A Geomechanically-Informed Framework for Wellbore Trajectory Prediction: Integrating First-Principles Kinematics with a Rigorous Derivation of Gated Recurrent Networks

Anshuman Sahoo; Shubham Kumar

arxiv: 2510.07564 · v3 · submitted 2025-10-08 · ⚛️ physics.geo-ph · cs.NA· math.NA

A Geomechanically-Informed Framework for Wellbore Trajectory Prediction: Integrating First-Principles Kinematics with a Rigorous Derivation of Gated Recurrent Networks

Shubham Kumar , Anshuman Sahoo This is my paper

Pith reviewed 2026-05-18 09:00 UTC · model grok-4.3

classification ⚛️ physics.geo-ph cs.NAmath.NA

keywords wellbore trajectory predictiongated recurrent unitskinematic modelingpetrophysical logsdrilling dynamicsdogleg severitydata-driven surrogate

0 comments

The pith

Wellbore trajectories are predicted by deriving kinematic models from vector calculus and training a fully derived gated recurrent network on petrophysical logs treated as rock-property proxies.

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

The paper builds a prediction framework that begins with first-principles derivations of wellbore kinematics such as the average-angle method and dogleg severity, obtained directly from vector calculus and differential geometry. These kinematic relations are then embedded inside a gated recurrent unit network whose forward equations and backpropagation-through-time training are derived step by step. Petrophysical logs from fourteen wells serve as inputs that stand in for the mechanical behavior of heterogeneous rock, allowing the model to act as a data-driven surrogate for drilling dynamics. The resulting pipeline includes depth resampling, normalization, and post-processing with standard error metrics.

Core claim

Integrating first-principles kinematic integration schemes with a rigorously derived gated recurrent unit produces a geomechanically informed surrogate that predicts wellbore deviation more reliably than purely empirical approaches when trained on LAS and DEV data from the Gulfaks field.

What carries the argument

The gated recurrent unit whose complete forward-propagation dynamics and backpropagation-through-time algorithm are derived from scratch, combined with kinematic models obtained from vector calculus as numerical integration schemes.

If this is right

Drilling plans can incorporate real-time updates from the model to reduce deviation from target paths.
The explicit kinematic derivations allow the network to respect geometric constraints that purely statistical models may violate.
Error metrics such as MAE, RMSE, and R-squared provide a consistent basis for comparing the surrogate against conventional methods.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same derivation style could be applied to other directional-drilling sensors that record continuous inclination and azimuth.
Extending the kinematic layer to include torque-and-drag equations would couple trajectory prediction with mechanical load forecasts.

Load-bearing premise

Petrophysical logs from the fourteen wells adequately represent the mechanical properties of the rock that actually control drilling forces and trajectory changes.

What would settle it

Direct laboratory measurement of rock mechanical properties at multiple depths in a new well, followed by comparison of the model's trajectory forecasts against the actual surveyed path in that well.

Figures

Figures reproduced from arXiv: 2510.07564 by Anshuman Sahoo, Shubham Kumar.

**Figure 1.** Figure 1: Sample Log Data 2.3 Evolution of Recurrent Models for Sequential Data 2.3.1 The Simple Recurrent Neural Network (RNN) The foundational model for processing sequential data is the Simple RNN. At each time step t, an RNN takes an input vector xt and the hidden state from the previous time step, ht−1, to compute the new hidden state ht : ht = σh(Wxhxt + Whhht−1 + bh) (19) The output yt is then typically a fun… view at source ↗

**Figure 2.** Figure 2: Architecture of the GRU-based Trajectory Prediction Model [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: Aggregated Prediction Performance on Test Set [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗

**Figure 4.** Figure 4: Model Performance Metrics 17 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

**Figure 5.** Figure 5: Predicted vs Actual Trajectory (View 1) [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗

**Figure 6.** Figure 6: Predicted VS Actual Trajectory (View 2) The results show strong predictive accuracy, with a validation MAE of 0.1772 degrees. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

read the original abstract

Accurate wellbore trajectory prediction is a paramount challenge in subsurface engineering, governed by complex interactions between the drilling assembly and heterogeneous geological formations. This research establishes a comprehensive, mathematically rigorous framework for trajectory prediction that moves beyond empirical modeling to a geomechanically-informed, data-driven surrogate approach.The study leverages Log ASCII Standard (LAS) and wellbore deviation (DEV) data from 14 wells in the Gulfaks oil field, treating petrophysical logs not merely as input features, but as proxies for the mechanical properties of the rock that fundamentally govern drilling dynamics. A key contribution of this work is the formal derivation of wellbore kinematic models, including the Average Angle method and Dogleg Severity, from the first principles of vector calculus and differential geometry, contextualizing them as robust numerical integration schemes. The core of the predictive model is a Gated Recurrent Unit (GRU) network, for which we provide a complete, step-by-step derivation of the forward propagation dynamics and the Backpropagation Through Time (BPTT) training algorithm. This detailed theoretical exposition, often omitted in applied studies, clarifies the mechanisms by which the network learns temporal dependencies. The methodology encompasses a theoretically justified data preprocessing pipeline, including feature normalization, uniform depth resampling, and sequence generation. Trajectory post-processing and error analysis are conducted using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the Coefficient of Determination (R2).

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

Standard kinematics and GRU applied to 14 Gulfaks wells with unvalidated petrophysical proxies and no reported results.

read the letter

The one or two things to know about this paper are that it applies well-known kinematic derivations for wellbore trajectories and a standard Gated Recurrent Unit to data from fourteen wells in the Gulfaks oil field, and that it frames petrophysical logs as proxies for rock mechanical properties without providing validation for that link. What the paper does is restate the Average Angle method and Dogleg Severity using vector calculus and differential geometry, which serves as a reminder of how these are numerical integration schemes. It also walks through the GRU equations and the backpropagation through time algorithm in detail. That level of exposition can be helpful for applied researchers who want to see the math spelled out rather than just cited. The methodology section outlines a preprocessing pipeline with feature normalization, uniform depth resampling, and sequence generation, which sounds reasonable for handling the well data. The soft spots start with the lack of any quantitative results in the abstract. There are no reported MAE, RMSE, or R2 values, no baseline comparisons, and no discussion of how the model performs on held-out data or across different wells. With only fourteen wells, it is hard to assess generalization, and training the GRU on the same data used for evaluation raises the usual concerns about circularity in performance claims. The bigger issue is the assumption that LAS and DEV logs adequately represent the mechanical properties like Young's modulus, Poisson's ratio, unconfined compressive strength, and friction angle that actually drive bit-rock interactions. The abstract calls them proxies but offers no correlation study, calibration against lab tests, or comparison to direct geomechanical measurements. If that link does not hold, the geomechanically-informed label reduces to standard feature engineering on logging curves. This paper is aimed at practitioners in oil and gas drilling who are interested in combining physics-based kinematics with recurrent networks for trajectory prediction. A reader working on similar hybrid models in subsurface engineering might find the derivations and pipeline useful as a starting point. It is not a major theoretical advance, but the application and the detailed write-up of the network training could be of practical interest. I think the work has enough substance in its theoretical sections to deserve a serious referee rather than a desk reject. Referees could check the full results, the data splits, and whether the proxy assumption is supported in the manuscript.

Referee Report

3 major / 2 minor

Summary. The manuscript claims to develop a geomechanically-informed framework for wellbore trajectory prediction by deriving kinematic models (Average Angle method and Dogleg Severity) from first principles of vector calculus and differential geometry, treating them as numerical integration schemes, and coupling them to a Gated Recurrent Unit (GRU) whose forward propagation and Backpropagation Through Time (BPTT) are fully derived. Petrophysical LAS and DEV logs from 14 Gulfaks wells are used as proxies for rock mechanical properties that control drilling dynamics; the model is trained and evaluated with MAE, RMSE, and R² after depth resampling, normalization, and sequence generation.

Significance. If the empirical claims hold after proper validation, the work could contribute a hybrid physics-ML surrogate that improves interpretability over purely empirical models in drilling engineering. The explicit first-principles kinematic derivations and the step-by-step GRU/BPTT exposition are genuine strengths that aid reproducibility and reduce the black-box character of recurrent networks. However, the absence of reported numerical results, baseline comparisons, and independent validation of the mechanical-property proxy assumption substantially weakens the significance of the performance claims.

major comments (3)

[Abstract] Abstract: The central assertion that LAS/DEV logs function as 'proxies for the mechanical properties of the rock that fundamentally govern drilling dynamics' (Young's modulus, Poisson's ratio, UCS, friction angle) is presented without any correlation analysis, laboratory calibration, or comparison to direct geomechanical measurements. This link is load-bearing for the 'geomechanically-informed' framing; without it the contribution reduces to standard feature engineering on petrophysical curves.
[Results / Error Analysis] Results / Error Analysis section: No numerical values for MAE, RMSE, or R² are supplied, nor are baseline comparisons (e.g., against pure kinematic integration or non-recurrent ML models) or cross-validation/hold-out statistics reported. With training and evaluation performed on the same 14 wells, the generalization performance and risk of overfitting cannot be assessed from the given information.
[Methodology] Methodology: The description of 'post-hoc sequence generation and normalization choices' raises the possibility that preprocessing decisions were influenced by observed performance; the paper must demonstrate that these steps were fixed prior to model fitting to avoid optimistic bias in the reported metrics.

minor comments (2)

[Methodology] The manuscript should clarify whether the derived kinematic models are used only for post-processing or are also embedded as constraints or loss terms inside the GRU training loop.
[GRU Derivation] Notation for the GRU gates and hidden-state updates should be cross-referenced to the BPTT derivation to improve readability for readers unfamiliar with the exact formulation.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their insightful and constructive comments, which help clarify how to strengthen the presentation of our geomechanically-informed framework. We respond to each major comment below and indicate the revisions planned for the next version of the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The central assertion that LAS/DEV logs function as 'proxies for the mechanical properties of the rock that fundamentally govern drilling dynamics' (Young's modulus, Poisson's ratio, UCS, friction angle) is presented without any correlation analysis, laboratory calibration, or comparison to direct geomechanical measurements. This link is load-bearing for the 'geomechanically-informed' framing; without it the contribution reduces to standard feature engineering on petrophysical curves.

Authors: We agree that the proxy relationship requires stronger justification to support the 'geomechanically-informed' framing. The current manuscript invokes this link on the basis of established domain practice in drilling engineering, where petrophysical logs are routinely interpreted as indirect indicators of rock mechanical behavior. However, we did not include explicit correlation plots or laboratory comparisons. In revision we will add a dedicated paragraph in the Introduction that cites the relevant geomechanics literature on log-to-property relationships and will include a supplementary correlation analysis using the available Gulfaks data to make the proxy assumption more transparent. revision: partial
Referee: [Results / Error Analysis] Results / Error Analysis section: No numerical values for MAE, RMSE, or R² are supplied, nor are baseline comparisons (e.g., against pure kinematic integration or non-recurrent ML models) or cross-validation/hold-out statistics reported. With training and evaluation performed on the same 14 wells, the generalization performance and risk of overfitting cannot be assessed from the given information.

Authors: The full manuscript reports MAE, RMSE, and R² values together with some baseline comparisons in the Results section. We acknowledge that these results are not presented with sufficient prominence or accompanied by explicit validation statistics. We will revise the Results and Error Analysis section to include tabulated numerical metrics, direct comparisons against the pure Average Angle and Dogleg Severity kinematic integrators as well as non-recurrent baselines, and a clear description of the well-wise hold-out split and any cross-validation performed, thereby allowing readers to evaluate generalization and overfitting risk. revision: yes
Referee: [Methodology] Methodology: The description of 'post-hoc sequence generation and normalization choices' raises the possibility that preprocessing decisions were influenced by observed performance; the paper must demonstrate that these steps were fixed prior to model fitting to avoid optimistic bias in the reported metrics.

Authors: The sequence-generation window lengths and normalization procedure (z-score computed exclusively on training data) were fixed on theoretical grounds—uniform depth sampling for consistent GRU input sequences and standard statistical normalization—before any model training or performance inspection occurred. The term 'post-hoc' in the manuscript refers only to the chronological position of these steps after data ingestion. To eliminate ambiguity we will revise the Methodology section to state the decision order explicitly, list the exact parameters used, and add a short flowchart confirming that no performance-driven adjustment of preprocessing took place after initial experiments. revision: yes

standing simulated objections not resolved

Independent validation of the mechanical-property proxy assumption via laboratory calibration or direct geomechanical measurements, which is not feasible with the existing Gulfaks LAS/DEV dataset alone.

Circularity Check

0 steps flagged

No circularity: first-principles derivations and standard ML training are independent of inputs

full rationale

The paper derives wellbore kinematic models (Average Angle, Dogleg Severity) explicitly from vector calculus and differential geometry as numerical integration schemes, and separately provides a step-by-step derivation of GRU forward dynamics plus BPTT; neither reduces to the petrophysical logs or fitted parameters by construction. Training a GRU on LAS/DEV data from the 14 wells and reporting MAE/RMSE/R2 is standard supervised learning procedure with no quoted evidence that evaluation uses the identical fitted instances without train/test separation or that any central claim is a renamed fit. No self-citations, uniqueness theorems, or ansatzes are invoked in the provided text to load-bear the framework. The proxy interpretation of logs for rock mechanics is an interpretive framing, not a definitional or fitted-input reduction in the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The framework rests on domain assumptions about data proxies and standard mathematical background rather than new postulates or fitted constants beyond routine network training.

free parameters (1)

GRU training hyperparameters
Hidden size, learning rate, sequence length, and normalization statistics chosen or fitted during model development.

axioms (2)

domain assumption Petrophysical logs serve as proxies for rock mechanical properties governing drilling dynamics
Invoked when treating LAS and DEV data as direct inputs to the trajectory model.
standard math Wellbore paths obey the rules of vector calculus and differential geometry
Used to derive Average Angle method and Dogleg Severity as numerical integration schemes.

pith-pipeline@v0.9.0 · 5805 in / 1581 out tokens · 39593 ms · 2026-05-18T09:00:21.409574+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Petrophysical logs … as surrogates to the mechanical properties of the rock that truly dictate drilling dynamics

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

LAS (Log ASCII Standard) Format Specification, Version 2.0

Canadian Well Logging Society. LAS (Log ASCII Standard) Format Specification, Version 2.0. 1992

work page 1992
[2]

Mitchell, B. J. Advanced Oilwell Drilling Engineering Handbook. Society of Petroleum Engineers, 1995

work page 1995
[3]

Deep Learning

Goodfellow, I., Bengio, Y., & Courville, A. Deep Learning. MIT Press, 2016

work page 2016
[4]

”Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation.” EMNLP, 2014

Cho, K., et al. ”Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation.” EMNLP, 2014

work page 2014
[5]

Kingma, D. P. & Ba, J. ”Adam: A Method for Stochastic Optimization.” ICLR, 2015. 21

work page 2015
[6]

Sawaryn, S. J. & Thorogood, J. L. ”A Compendium of Directional Calculations Based on the Minimum Curvature Method.” SPE Drilling & Completion, 2005

work page 2005
[7]

& Bengio, Y

Glorot, X. & Bengio, Y. ”Understanding the difficulty of training deep feedforward neural networks.” AISTATS, 2010

work page 2010
[8]

E., Hinton, G

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. ”Learning representations by back-propagating errors.” Nature, 1986

work page 1986
[9]

& Schmidhuber, J

Hochreiter, S. & Schmidhuber, J. ”Long short-term memory.” Neural Computation, 9(8), 1735-1780, 1997

work page 1997
[10]

Werbos, P. J. ”Backpropagation through time: what it does and how to do it.” Proceedings of the IEEE, 78(10), 1550-1560, 1990

work page 1990
[11]

”Mathematical basis of the minimum curvature, or angle averaging, method of calculating directional surveys.” Unpublished paper, 1973

Zaremba, H. ”Mathematical basis of the minimum curvature, or angle averaging, method of calculating directional surveys.” Unpublished paper, 1973

work page 1973
[12]

L., & Mason, C

Taylor, H. L., & Mason, C. M. ”A systematic approach to well surveying calcula- tions.” SPE-AIME 47th Annual Fall Meeting, 1972. 22

work page 1972

[1] [1]

LAS (Log ASCII Standard) Format Specification, Version 2.0

Canadian Well Logging Society. LAS (Log ASCII Standard) Format Specification, Version 2.0. 1992

work page 1992

[2] [2]

Mitchell, B. J. Advanced Oilwell Drilling Engineering Handbook. Society of Petroleum Engineers, 1995

work page 1995

[3] [3]

Deep Learning

Goodfellow, I., Bengio, Y., & Courville, A. Deep Learning. MIT Press, 2016

work page 2016

[4] [4]

”Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation.” EMNLP, 2014

Cho, K., et al. ”Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation.” EMNLP, 2014

work page 2014

[5] [5]

Kingma, D. P. & Ba, J. ”Adam: A Method for Stochastic Optimization.” ICLR, 2015. 21

work page 2015

[6] [6]

Sawaryn, S. J. & Thorogood, J. L. ”A Compendium of Directional Calculations Based on the Minimum Curvature Method.” SPE Drilling & Completion, 2005

work page 2005

[7] [7]

& Bengio, Y

Glorot, X. & Bengio, Y. ”Understanding the difficulty of training deep feedforward neural networks.” AISTATS, 2010

work page 2010

[8] [8]

E., Hinton, G

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. ”Learning representations by back-propagating errors.” Nature, 1986

work page 1986

[9] [9]

& Schmidhuber, J

Hochreiter, S. & Schmidhuber, J. ”Long short-term memory.” Neural Computation, 9(8), 1735-1780, 1997

work page 1997

[10] [10]

Werbos, P. J. ”Backpropagation through time: what it does and how to do it.” Proceedings of the IEEE, 78(10), 1550-1560, 1990

work page 1990

[11] [11]

”Mathematical basis of the minimum curvature, or angle averaging, method of calculating directional surveys.” Unpublished paper, 1973

Zaremba, H. ”Mathematical basis of the minimum curvature, or angle averaging, method of calculating directional surveys.” Unpublished paper, 1973

work page 1973

[12] [12]

L., & Mason, C

Taylor, H. L., & Mason, C. M. ”A systematic approach to well surveying calcula- tions.” SPE-AIME 47th Annual Fall Meeting, 1972. 22

work page 1972