across 30 tectonic-province labels; about 112k holes with at least two lithology layers are used for pretraining.The archive is used for pretraining and retrieval rather than benchmark evaluation, and evaluation holes plus archive twins within 500 m are excluded from those uses.Source: §4.4 · p. 5

5 A TRANSFER-SAFE TRAINING-AND-INFERENCE RECIPE

The recipe combines layer-level autoregression, archive pretraining, and retrieval of neighbouring lithology while keeping absolute coordinates and assays out of the prediction input. Fine-tuning and evaluation place these predictions in a shared cell-level format across next-layer prediction and free-running generation.

5 A TRANSFER-SAFE TRAINING-AND-INFERENCE RECIPE

  • The recipe combines a layer-level autoregressive backbone, large-scale pretraining, and spatial retrieval, using coordinates only to retrieve neighbours.Absolute coordinates and assay values are not provided in the prediction input.Source: §5 · p. 5

5.1 LAYER-LEVEL AUTOREGRESSIVE MODEL

  • The backbone predicts each next layer’s lithology and thickness from the observed layer sequence, then appends the prediction for autoregressive generation.A lithology head predicts a categorical distribution; a separate head predicts thickness with a log-normal distribution.Source: §5.1 · p. 6

5.2 PRETRAINING ON UNCURATED ARCHIVES

  • Pretraining uses gap tokens to retain the order of incomplete logs and auxiliary assay objectives to learn from intervals without reliable mapped lithology.Assays are used only for pretraining and removed during fine-tuning and inference.Source: §5.2 · p. 6

5.3 SPATIAL RETRIEVAL WITHOUT EXPLICIT COORDINATE INPUTS

  • Retrieval combines votes from nearby logged holes at the corresponding elevation, or depth when elevation is unavailable, and fuses their lithology context with the layer representation.The retrieval pool excludes evaluation holes and skips neighbours within 25 m; no valid vote produces a zero retrieval signal.Source: §5.3 · p. 6

5.4 FINE-TUNING, DECODING, AND EVALUATION

  • Fine-tuning removes the chemistry pathway, jointly learns the retrieval encoder, and zeros retrieval for 30% of training holes so the model can fall back on sequence context.Evaluation converts predictions to a shared cell-level output and tests both next-layer prediction and autoregressive generation from a hole’s upper half.Source: §5.4 · p. 7

6 RESULTS AND DISCUSSION

Across next-layer prediction and free-running generation, autoregressive neural models are more stable under spatial shift than spatial methods, while retrieval and pretraining contribute differently across local and broader shifts. The results identify a transfer boundary for direct conditioning and show remaining weaknesses on poorly supported lithologies.

6.1 BENCHMARK PERFORMANCE

  • Across both tasks, layer-level autoregressive models degrade more gradually under shift; retrieval favors local prediction, while pretraining and retrieval provide complementary transfer gains.For next-layer prediction, GRU+R reaches 70.0% macro-F1 in-distribution; pretraining gives the best OOD-R and Pilbara-test results, and pretraining plus retrieval gives the best OOD-D result at 37.2%. For generation, GRU+PT+R has joint KL of 0.408 on OOD-D and 0.458 on Pilbara-test.Source: Table 2 · p. 7, §6.1 · pp. 7–8

6.2 TRANSFER BOUNDARY AND MITIGATION STRATEGIES - OUR TRANSFER SAFE RECIPE

  • Direct location and chemistry conditioning raises TEST performance but drops below 3% on PILBARA-TEST, while the lithology-only model stays above 20% on shifted splits.Retrieval helps most in-distribution; pretraining helps on every shifted split, particularly in bedrock and under broader geological shift.Source: Figure 4 · p. 8, §6.2 · p. 8

6.3 FAILURE MODES AND CASE STUDIES.

  • Cross-province errors concentrate in lithologies with little source-region support, and every zero-recall class on PILBARA-TEST has at most 71 training layers.The passage attributes the difficulty to class-support shift: rarely observed lithologies remain hard to recover from sequence context alone.Source: §6.3 · p. 8, Figure 5 · p. 9
  • Without retrieved context, zero-shot GPT-6 rises from 30.5% to 47.4% top-1 accuracy under cross-province shift, slightly above the trained model’s 44.9%.The authors note that some performance may reflect exposure to public DASC records during pretraining.Source: §6.3 · p. 9
  • Drillhole case studies show retrieval helping when nearby holes provide relevant structure and pretraining becoming more useful as geological shift increases.Free-running examples also retain lithology-identity and layer-boundary errors as predictions accumulate.Source: §6.3 · p. 9

7 CONCLUSION

The conclusion presents DrillBench and a transfer-safe recipe as steps toward evaluating and improving autoregressive drillhole modelling across spatial and geological shifts. Its central finding is that sequence structure learned from historical archives transfers where location-based conditioning does not.

7 CONCLUSION

  • The paper introduces DrillBench to evaluate next-layer prediction and free-running generation, and reports that location- and assay-based conditioning can create a transfer boundary.The study targets settings with strong spatial and geological shifts.Source: §7 · pp. 9–10
  • The proposed recipe combines layer-level autoregression, large-scale geological pretraining, and spatial retrieval to use local evidence and broader geological knowledge.Retrieval exploits local continuity, while pretraining supports broader geological transfer.Source: §7 · p. 10
  • The conclusion argues that subsurface modelling needs mechanisms designed around geological structure, data sparsity, and regional shift, not only stronger sequence models.Source: §7 · p. 10

A SCOPE AND LIMITATIONS

The benchmark’s scope reflects choices about available exploration records, provincial coverage, and model scale. These choices constrain what the reported results establish and identify extensions for future work.

A SCOPE AND LIMITATIONS

  • Depths and thicknesses are measured along company drillholes without orientation correction, so thicknesses are apparent rather than true.Lithology labels come from rule-based harmonisation of free-text descriptions.Source: A SCOPE AND LIMITATIONS · p. 13
  • Cross-province evaluation covers Pilbara; extending the benchmark to other WA-DASC provinces requires reconciling company vocabularies, since only about 15% of intervals map to the target vocabulary.Source: A SCOPE AND LIMITATIONS · p. 13
  • The study keeps models small and inputs to lithology so gains can be attributed to the formulation rather than scale or region-specific signals.Larger backbones and transferable encodings of location and chemistry are identified as next steps.Source: A SCOPE AND LIMITATIONS · p. 13
  • The paper identifies orientation-corrected depths and per-hole likelihood evaluation of generation as natural next steps.Source: A SCOPE AND LIMITATIONS · p. 13

B.1 DATASET CURATION DETAILS

The appendix documents data cleaning and harmonisation, then specifies the statistical, geostatistical, feature-based, and autoregressive baselines alongside the proposed layer-level model.

B.1 DATASET CURATION DETAILS

  • The dominant curation step filters holes by depth, interval count, substantive lithology text, and vocabulary coverage; coordinates are harmonised and duplicate twins removed.Unmapped lithology remains unobserved rather than receiving a fallback code, while assay values beyond physical bounds are masked.Source: §B.1 · p. 14

B.2 LITHOLOGY VOCABULARY DETAILS

  • The shared 37-code vocabulary makes provider descriptions comparable; unmapped labels remain unobserved, and COVER/BEDROCK zones are used only for analysis.COVER includes transported material and weathering products such as saprolite and clay; BEDROCK covers fresh lithologies.Source: §B.2 · p. 15, Table 7 · p. 16

C.1 BASELINE IMPLEMENTATION DETAILS

  • The baseline suite includes cell- and layer-level methods, with cell methods operating on 1 m sequences and layer methods on stratum sequences.Count tables generally use add-one smoothing, and UNK receives smoothing mass but is excluded from count estimation.Source: §C.1 · p. 16, §C.1.1 · p. 16

C.1.1 FREQUENCY AND MARKOV MODELS

  • Frequency and Markov baselines range from context-free marginals to sequence models, including order-3 back-off and layer transitions with thickness hazards.Layer Markov uses hierarchical order-2 back-off; layer marginal provides a generation reference without sequence dependence.Source: §C.1.1 · pp. 16–17

C.1.2 GEOSTATISTICAL MODELS

  • Geostatistical baselines either use TRAIN-derived regional patterns or retrieve neighbouring holes; k-NN, kriging, and SIS use depth-wise neighbour evidence.MPS-lite matches up to five preceding tokens and backs off to shorter contexts when patterns are sparse.Source: §C.1.2 · pp. 17–18

C.1.3 FEATURE-BASED MACHINE LEARNING

  • XGBoost is the feature-based baseline, using prediction-time coordinates, region, depth, run length, and previous lithologies.Source: §C.1.3 · p. 18

C.1.4 CELL-LEVEL AUTOREGRESSIVE MODELS.

  • Five neural baselines model the 1 m lithology sequence using history alone and train with teacher forcing and cross-entropy.The set includes recurrent, attention, convolutional, and state-space architectures.Source: §C.1.4 · pp. 18–19

C.1.5 OUR LAYER-LEVEL MODEL.

  • The proposed layer-level model combines a small sequence backbone with optional retrieval and pretraining, while withholding thickness and depth inputs as province fingerprints.Retrieval uses inverse-distance-weighted lithology from ten neighbours; pretraining uses a deduplicated multi-province corpus and auxiliary thickness, masked-token, and chemistry-anchor losses.Source: §C.1.5 · p. 19

C.2 EVALUATION PROTOCOL DETAILS

The protocol puts cell- and layer-level methods into a shared evaluation space and measures both next-stratum prediction and free-running generation. Across the reported comparisons, the combined recipe retains local performance while improving shifted-split results, with retrieval and pretraining contributing differently across settings.

C.2 EVALUATION PROTOCOL DETAILS

  • Task 1 evaluates next lithology at true stratum starts using macro-F1, while Task 2 generates from a true 50% prefix and scores joint lithology-thickness distributions.The joint KL metric avoids treating independently matched lithology and thickness marginals as sufficient for realistic stratigraphy.Source: §C.2 · p. 20

D.1.1 PROTOCOL AND UNCERTAINTY

  • Paired comparisons use shared hole resamples, while generation confidence intervals resample whole reports to account for holes sharing projects and logging vocabularies.Source: §C.2 · p. 20, §D.1.1 · p. 21

D.1.2 ALL BASELINES

  • Spatial baselines reach 66–72% macro-F1 on TEST but fall to 1–5% on shifted splits with the training pool and only 6–12% with the archive pool.Location-conditioned networks also reverse from local gains to cross-province losses; geochemistry shows a milder version of that transfer boundary.Source: §D.1.2 · p. 21

D.1.3 PAIRED DIFFERENCES

  • The full recipe improves over the plain GRU on all four splits and is 23–36 points ahead of the strongest spatial baselines under shift.For OOD-R and PILBARA-TEST, its lower macro-F1 than GRU + PT is not significant; under shift, it leads on top-1, balanced accuracy, support-weighted F1, and Cohen’s κ.Source: §D.1.3 · pp. 21–22

D.1.4 BACKBONES AND TOKEN GRANULARITY

  • Across five layer-level backbones, retrieval adds 10–11 macro-F1 points on TEST while pretraining adds 2.1–4.2 on PILBARA-TEST; cell-level pretraining instead lowers macro-F1 on three splits.The authors attribute the layer-level recipe’s advantage to avoiding the CONT-dominated pretraining signal at one-metre resolution.Source: §D.1.4 · p. 22

D.1.5 COVER AND BEDROCK

  • Retrieval gains are strongest in cover within the Eastern Goldfields, whereas pretraining gains are generally larger in bedrock on shifted splits, with OOD-D an exception.Across provinces, retrieval adds only +0.8 in both zones; the proposed explanation for OOD-D is explicitly untested.Source: §D.1.5 · p. 22

D.1.6 ERROR ANALYSIS

The full recipe’s cross-province gains are strongest on rare bedrock classes but remain uneven across Pilbara reports. In generation, its advantage is concentrated on the most shifted splits, while component effects vary by split and backbone.

D.1.6 ERROR ANALYSIS

  • On PILBARA-TEST, the recipe gains recall on rare bedrock classes, but TOPS and chert remain unsolved and project-level wins are uneven.It beats layer Markov in 20 and loses in 18 of 43 reports; the +7.9-point margin is carried by a few large reports.Source: §D.1.6 · p. 22, Figure 6 · p. 23

D.2.1 PROTOCOL AND UNCERTAINTY

  • The protocol generates each hole’s lower half from its true upper half and uses report-cluster bootstrap intervals, which are wide on OOD-D.A single camp contributes more than half of OOD-D holes, so resampling can omit or double it.Source: §D.2.1 · p. 23

D.2.2 ALL BASELINES

  • On shifted splits, layer-level neural models lead generation; location-conditioned cell-level networks produce especially implausible rollouts under stronger shift.GRU + location has KL 2.64 on OOD-D and 6.02 on PILBARA-TEST, versus 0.56 and 1.00 without coordinates.Source: §D.2.2 · p. 23

D.2.3 PAIRED DIFFERENCES

  • On OOD-D and PILBARA-TEST, the full recipe lowers joint KL versus the plain GRU significantly only on PILBARA-TEST, where both pretraining and retrieval contribute.Against the layer-level Transformer it is better on both shifted splits; its advantage is largest against fitted and cell-level baselines.Source: §D.2.3 · pp. 23–24

D.2.4 BACKBONES

  • Across five layer-level backbones, no per-component generation gain holds for every backbone, and the ordering of augmented arms is unstable with three seeds.The GRU is used in the main text because it is the most even across the two tasks.Source: §D.2.4 · p. 24

D.2.5 OTHER METRICS

Alternative metrics preserve the main generation comparisons but reveal that pooled divergence does not measure fit to each hole’s prefix. The LLM baselines use a shared prediction mask and prompts, with different read-outs for open-weight and hosted models.

D.2.5 OTHER METRICS

  • Under KL, KL-13, JSD and report-balanced KL, the full recipe leads on OOD-D and PILBARA-TEST, while layer Markov leads on TEST.Per-hole metrics are flat across layer-level rows, and a pooled law does not score each continuation’s fit to its own prefix.Source: §D.2.5 · p. 24

D.3.1 PIPELINE

  • LLM predictions are scored at the same layer-start points as other methods, using only layers strictly above each point as context.A whole-hole prompt would leak later target layers, so each point is an independent request.Source: §D.3.1 · p. 25

D.3.2 PROMPTS

  • The prompt distinguishes unlogged gaps from the UNKN lithology code and spells out the 37 candidate codes.Cold starts receive a dedicated prompt; retrieval prompts summarize nearby lithology, coverage and distance.Source: §D.3.2 · pp. 25–26

D.3.3 OPEN-WEIGHT MODELS: LIKELIHOOD SCORING

  • Open-weight models rank candidate codes by teacher-forced likelihood, with domain-conditional PMI correcting for each code’s surface preference.The neutral prompt removes hole evidence while retaining the system prompt and question.Source: §D.3.3 · p. 26

D.3.4 HOSTED MODELS: GENERATIVE SCORING

  • Hosted models return top-three codes in JSON, and rows with unparseable responses exclude those points from their denominator.Hosted rows are compared with the full recipe under the shared top-1 mask, not directly with likelihood-scored rows.Source: §D.3.4 · p. 27

D.3.6 LLM LIMITATIONS

Hosted-model comparisons are qualified by possible training-data contamination, structured-output failures, sampling uncertainty and cost. Component ablations show retrieval and pretraining contribute differently across prediction and generation settings.

D.3.6 LLM LIMITATIONS

  • Hosted-model cross-province scores may partly reflect recall of public WA-DASC records because closed training corpora cannot be audited.The caveat applies to every hosted row.Source: §D.3.6 · p. 27
  • Gemini-pro-latest failed to parse 65.2% of retrieval-arm points, leaving a biased subset of responses.It failed on up to 97% of Pilbara test points.Source: §D.3.6 · p. 27
  • LLM hole-sampling intervals are about ±3–4 points, so smaller differences should not be treated as an ordering.Likelihood and greedy read-outs are deterministic given the prompt, but hole-sampling variance remains.Source: §D.3.6 · p. 27
  • One GPT-6-astra evaluation pass consumed 3.36 million input tokens, while the 0.68-million-parameter full recipe runs locally.A likelihood-scored open-weight arm takes about two hours on one GPU.Source: §D.3.6 · p. 27

D.4.1 WHAT EACH COMPONENT CONTRIBUTES

  • For next-layer prediction, retrieval yields local and distal-ring gains but no measurable cross-province gain, whereas pretraining helps every shifted split.For generation, both components help on distal and cross-province splits, and the full recipe is the only arm with significant gains on both.Source: §D.4.1 · p. 28

D.4.2 PRETRAINING

Pretraining transfer depends on corpus size, optimisation and geological diversity. Fine-grained lithology targets matter, and pretraining is particularly valuable when labelled training data are scarce.

D.4.2 PRETRAINING

  • Under shift, pretraining benefits grow with corpus size, but the distal ring benefits from thousands of holes while PILBARA-TEST requires tens of thousands.In-distribution performance does not depend on corpus size.Source: §D.4.2 · pp. 28–29
  • With optimisation steps held equal, performance saturates at about one-tenth of the corpus, so much of the larger-corpus curve reflects optimisation rather than data.Only the two smallest fractions remain clearly worse.Source: §D.4.2 · p. 29
  • At fixed corpus size, PILBARA-TEST gains respond to province diversity, while the distal ring responds much less and in-distribution scores do not move.The deficit against the full archive at one or two provinces is at least halved at thirty provinces.Source: §D.4.2 · p. 29
  • Most cross-province pretraining gain persists when Pilbara archive holes are held out, and removing Eastern Goldfields as well costs nothing further.With retrieval, 82% of the gain remains with Pilbara held out and 96% when Eastern Goldfields is also held out.Source: §D.4.2 · p. 29
  • Replacing fine-grained lithology targets with a 13-class objective significantly reduces performance on all shifted splits, while chemistry losses show no measurable transfer benefit.The authors retain chemistry objectives because they do no harm and make the archive’s geochemistry channel available.Source: §D.4.2 · p. 29
  • Pretraining helps most when labels are scarce: with 1,000 labelled holes, it lifts every split and beats training from scratch on all TRAIN holes across provinces.Retrieval’s in-distribution gain grows with the labelled set, but it adds nothing on PILBARA-TEST at any size.Source: §D.4.2 · p. 29

D.4.3 RETRIEVAL

Retrieval benefits depend on which neighbours are available and how their lithology is used. Pretraining and retrieval act differently across geological regimes, while stratum tokens and the 1 m grid support the deployed recipe.

D.4.3 RETRIEVAL

  • Across provinces, larger k can dilute neighbour evidence; one neighbour gives about half the distal-ring gain, and distance beats mutual-agreement reranking.In-distribution accuracy keeps rising up to the largest k tried, while across-province performance moves the other way; deployed k=10 sits on the shifted-split plateau.Source: §D.4.3 · pp. 29–30
  • Training-only retrieval leaves in-distribution accuracy unchanged but costs both spatial rings; across provinces it is no worse, and thinning Pilbara archive holes leaves PILBARA-TEST unchanged within noise.The archive supplies the neighbours used in the rings, but its Pilbara holes do not carry the cross-province result.Source: §D.4.3 · p. 30
  • The recipe remains far above spatial baselines on every shifted split, even when those baselines use the same train-and-archive pool.The archive lifts the spatial baselines under every shift, but the sequence model adds value by reading neighbours as context.Source: §D.4.3 · p. 30
  • Letting every neighbour bed vote across its full thickness gives the largest in-distribution gain, but significantly harms PILBARA-TEST.The deployed rule lets neighbours vote only where a bed starts; denser evidence helps where neighbours are right and hurts where they are not.Source: §D.4.3 · p. 30

D.4.4 WHERE EACH COMPONENT ACTS

  • Retrieval mainly helps in cover and pretraining mainly in bedrock, with the distal ring as the exception; this division holds across five backbones on both tasks.Source: §D.4.4 · p. 30

D.4.5 TOKENISATION GRID

  • Retrieval helps with both token units, but pretraining raises macro-F1 on all three shifted splits only with stratum tokens.At cell level, pretraining lowers macro-F1 on most splits; most 1 m targets are CONT, so its signal is dominated by bed thickness rather than lithological succession.Source: §D.4.4 · p. 30
  • Neither 0.5 m nor 2 m spacing significantly improves a split, and each costs on at least one; the benchmark therefore stays at 1 m.The 1 m grid matches source logging resolution; finer spacing can miss bed starts after elevation-aligned lookup, while coarsening erases layer starts.Source: §D.4.5 · p. 30
Loading 2610.01204v1…