Robust Well-Log Reconstruction Across Point Gaps and Whole Wells via Boundary-Gated Soft-Thresholding Kalman Fusion
Authors: Yongan Zhang, Zhihan Qiu, Yizheng Li, Youzhuang Sun
Organizations: 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
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.
Figures & tables
Fig. 1: Architecture of ST-KF and its training–inference decoupling. (a) Multiscale wavelet denoising of predictor logs and measured anchors. (b) Distance-gated anchor bridge on the Transformer backbone. (c) Soft-thresholding Kalman fusion solved in closed form. Red dashed boxes mark inference-only despiking (a) and training-only loss supervision (c).
Fig. 2: Study areas, well partition and gap protocol used to test whether one model generalizes across wells, regions and gap lengths. (a) FORCE 2020 wells in the northern North Sea coloured by split. (b) Taranaki Basin wells used for zero-shot transfer. Maps use WGS84 coordinates (FORCE locations converted from ED50/UTM zone 31N) and Natural Earth coastlines, and insets show the regional setting. (c) Depth ranges of the lithostratigraphic groups in the FORCE wells. (d) Gap intervals imposed by each scenario on one validation segment.
In-distribution blind (18 wells)
Spatially blind (10 wells)
Official hidden (10 wells)
RHOB
DTC
RHOB
DTC
RHOB
DTC
Method
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
RMSE
R2
ST-KF
0.0663
0.885
5.374
0.951
0.0585
0.932
4.464
0.974
0.0508
0.930
4.375
0.977
Transformer
0.0706
0.870
5.741
0.944
0.0617
0.925
5.114
0.966
0.0554
0.917
5.345
0.966
SAITS
0.0704
0.871
5.835
0.942
0.0640
0.919
5.249
0.965
0.0557
0.916
5.181
0.968
BiGRU
0.0753
0.852
6.090
0.937
0.0675
0.910
5.545
0.961
0.0594
0.904
6.228
0.954
TABLE I: Pooled reconstruction accuracy over the 13 gap scenarios with anchors ( whole-well loss excluded) on the three blind test sets. RMSE in g/cm 3 (RHOB) and μ s/ft (DTC), mean of five seeds, with best values in bold.
Fig. 3: Showcase of how ST-KF reconstructs short gaps between measured anchors in blind wells. Each row shows a 30-m context track with the gap marked by a red dashed box (left) and the magnified gap with its observed neighbourhood (right). Black dots mark the measured anchors and shading separates the observed sections (blue) from the gap (cream). (a) 1-m and (b) 5-m RHOB gaps, (c) 1-m and (d) 3-m DTC gaps. The four cases were selected by a fixed rule as the gaps with the largest advantage of ST-KF over all baselines (top 0.5% of 3638 candidate gaps), and representative statistics are given in Fig. 4 and Table II.
Fig. 4: Dependence of reconstruction error on gap length for the three blind test sets. Rows show RHOB and DTC, and shaded bands are 95% well-level bootstrap intervals. Gap lengths range from one to three samples (0.15–0.46 m) to 100 m.
RHOB (g/cm 3 )
DTC ( μ s/ft)
Scenario
ST-KF
Transf.
SAITS
BiGRU
U-Net
Linear
ST-KF
Transf.
SAITS
BiGRU
U-Net
Linear
Point (1–3 samples)
0.0216
0.0354*
0.0418*
0.0274
0.0370*
0.0214
1.397
3.153*
3.522*
2.722*
3.971*
1.438
1 m
0.0326
0.0401*
0.0416*
0.0358
0.0437*
0.0348
2.451
3.797*
3.990*
3.546*
4.759*
2.543
3 m
0.0352
0.0411*
0.0417*
0.0407*
0.0447*
0.0487*
3.104
4.099*
4.019*
4.002*
4.856*
4.580*
5 m
0.0453
0.0509*
0.0503*
0.0527*
0.0595*
0.0734*
3.417
3.997*
4.205*
4.310*
4.991*
4.704*
10 m
0.0440
0.0488*
0.0482*
0.0503*
0.0527*
0.0648*
3.757
4.306
4.310*
4.826*
5.254*
5.579*
TABLE II: Scenario-level RMSE on the official hidden wells (mean of five seeds). * and † mark ST-KF as significantly better or worse than that baseline ( well-paired Wilcoxon test, Holm-corrected p<0.05 ), and best values are in bold. Noise level in standardized units of the predictor logs. The other two test sets are in Supplementary Tables S1–S2.
Fig. 5: Reconstruction of a 20-m gap through interbedded sandstone, shale and limestone in blind well 31/3-3 (1885.6–1905.3 m). Tracks show (a) lithology, (b) gamma ray, (c) neutron porosity, (d) RHOB and (e) DTC. In (d) and (e) the observed sections are shaded blue, the gap cream, and black dots mark the gap anchors. The gap was selected by a fixed rule among 273 candidate 20-m gaps. Five gaps in which ST-KF has R2≥0.7 and the lowest error on both logs qualified, and the one with the most varied lithology was taken.
Fig. 6: Reconstruction under corrupted anchors, noisy predictor logs and a new basin. (a) Spiky anchors next to a gap and the firm outlier term. (b) Point-wise transition coefficient along the same interval, with injected spikes and the largest true step marked. (c) DTC RMSE versus predictor-noise level for 20-m and whole-well gaps, averaged over the three blind test sets. (d) Well-level RMSE in the Taranaki Basin without fine-tuning .
Model
Params
Train (min)
ms/well
DTC RMSE
ST-KF (frozen)
597,793
13.9
22.4¶
4.737
ST-KF w/o inference-time modules
597,793
13.9 ± 1.0
17.8
4.771
Transformer
597,250
7.1 ± 0.7
5.4
5.400
SAITS
540,798
24.1 ± 5.5
24.7
5.422
BiGRU
407,554
6.5 ± 1.0
10.8
5.954
BiLSTM
543,234
7.7 ± 1.1
22.4
6.905
TABLE III: Computational cost on one NVIDIA RTX 5080, including parameters, training time per seed (mean ± s.d. over five seeds) and inference latency per well (18 in-distribution wells, 20-m gaps), with the mean pooled DTC RMSE of Table I.
≤ 1-m gaps
3–100-m gaps
Whole well
Spiky anchors
Noisy inputs
Variant
RHOB
DTC
RHOB
DTC
RHOB
DTC
RHOB
DTC
RHOB
DTC
(a) Inference-time modules
w/o anchor bridge cleaning
+0.0
-0.0
-0.0
+0.1
+0.0
+0.0
+2.7
+5.7
+0.0
+0.0
w/o anchor cleaning and Firm-IRLS
-0.1
-0.0
-0.0
+0.1
+0.0
+0.0
+3.0
+7.1
+0.0
+0.0
w/o all inference-time modules
-0.1
-0.2
-0.2
-0.0
-0.1
+0.0
+3.0
+7.1
+26.4
+16.5
(b) Trained components
TABLE IV: Ablation on the three blind test sets, given as the mean change of RMSE (%) when a component is removed or replaced (positive values mean worse without the component), averaged over test sets and over the scenarios of each group. (a) Inference-time modules, relative to ST-KF . (b) Trained components, relative to the same network without inference-time modules, with each variant was retrained with five seeds.
Fig. 7: Reliability of the predicted uncertainty and stability of the frozen configuration. (a) Backbone standard deviation versus empirical RMS error on masked samples of the in-distribution blind wells, binned by deciles of σt , with 95% well-bootstrap intervals and the distribution of σt . (b) Sensitivity to the persistence ϕ0 and the threshold floor τmin on the validation wells (mean ±1 s.d. over seeds, flat region shaded, network without inference-time modules). (c) RMSE change of 20-m and whole-well reconstruction relative to disabled input despiking, for each despiking threshold t1 and predictor-noise level (validation wells, mean of three seeds), with the absolute RMSE without despiking in the top row and the frozen t1 in bold. (d) Learned distance gate g for every seed and the resulting DTC RMSE for noisy and cleaned training bridges and for the fixed-gate control (validation wells), with dots marking g one sample from the anchor.
5-m gaps
20-m gaps
Natural pattern
Whole well
Method
RHOB
DTC
RHOB
DTC
RHOB
DTC
RHOB
DTC
ST-KF
0.0373
3.373
0.0462
4.254
0.0476
4.168
0.1314
11.255
Transformer
0.0499*
4.281*
0.0594*
5.075*
0.0650*
5.371*
0.1317
11.031†
SAITS
0.0450*
4.301*
0.0538*
5.008*
0.0594*
5.435*
0.1360*
11.386
BiGRU
0.0538*
4.541*
0.0659*
5.530*
0.0680*
5.632*
0.1352*
11.507
Linear interp.
0.0498*
4.295*
0.0654*
5.873*
0.0680*
5.356*
–
–
TABLE V: Zero-shot RMSE on 149 Taranaki Basin wells (models trained on FORCE 2020 only, mean of five seeds). Markers as in Table II.
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.
Yongan Zhang, Yitao Yu, Yizheng Li +2
College of Computer Science and Technology, China University of Petroleum (East China), Qingdao, 266580, China
Learning from scientific measurements often requires aligning modalities with different spatial support and resolution. Subsurface characterization is an extreme dense-sparse problem in which 3D seismic provides volumetric but indirect measurements and well logs provide high-resolution 1D measurements at sparse borehole locations. We introduce a physically grounded cross-modal masked autoencoder (CM-MAE) for seismic-to-well representation learning. The model jointly tokenizes seismic volumes and well-log depth patches, embeds both modalities in continuous physical coordinates using four-axis rotary position embeddings, and reconstructs masked targets with a cross-modal decoder. Sparse Mixture-of-Experts layers provide modality-specific capacity while dense attention allows information exchange between modalities. Pretraining uses 23 seismic volumes covering approximately 178,000 square km across U.S. offshore and onshore basins, together with approximately 92,000 wells. Matched-mask ablations show strongly asymmetric information flow. Seismic context improves held-out well-log reconstruction by 9.73%, while well-log context improves seismic reconstruction by 0.65%. We evaluate seismic-only pseudo-log predictions against independent interpreter-drawn salt-geobody masks to determine whether they contain geologic signal. Across offshore U.S. surveys, compressional-slowness-derived salt scores reach an AUROC of 0.910. In onshore basins, predicted compressional slowness preserves formation-scale structure and partial relative organization in an unseen survey despite calibration drift. CM-MAE learns useful seismic-to-well representations under extreme modality asymmetry, although absolute pseudo-log calibration remains survey-dependent.
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