A Generative-AI Modeling Framework for Explainable Decision Support in Complex Geosteering Scenarios

Sergey Alyaev, Kristian Fossum, Hibat Errahmen Djecta, Jan Tveranger, Ahmed H. Elsheikh · 2025

Read A Generative-AI Modeling Framework for Explainable Decision Support in Complex Geosteering Scenarios by Sergey Alyaev, Kristian Fossum, Hibat Errahmen Djecta, Jan Tveranger, Ahmed H. Elsheikh (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-08509.txt.

Introduction

Sergey Alyaev and co-authors present an AI framework that helps drillers steer wells through rock in real time, in a 2025 paper. Geosteering means adjusting a well's path while drilling. Good steering keeps the bit in the best rock layer. The same skill now serves geothermal energy and carbon storage.

Their workflow joins several AI parts into one loop. Generative networks propose plausible geology ahead of the bit. Ensemble methods update those models with live sensor data. A dynamic programming optimizer then picks the best steering moves. Crucially, the system explains its advice in human terms. Drillers gain decision support they can question, not a black box. The design targets fewer costly drilling errors.

Catalog record

CREATORSergey Alyaev, Kristian Fossum, Hibat Errahmen Djecta, Jan Tveranger, Ahmed H. Elsheikh
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-08509

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Generative-AI Modeling Framework for Explainable Decision Support in Complex Geosteering Scenarios Sergey Alyaev, Kristian Fossum, Hibat Errahmen Djecta, Jan Tveranger, Ahmed H. Elsheikh Abstract The real-time process of directional changes while drilling, known as geosteering, is crucial for hydrocarbon extraction and emerging directional drilling applications such as geothermal energy, civil infrastructure, and CO2 storage. The geo-energy industry seeks an automatic geosteering workflow that continually updates subsurface uncertainties and captures the latest geological understanding, informed by real-time observations. We propose a real-time, AI-driven geosteering workflow that integrates Generative Adversarial Networks (GANs) for geological parameterization, ensemble methods for model updating, and global discrete dynamic programming (DDP) optimization for complex decision-making during directional drilling operations. Our framework relies on offline training of a GAN model to reproduce relevant geology realizations and a Forward Neural Network (FNN) to model the response of Logging-While-Drilling (LWD) tools for a given geomodel. This paper introduces a first-of-its-kind workflow that progressively reduces GAN-geomodel uncertainty around and ahead of the drilling bit and adjusts the well plan accordingly. The workflow automatically integrates real-time around-bit LWD, which, through learned geological correlations, reduces uncertainty in predicted geology ahead of drilling. A DDP-based decision support system leverages probabilistic look-ahead predictions to suggest better steering strategies. We test the workflow prototype on a small yet challenging low-net-to-gross drilling scenario with several possible targets. The results show that the workflow produces meaningful steering recommendations and, through its probabilistic updates, automatically maps formation boundaries along the drilled well. 1 Introduction Dynamic real-time adjustment of well trajectory, or geosteering, is established in hydrocarbon drilling operations to maximize efficiency and safety ( Hermanrud et al., 2019 ) . Traditionally, geosteering has depended heavily on the manual interpretation of real-time Logging-While-Drilling (LWD) data, a challenging task under the pressure of rapid drilling operations and complex geological uncertainties ( Cheraghi et al., 2022 ; Rassenfoss, 2022 ) . The application of geosteering extends beyond hydrocarbon wells to geothermal ( Ungemach et al., 2021 ) , civil infrastructure, and emerging domains like CO2 storage ( Miotti et al., 2022 ) , further underscoring its importance. Current methodologies developed for well-planning incorporate Ensemble-based Closed-Loop Reservoir Management (EnCLRM), which has set a new standard for probabilistic estimate of subsurface uncertainties ( Hanea et al., 2016 ; Skjervheim et al., 2012 ; Abadpour et al., 2018 ) and informed decision support ( Hanea et al., 2019 ) . Despite its effectiveness, implementing EnCLRM in real-time scenarios is limited by the computational demands of its complex modeling sequences. Within these sequences, the geomodeling that involves depth conversion, structural modeling, faults, facies modeling on a grid, and property modeling ( Skjervheim et al., 2012 ) is among the hardest to automate and adapt for real-time operations. Recent technological innovations have introduced generative machine learning, including variational auto-encoders ( Canchumuni et al., 2019 ) , Generative Adversarial Networks (GAN) ( Chan and Elsheikh, 2017 ) , and, lately, latent diffusion models Lee et al. (2025) ; Di Federico and Durlofsky (2025) as powerful tools for geomodeling. We will refer to them as GANs despite meaning a larger class of models. These models, capable of rapidly producing realistic geomodel realizations from a reduced set of parameters, provide a promising approach to overcoming traditional limitations. Alyaev et al. (2021c) and Fossum et al. (2022) demonstrated that GANs combined with ensemble methods enable quick updates to geological models in response to acquired drilling data. Thus, GAN-based workflows can potentially transform geosteering into a more quantifiable and precise process. While in field operations, decisions remain the responsibility of a geosteering team ( Hermanrud et al., 2019 ; Rassenfoss, 2022 ) , academic studies have advanced decision optimization methods to enhance efficiency and consistency. Initial approaches, such as greedy optimization, were employed by Chen et al. (2015) and Kullawan et al. (2016) , focusing on immediate gains at each decision stage, but often resulted in sub-optimal outcomes due to the lack of future consideration ( Kullawan et al., 2018 ) . Pavlov et al. (2024) expanded on greedy optimization by integrating real-time formation evaluation data and differential evolution algorithms to optimize drilling trajectories, showing improvements over traditional methods. To address the limitations of greedy optimization, Kullawan et al. (2018) introduced Approximate Dynamic Programming (ADP) for geosteering, which incorporates future decisions and learning for proactive, globally optimized decision-making, superseding the greedy approaches in many scenarios. Despite its effectiveness, ADP’s high computational demands and scenario-specific design pose challenges for real-time applications. Alyaev et al. (2019) proposed a version of ADP, termed ’naive-optimistic’ ( Alyaev et al., 2018 ) , that adapted the method to a general updatable ensemble representation of uncertainty, albeit with a loss of proven dynamic-programming (DP) optimality. Nevertheless, it outperformed most experts in a synthetic experiment ( Alyaev et al., 2021a ) . Outside of geosteering, for sequential well location selection to avoid subsurface caves, Kanfar et al. (2024) explored the application of GANs combined with the Fast Informed Bound (FIB) algorithm, which utilizes elements of ADP to compute upper bounds on the value function. Recently, reinforcement learning (RL) has emerged as a promising alternative to ADP methods, offering flexibility and computational efficiency once trained. RL trains an agent through iterative learning from environmental interactions, thereby developing optimal decision-making strategies suited for real-time scenarios. Kristoffersen et al. (2021) presented the use of the ’NEAT’ evolutionary RL algorithm to incorporate the effects of geosteering during well planning, demonstrating another aspect of decision optimization. Muhammad et al. (2023) demonstrated that Deep Q-Network (DQN) algorithm achieves comparable and sometimes better results than ADP in geosteering optimization. Muhammad et al. (2025b) combined DQN with particle filter, that in several experiments outperformed all the experts in the close-to-realistic ROGII Geosteering World Cup 2021, unconventional stage ( Muhammad et al., 2025a ) . This paper introduces a holistic geosteering approach that integrates an updatable ensemble of GANs within a probabilistic framework for dynamic well placement. The workflow is modular, allowing for the combination of different generative geomodels and various measurements, with diverse data assimilation and decision support algorithms, making it easily adaptable to other geosteering scenarios. Our initial implementation utilizes real-time ultra-deep electromagnetic (UDAR) data and a robust Decision Support System (DSS), combining the Ensemble Kalman filter (EnKF), GAN-based geomodels, and the adapted ’naive-optimistic’ dynamic programming. The workflow enhances the precision of real-time subsurface modeling and decision-making by automating geosteering through a closed loop of sequential decisions and data acquisition. It moves towards an ’autopilot’ capability for complex geosteering challenges. This paper extends a conference paper Alyaev et al. (2024) and reuses parts of the methodological description and the text. The rest of the manuscript is organized as follows: Section 2 details the proposed workflow, including the multi-step forward mapping sequence, the EnKF data-assimilation procedure, and the dynamic-programming-based DSS under uncertainty. Section 3 presents the synthetic reference case: prior construction, objective and constraint formulation, data-assimilation setup, sequential decision recommendations, and a discussion of performance and limitations. Section 4 summarizes the main contributions, outlines observed artifacts and limitations, and highlights avenues for future developments. 2 Method The proposed workflow consists of two parts: the offline (pre-job) phase and the online phase. The online phase relies on selecting a geological prior and is enabled by offline training:

Read the complete text