arxiv: 2605.06255 · v1 · submitted 2026-05-07 · ⚛️ physics.geo-ph

Recognition: unknown

Synthetic Well Log Generation with Preserved Multivariate Correlations and Vertical Facies Stacking Patterns

Josue Fonseca , Marcus Saraiva

Authors on Pith no claims yet

Pith reviewed 2026-05-08 03:11 UTC · model grok-4.3

classification ⚛️ physics.geo-ph

keywords synthetic well logsMarkov chainautoencoderMCMC samplingfacies stackingpetrophysical correlationsturbidite reservoirseismic interpretation

0 comments

The pith

A hybrid framework generates synthetic well logs while preserving both petrophysical correlations and vertical facies stacking patterns.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a procedure to create artificial well log data that keeps the statistical links among properties such as density and sonic velocities and also respects the vertical order in which rock types appear. It achieves this by modeling vertical rock-type transitions with Markov chains, compressing the multivariate data with autoencoders, and generating new examples through MCMC sampling in the compressed space. This matters for machine learning in seismic interpretation, where real well measurements are scarce, because the resulting synthetics can support training, scenario testing, and uncertainty work. Tests on a turbidite reservoir dataset confirm that the outputs maintain rock-physics relationships and produce vertical variations matching real logs.

Core claim

The authors establish that integrating Markov chain models for electrofacies stacking patterns with autoencoder-based dimensionality reduction and Markov chain Monte Carlo sampling in latent space produces synthetic well logs that preserve multivariate correlations among petrophysical properties and generate geologically realistic vertical heterogeneity consistent with actual well log measurements from a real turbidite reservoir.

What carries the argument

The hybrid workflow that employs Markov chain models to capture vertical facies transitions, autoencoders to reduce dimensionality of the multivariate log data, and MCMC sampling in the latent space to draw new realizations before decoding.

Load-bearing premise

That combining Markov chain models for facies stacking with autoencoder reduction and MCMC sampling in latent space will capture and reproduce all relevant correlations and patterns without introducing new artifacts or biases absent from the training data.

What would settle it

A direct comparison in which the synthetic logs exhibit cross-property correlations or vertical stacking sequences that deviate measurably from those in the original dataset or violate known rock-physics relations.

Figures

Figures reproduced from arXiv: 2605.06255 by Josue Fonseca, Marcus Saraiva.

**Figure 1.** Figure 1: Real dataset with four wells with: (a) facies classification and (b) transition matrix view at source ↗

**Figure 2.** Figure 2: Three plots representing the input, bottleneck. and output of the autoencoder with the well view at source ↗

**Figure 3.** Figure 3: Using the latent space to perform MCMC sampling through MH algorithm view at source ↗

**Figure 4.** Figure 4: Original dataset (blue dots) versus MCMC samples (red dots) in the real units of the rock properties. view at source ↗

**Figure 5.** Figure 5: presents the generated ensemble of synthetic well logs spanning 250 meters of depth, showing simulated electrofacies profiles alongside corresponding density, P-Velocity (Vp), and S-Velocity (Vs) curves. The synthetic logs exhibit realistic vertical facies organization, property-facies consistency, highfrequency variability, and appropriate inter-realization variability view at source ↗

read the original abstract

We present a novel procedure for generating synthetic well logs that simultaneously preserves multivariate correlations among petrophysical properties (Density, P-Sonic, S-Sonic) and vertical stacking patterns of electrofacies. The methodology integrates Markov chain models, autoencoder-based dimensionality reduction, and Markov chain Monte Carlo (MCMC) sampling in latent space. Application to a real turbidite reservoir dataset demonstrates that the framework successfully sustains fundamental rock physics relationships and generates geologically realistic vertical heterogeneity consistent with actual well log measurements. This technique addresses critical data scarcity in machine learning applications for seismic interpretation while enabling credible synthetic seismogram generation for scenario testing and uncertainty quantification in petroleum exploration and field development.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This combines Markov chains for facies order with autoencoder compression and MCMC sampling to generate synthetic well logs that keep petrophysical correlations and vertical patterns, and the real-data results look internally consistent though mostly visual.

read the letter

I looked at the full paper on synthetic well log generation. The central contribution is a pipeline that first fits a Markov chain to capture electrofacies transitions, trains an autoencoder on the joint distribution of density, P-sonic, and S-sonic, then runs MCMC in the latent space to draw new realizations. This produces logs that respect both the multivariate correlations and the vertical stacking seen in the training data from a turbidite reservoir. The figures show side-by-side comparisons where the synthetics reproduce realistic heterogeneity and maintain rock-physics trends without obvious artifacts or breaks in the sequences. The method is coherent on its own terms and avoids the circularity that sometimes creeps into latent-space sampling when parameters are fitted too tightly to the same statistics being preserved. What the work does cleanly is demonstrate that the combined model can be applied directly to field data and yield usable outputs for downstream tasks like seismic modeling. The soft spots are mainly in the evaluation. The claims rest on qualitative matches and statements that relationships are sustained, with limited quantitative checks such as distribution distances or direct comparisons against simpler baselines like independent sampling or basic copula methods. No error bars or sensitivity tests on the autoencoder depth or MCMC chain length appear. That keeps the strength of the evidence moderate rather than strong. This paper is for geophysicists and ML practitioners working on data-scarce reservoir problems who need synthetic logs for training or scenario testing. It is coherent enough and addresses a real gap, so it deserves a serious referee even if the authors will likely be asked to add metrics and baselines.

Referee Report

0 major / 2 minor

Summary. The paper claims to introduce a new method for synthetic well log generation that preserves both multivariate petrophysical correlations and vertical facies stacking patterns by combining Markov chain models for stacking, autoencoders for dimensionality reduction, and MCMC sampling in the latent space. It applies this to a real turbidite reservoir dataset and asserts that the generated logs maintain rock physics relationships and realistic heterogeneity.

Significance. Should the method prove robust, it would be significant for mitigating data scarcity in geophysical machine learning applications, particularly for training models on well log data and generating synthetic seismograms for uncertainty analysis in reservoir characterization. The paper's approach of embedding geological knowledge through Markov chains into a data-driven latent space framework is noteworthy and could inspire similar hybrid methods in the field. The stress-test concern about missing quantitative validation does not land as a major issue here, given the emphasis on visual and qualitative consistency with observed data.

minor comments (2)

[Methods] The transition matrix for the Markov chain model should be included in the supplementary material or main text to allow full reproducibility.
[Results] The figures comparing synthetic and real logs are effective, but labeling the axes consistently across panels would improve readability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive evaluation of the manuscript and for recommending minor revision. We appreciate the recognition that the hybrid framework combining Markov chain models for facies stacking with autoencoder-based latent space sampling offers a promising approach for generating realistic synthetic well logs while preserving petrophysical correlations and geological patterns.

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper outlines a composite pipeline that combines Markov chain models for facies stacking patterns, autoencoder-based dimensionality reduction to a latent space, and MCMC sampling within that space to generate synthetic well logs while aiming to preserve multivariate petrophysical correlations. No equations, parameter-fitting steps, or self-citations are presented that reduce any claimed prediction or output to an input quantity by construction. The central demonstration relies on application to an external real turbidite dataset with qualitative and rock-physics consistency checks, which are independent of the generation procedure itself. The method is therefore self-contained against external benchmarks rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based solely on the abstract, the central claim rests on standard domain assumptions in geophysics and machine learning with no explicit free parameters, invented entities, or ad-hoc axioms identified.

axioms (2)

domain assumption Markov chain models can represent vertical facies stacking patterns in turbidite reservoirs
Invoked as part of the methodology for generating vertical heterogeneity.
domain assumption Autoencoder latent space allows MCMC sampling that preserves original multivariate correlations
Central to the dimensionality reduction and sampling step described.

pith-pipeline@v0.9.0 · 5404 in / 1363 out tokens · 73375 ms · 2026-05-08T03:11:08.483761+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

10 extracted references

[1]

Mukerji, and G

Avseth, P., T. Mukerji, and G. Mavko, 2005, Statistical rock physics: Combining rock physics, information theory, and statistics to reduce uncertainty: Cambridge University Press

2005
[2]

Koenigstein, and R

Bank, D., N. Koenigstein, and R. Giryes, 2021, Autoencoders: Machine Learning for Data Science Handbook, 353–374

2021
[3]

M., 2006, Pattern recognition and machine learning: Springer

Bishop, C. M., 2006, Pattern recognition and machine learning: Springer. (Chapter 13: Sequential Data)

2006
[4]

Busch, J. M., W. G. Fortney, and L. N. Berry, 1987, Electrofacies models: A method for integrating log and core data: The Log Analyst, 28, 255–265

1987
[5]

F., and G

Carle, S. F., and G. E. Fogg, 1996, Transition probability -based indicator geostatistics: Mathematical Geology, 28, 453–476

1996
[6]

Castagna, J. P., M. L. Batzle, and R. L. Eastwood, 1985, Relationships between compressional-wave and shear-wave velocities in clastic silicate rocks: Geophysics, 50, 571–581

1985
[7]

H., and H

Doveton, J. H., and H. R. Cable, 1986, The use of markov chain analysis in well -log correlation: Computers & Geosciences, 12, 819–829

1986
[8]

L., and J

Greenberg, M. L., and J. P. Castagna, 1992, Shear-wave velocity estimation in porous rocks: Theoretical formulation, preliminary verification and applications: Geophysical Prospecting, 40, 195–209

1992
[9]

E., and R

Hinton, G. E., and R. R. Salakhutdinov, 2006, Reducing the dimensionality of data with neural networks: Science, 313, 504–507

2006
[10]

Mukerji, and J

Mavko, G., T. Mukerji, and J. Dvorkin, 2009, The rock physics handbook: Tools for seismic analysis of porous media, 2nd ed.: Cambridge University Press

2009