Missing well-log intervals make subsurface reconstruction ill-posed. Existing numerical regression models lack geological prior knowledge to constrain ambiguous inverse mappings. We propose GeoLog-Agent, a knowledge-conditioned generative agent using pre-trained large language models (LLMs) as a source of geological priors. A universal Geo-Harness translates LLM stratigraphic reasoning into formation-anchored petrophysical trajectories, feature-wise linear modulation, and conditional adversarial constraints. On the Norwegian North Sea benchmark, GeoLog-Agent achieves a mean R2 of 0.9307 and a peak R2 of 0.9713. It exceeds the nine-baseline average R2 of 0.8387 by 0.0920, with a 30.97% average RMSE reduction, and exceeds the strongest baseline by 0.0274--0.0498 in R2, with 16.25%--24.05% lower RMSE. Ablations confirm the effectiveness and synergy of geological prior reasoning, multi-scale dilated generation, and conditional adversarial modules. Transferring the coupling mechanism to six conventional deep learning models improves their R2 by 0.0162--0.0789 with cross-well stability. These results demonstrate how foundation-model domain reasoning can constrain deep generative reconstruction of missing petrophysical measurements.
Figures & tables
Figure 1: Architecture of GeoLog-Agent and the Geo-Harness coupling bridge: ( a ) two-stage collaborative workflow, ( b ) Stage II knowledge-modulated adversarial training and observation fusion, ( c ) Stage I pre-trained LLM geological prior reasoning, and ( d ) FiLM-modulated cyclic dilated generator G and spectral-normalized discriminator D .
Figure 2: Full-well reconstruction overview and local trajectory comparison: ( a ) observed GR, 50% contiguous missing zone ( 1654.6–2690.0 m ), and reconstructed DEN/AC profiles in Well 1; ( b ) local zoom across the Balder–Sele volcaniclastic transition in Well 1 ( 2236–2326 m , M50) and Well 2 ( 2210–2300 m , M100).
Model
M50 Protocol (50% Contiguous Missing)
M100 Protocol (100% Full-Well Blind)
Overall Mean R2↑
Mean R2↑
DEN RMSE ↓
AC RMSE ↓
Mean R2↑
DEN RMSE ↓
AC RMSE ↓
DLinear
0.7893
0.1254
38.7793
0.5573
0.1754
59.7646
0.6733
Autoformer
0.8588
0.0773
41.7653
0.6347
0.1320
64.5900
0.7468
LSTM
0.8955
0.0750
34.5233
0.8054
0.1022
47.2646
0.8505
BiLSTM
0.9057
0.0715
32.8235
0.8220
0.0955
46.2017
0.8638
TimesNet
0.9166
0.0683
30.3651
0.8249
0.0970
44.9000
0.8708
Table 1: Quantitative comparison of well-log reconstruction across blind test Well 1 (15-9-15) and Well 2 (15-9-17) under M50 and M100 protocols (two-well mean R2 , DEN RMSE in g/cm3 , and AC RMSE in μs/m ; per-well breakdowns are in Table 1 ). Bold indicates the best result.
Figure 3: Multi-window log reconstruction across all ten models: ( a , b ) Well 1 under M50 across windows Z1 ( 2020–2075 m ), Z2 ( 2236–2296 m ), and Z3 ( 2405–2465 m ); ( c , d ) Well 2 under M100 across windows Z1 ( 2210–2270 m ), Z2 ( 2360–2420 m ), and Z3 ( 2450–2510 m ).
Figure 4: Cross-well correlation fidelity and error distributions under M50 and M100: ( a ) predicted-versus-true DEN and AC scatters for the top four models, and ( b ) pointwise absolute error boxplots across all ten models.
Figure 5: Benchmark R2 matrix, cross-backbone transferability, and component ablation: ( a ) R2 heatmap across all eight scenarios, and ( b ) M100 R2 gains when attaching Geo-Harness to six conventional backbones (Panel a) alongside the seven-tier ablation hierarchy of GeoLog-Agent (Panel b).
Table 2: Component-wise ablation study of Stage I geological prior reasoning and Stage II architectural modules in GeoLog-Agent across blind test Well 1 (15-9-15) and Well 2 (15-9-17) under M50 and M100 protocols. Bold values denote the best non-oracle performance; italicized rows indicate the theoretical Oracle upper bound.
Protocol
Backbone Architecture
Vanilla R2
With Geo-Harness R2
ΔR2 Gain
RMSE Reduction (%)
M50
LSTM
0.8955
0.9424
0.0469
26.28%
BiLSTM
0.9057
0.9453
0.0396
24.54%
PatchTST
0.9178
0.9479
0.0301
21.04%
Informer
0.9209
0.9490
0.0281
20.30%
Transformer
0.9192
0.9480
0.0288
20.43%
GAN
0.9357
0.9519
0.0162
14.17%
Table 3: Cross-backbone transferability of the pre-trained geological knowledge coupling mechanism ( Geo-Harness ) attached to six conventional deep learning models (averaged across DEN and AC on blind test Wells 15-9-15 and 15-9-17).
Figure 6: Sensitivity analysis and downstream reservoir evaluation: ( a ) R2 across the 20%–100% missingness spectrum, ( b ) Stage I prior sensitivity on Macro-F1 and sandstone/tuff recall, ( c ) Balder–Sele acoustic impedance ( Zp ) profile in Well 1 ( 2236–2306 m ), and ( d ) downstream Zp , porosity ( ϕD ), and pay-zone Macro-F1 accuracy.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Protocol
Model
Bulk Density (DEN, g/cm3 )
Compressional Transit Time (AC, μs/m )
Well 1 (15-9-15)
Well 2 (15-9-17)
Well 1 (15-9-15)
Well 2 (15-9-17)
R2↑
MAE ↓
RMSE ↓
R2↑
MAE ↓
RMSE ↓
R2↑
MAE ↓
RMSE ↓
R2↑
MAE ↓
RMSE ↓
M50
DLinear
0.7343
0.0711
0.1215
0.6998
0.0743
0.1293
0.8712
20.9275
42.4008
0.8519
17.9351
35.1579
Autoformer
0.9057
0.0377
0.0724
0.8786
0.0446
0.0822
0.8936
17.8242
38.5345
0.7574
26.8737
44.9962
LSTM
0.8974
0.0387
0.0755
0.9002
0.0361
0.0745
0.8846
18.7295
40.1211
0.8998
14.3537
28.9255
BiLSTM
0.9099
0.0372
0.0708
0.9065
0.0353
0.0722
0.8872
18.8382
39.6770
0.9192
12.6309
25.9700
Appendix
Table 1: Consolidated per-well and per-curve reconstruction performance for bulk density (DEN, g/cm3 ) and compressional acoustic transit time (AC, μs/m ) under the M50 (50% contiguous missing, evaluated on the full-well reconstructed profile Y^full after observation-preserving fusion) and M100 (100% full-well blind) protocols across blind test Well 1 (15-9-15) and Well 2 (15-9-17). Bold values indicate the best performance.
Model
Acoustic Impedance Zp ( MPa⋅s/m )
Reservoir Porosity ϕD (p.u.)
Pay-Zone Delineation
R2↑
RMSE ↓
R2↑
RMSE ↓
Macro-F1 (%) ↑
Accuracy (%) ↑
DLinear
0.5904
1.7655
0.3710
9.9259
38.39
71.99
Autoformer
0.6134
1.7342
0.7240
6.4740
57.38
78.68
LSTM
0.7654
1.3422
0.7998
5.6444
56.94
84.55
BiLSTM
0.7743
1.3072
0.8115
5.4497
58.19
85.20
TimesNet
0.7841
1.2753
0.8287
5.2316
59.44
85.67
Appendix
Table 2: Downstream reservoir characterization and seismic inversion performance evaluated directly on the reconstructed target intervals ( Mt,j=1 , averaged across M50 and M100 on blind test Wells 15-9-15 and 15-9-17; effective porosity ϕD=(ρma−ρb)/(ρma−ρfl) with ρma=2.65 g/cm3,ρfl=1.00 g/cm3 , and four-zone pay delineation defined by standard North Sea shale-volume, porosity, and resistivity cutoffs).
Figure 1: Multi-well grouped bar comparisons across M50 and M100 scenarios: ( a ) MAE and RMSE, and ( b ) R2 across Well 1 (15-9-15) and Well 2 (15-9-17).
Figure 2: True versus predicted hexbin density scatter matrices across all ten models for Well 1 (15-9-15) under M50: ( a ) AC and ( b ) DEN.
Figure 3: True versus predicted hexbin density scatter matrices across all ten models for Well 2 (15-9-17) under M50: ( a ) AC and ( b ) DEN.
Figure 4: Supplementary error and boundary diagnostics: ( a ) per-scenario model ranking matrix, ( b ) depth-sequence DEN absolute error heatmap in Well 1 (15-9-15), and ( c ) boundary transition across 2690.0 m .
Figure 5: Normalized six-axis radar comparisons of the top five models under ( a ) M50 and ( b ) M100 protocols.
Missing wireline logs hinder subsurface characterization. Deep reconstruction networks suffer from gap-boundary steps and cross-basin shifts, while interpolation and Gaussian smoothers struggle with long intervals and outliers. We propose boundary-gated soft-thresholding Kalman fusion (ST-KF), unifying multiresolution wavelet shrinkage, distance-gated residual prior learning, and closed-form robust Kalman smoothing with learning--inference decoupling. Across two public benchmarks totaling 263 wells in the Norwegian North Sea and New Zealand's Taranaki Basin, ST-KF ranks first over 0.15--100-m anchored gaps with 0.09% parameter overhead. Against the strongest of nine baselines, pooled RMSE falls by 5.0%--8.3% for bulk density and 6.4%--15.6% for compressional slowness; against the four strongest baselines, reductions are 0.7%--4.1% for neutron porosity and 1.3%--2.9% for photoelectric factor. Ablations and sensitivity analyses confirm operator synergy and stable performance: removing state-space fusion increases short-gap errors by up to 58.8%, while zero-shot transfer cuts anchored errors by 14%--68%. The results show that local borehole continuity transfers across basins whereas cross-property mappings shift with regional geology, supporting measurement-guided log reconstruction.
Yongan Zhang, Zhihan Qiu, Yizheng Li +1
State Key Laboratory of Climate Resilience for Coastal Cities, Department of Building Environment and Energy Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China · Shandong University of Science and Technology, Qingdao, China · School of Computer Science and Technology, China University of Petroleum (East China), Qingdao 266580, China
Accurate geological characterization of subsurface reservoirs from well log data is essential to support projects such as carbon capture and storage (CCS), geothermal development, and extraction of natural resources. Existing automated techniques for geological characterization primarily use sliding-window classification, which limits their ability to understand broader geological contexts, often leading to misaligned formation layers. To overcome these limitations, we introduce LithoFormer, a robust framework for stratigraphic inference using a Seq2Seq transformer model that ingests entire multivariate well logs in a single pass. The framework utilizes a channel-independent PatchTST backbone enhanced with rotary positional embeddings (RoPE) to capture long-range geological dependencies across entire multivariate well logs. A decoupled multi-task head is employed to jointly predict geological zonation and precise boundary probabilities, while a geology-informed loss function enforces physical constraints such as the Law of Superposition. Validated and deployed on three real-world datasets, LithoFormer demonstrates a 90% reduction in median boundary error and eliminates stratigraphic order violations compared to traditional sliding-window baselines. It also achieves a 80% reduction in manual expert labor and eliminates stratigraphic inconsistencies, providing a scalable and reliable solution for large-scale subsurface modeling.
Laboratoire Interdisciplinaire des Sciences du Numérique, Université Paris-Saclay, SLB Montpellier, France · Laboratoire Interdisciplinaire des Sciences du Numérique, CentraleSupélec, CNRS, Université Paris-Saclay Gif-sur-Yvette, France · SLB Montpellier, France
While general-purpose Large Language Models (LLMs) applied to Geology often hallucinate when reasoning about subsurface structures and deep-time evolution, current AI in Earth sciences predominantly targets surface remote sensing and GIS. To bridge this gap, we introduce Geo-Expert, a family of parameter-efficient geological LLMs fine-tuned on a custom-curated, high-quality instruction dataset processed using our custom instruction synthesis pipeline. We investigate the impact of model scaling and architecture by fine-tuning three base models: Qwen3-8B, Qwen3-32B, and Gemma-3-27B, with Low-Rank Adaptation (LoRA) method. Our extensive evaluation on a novel domain-specific benchmark, Geo-Eval, reveals that a domain-aligned 8B model can outperform open-weight 70B generalists and proprietary GPT-4o on specialized geological reasoning, while a 32B variant approaches frontier reasoning models. The optimized 8B model further offers a competitive cost-performance ratio for deployment. This work provides a reproducible recipe for democratizing scientific LLMs and establishes a baseline for geological artificial intelligence.
Chenyou Guo, Zongqi Liu, Yizhou Zhang +2
Ocean University of China · Peking University · Monash University