The Energy Paradox: Artificial Intelligence for Seismic Computational Load Reduction in Oil and Gas — A Technical Survey and Position Analysis.

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Computational geophysics and AI researchers can use the survey of surrogate models and the Energy Paradox framing to position future work.

🏢実務担当者:O&G geoscientists and HPC teams can benchmark their AI adoption against the case studies (Petrobras, Aramco, Shearwater-NVIDIA).

🏛政策担当者:The Energy Paradox highlights regulatory gaps in AI-energy governance; policymakers may need to consider cross-sectoral impacts.

Abstract — The oil and gas (O&G) industry faces a compounding computational challenge: advanced seismic imaging techniques such as Full Waveform Inversion (FWI) and Reverse Time Migration (RTM) impose exponentially scaling workloads — doubling the maximum frequency of a 3D FWI run increases computational demand by a factor of sixteen, while a single high-resolution 3D FWI job on a modern supercomputer may require approximately 25,000 GPU-hours. This paper surveys the principal artificial intelligence (AI) and machine learning (ML) techniques being deployed to reduce these workloads, including deep neural network surrogate models for iterative solvers, physics-informed neural networks (PINNs), edge inference architectures, and alternative dataflow hardware paradigms. A structured review of institutional deployments — spanning the U.S. Department of Energy Genesis Mission (Executive Order, November 2025), the NETL Science-Informed Machine Learning for Accelerating Real-Time Decisions (SMART/SAMI) initiative, Petrobras SolverBR, Saudi Aramco AI-assisted seismic processing, and the Shearwater-NVIDIA collaboration — is presented alongside quantified performance outcomes. Beyond the technical review, this paper introduces and formalizes the Energy Paradox: the O&G sector simultaneously supplies the hydrocarbon energy that powers AI data centers globally and consumes AI-driven high-performance computing to reduce its own operational costs. We argue this closed-loop relationship has strategic, economic, and regulatory implications that have not been adequately addressed in the literature. A framework for evaluating AI adoption in seismic workflows — encompassing computational efficiency, energy footprint, and governance alignment — is proposed as a contribution to practitioners, engineers, and researchers operating at this intersection. Index Terms — Full Waveform Inversion, Reverse Time Migration, Seismic Processing, Deep Learning, Surrogate Models, Physics-Informed Neural Networks, Edge Computing, Oil and Gas, HPC, Energy Paradox, DOE Genesis Mission, NETL SAMI, Computational Geophysics, GPU Efficiency I. INTRODUCTION The oil and gas sector has long been one of the most computationally intensive industries outside of national security and pharmaceutical research. Seismic data acquisition surveys generate petabytes of raw measurements per campaign, and the physics-based algorithms used to convert those measurements into actionable subsurface imagery demand supercomputer-scale resources. Yet the economic pressure on this infrastructure has intensified significantly in the 2024–2026 period — not because of increased exploration activity, but because of a structural shift in global energy demand driven by artificial intelligence itself. The proliferation of large language models (LLMs), generative AI systems, and large-scale GPU clusters has produced an unprecedented surge in data center electricity consumption. AI hardware — primarily high-density GPU and TPU farms — is projected to account for a rapidly growing share of global electricity demand through 2030 [1]. A substantial fraction of this electricity is sourced from natural gas, either directly through gas-fired generation or indirectly through LNG supply chains managed by major O&G operators. The O&G sector therefore finds itself in a structurally paradoxical position: it supplies the energy that enables AI at global scale, while simultaneously seeking to deploy AI to reduce the computational cost of its own most expensive workflows. This paper formalizes this relationship as the Energy Paradox and situates it within a technical survey of the AI-driven computational reduction techniques being adopted at the frontier of seismic data processing. The survey covers four primary AI paradigms — DNN surrogate models, physics-informed neural networks (PINNs), edge inference architectures, and alternative compute hardware — with particular attention to quantified performance outcomes from industrial and governmental deployments. The remainder of this paper is organized as follows. Section II quantifies the computational problem. Section III surveys AI reduction techniques. Section IV reviews institutional deployments and case studies. Section V presents the Energy Paradox framework. Section VI discusses implementation considerations. Section VII addresses challenges and governance. Section VIII outlines future directions, and Section IX concludes. II. THE COMPUTATIONAL PROBLEM: QUANTIFYING SEISMIC WORKLOADS Understanding the magnitude of the computational challenge is a prerequisite for evaluating the impact of AI-driven solutions. Seismic processing encompasses a pipeline of increasingly expensive operations, from pre-processing through migration and inversion, each characterized by distinct scaling properties. A. Full Waveform Inversion (FWI) FWI is the gold standard for high-resolution subsurface velocity model building. It iteratively minimizes the misfit between simulated and observed wavefields by repeatedly solving the wave equation forward and backward through the model volume [2]. The computational cost scales as O(N^4) in 3D, where N is the number of grid points per dimension — a consequence of the Nyquist sampling requirement applied across three spatial dimensions and time. A defining characteristic of FWI workloads is their sensitivity to frequency content. Doubling the maximum frequency of a 3D FWI run increases the computational demand by a factor of sixteen [3]. A single high-resolution 3D FWI job on a state-of-the-art supercomputer requires approximately 25,000 GPU-hours distributed across hundreds of GPUs — an energy equivalent to powering an average U.S. household for fifteen months [3]. Despite massive GPU deployments, many operators report that seismic workloads achieve only 50% of theoretical GPU peak performance due to memory-bandwidth bottlenecks and irregular access patterns [3]. B. Reverse Time Migration (RTM) RTM is the preferred depth migration algorithm for imaging beneath complex geological structures such as salt bodies and thrust belts [4]. Like FWI, RTM involves repeated wave equation solving, but its computational profile is dominated by the cross-correlation imaging condition applied between forward and backward wavefields. Pre-stack depth migration and pre-stack time migration combined represent more than 70% of the total computational load of a seismic data processing pipeline [5]. C. Reservoir Simulation Reservoir flow simulation, while distinct from seismic processing, shares the same HPC infrastructure and is subject to analogous scaling penalties. Reducing the spatial discretization of a 3D grid by a factor of two in each dimension increases the total number of cells — and thus the computational requirements — by a factor of eight [3]. This sensitivity to resolution creates strong economic incentives to develop AI-assisted proxy models capable of delivering accuracy competitive with full-physics simulators at a fraction of the computational cost. The dominant workloads in the seismic processing pipeline are characterized by super-linear or polynomial scaling with respect to resolution parameters. This scaling behavior, combined with the secular trend toward higher-resolution surveys driven by increasingly complex reservoir targets, creates a compounding cost trajectory that GPU scaling alone cannot resolve economically. III. AI AND ML TECHNIQUES FOR SEISMIC COMPUTATIONAL REDUCTION This section surveys the principal AI and ML approaches being applied to reduce the computational burden of seismic workflows. Each approach targets a different stage of the pipeline or a different aspect of the cost structure. A. Deep Neural Network Surrogate Models for FWI The most computationally impactful class of AI techniques involves replacing — partially or fully — the iterative optimization loop of FWI with a learned forward or inverse mapping. In surrogate-based approaches, a deep neural network is trained to approximate the relationship between observed seismic data and subsurface velocity models, bypassing the need to repeatedly solve the wave equation during inference [6]. Three distinct integration strategies have been proposed in the literature [6]: (i)      Learning the optimization algorithm itself — the neural network is trained to mimic the update steps of gradient-based FWI; (ii)    Introducing data-driven components within a classical optimization framework — hybrid architectures that replace computationally expensive sub-steps with learned proxies; (iii)  Full end-to-end inversion — the neural network directly maps seismic observations to velocity models. Processing and inference times for deep learning models trained on seismic inversion tasks have been reported to be an order of magnitude lower than the optimization time of classical FWI approaches [6]. This reduction is particularly pronounced at inference time, where — after a computationally intensive offline training phase — the network produces velocity model predictions in seconds rather than hours. The dimensionality-reduction and representation-learning capabilities of deep neural networks are particularly well-suited to seismic inversion, where the relationship between wavefield observations and subsurface parameters is high-dimensional and non-linear [6]. Convolutional architectures have shown strong performance on velocity model reconstruction tasks, exploiting the spatial coherence of both seismic data and subsurface geology. B. Physics-Informed Neural Networks (PINNs) A limitation of purely data-driven surrogates is their dependence on large, labeled training datasets — a constraint that is challenging in the O&G context due to the commercial sensitivity of subsurface data and the high cost of generating synthetic training data at industrial resolution. Physics-Informed Neural Networks address this limitation by embedding the gover

• openaire https://doi.org/10.5281/zenodo.19652036first seen 2026-06-11 04:45:56 · last seen 2026-06-16 04:32:35