Speaking to the Earth: AI-Powered Automation for Seismic Modeling

Author: Denis Avetisyan


A new workflow leverages artificial intelligence to simplify and accelerate complex simulations of Earth’s interior.

A global simulation of a magnitude 9.1 earthquake in the Tohoku region utilized a high-resolution, cubed-sphere mesh-detailed through twelve globally distributed seismic stations-to model wave propagation across the Earth, demonstrating a comprehensive approach to seismic event analysis and forecasting.
A global simulation of a magnitude 9.1 earthquake in the Tohoku region utilized a high-resolution, cubed-sphere mesh-detailed through twelve globally distributed seismic stations-to model wave propagation across the Earth, demonstrating a comprehensive approach to seismic event analysis and forecasting.

This review details an integration of large language model agents with the SPECFEM seismic modeling suite using the Model Context Protocol.

Despite advances in computational geophysics, seismic wave simulations remain hampered by complex workflows and a steep learning curve. This paper, ‘Seismology modeling agent: A smart assistant for geophysical researchers’, introduces a novel framework leveraging Large Language Models (LLMs) to automate and simplify the SPECFEM seismic modeling suite through a Model Context Protocol (MCP). By decomposing simulations into agent-executable tools, this work enables intuitive, conversational control and facilitates both fully automated and interactive research experiences. Could this approach herald a new era of AI-assisted scientific discovery in computational geophysics, lowering barriers to entry and enhancing reproducibility?


The Illusion of Precision: Modeling Earth’s Chaos

Seismic waves, generated by earthquakes and controlled sources, act as a natural probe of Earth’s hidden interior, revealing the composition, temperature, and dynamics of layers ranging from the crust to the core. However, accurately simulating how these waves travel through our planet presents a significant computational hurdle. The extreme heterogeneity of Earth’s materials – varying densities, compositions, and phase transitions – demands models of incredible detail. Each wave cycle requires resolving features on scales from kilometers to mere meters, necessitating simulations with billions of grid points. This computational intensity arises because the governing equations – typically the elastic wave equation – must be solved at each point in space and time, quickly escalating the processing demands. Consequently, achieving high-resolution, global simulations remains a substantial challenge, limiting the ability to fully interpret seismic data and construct a comprehensive understanding of Earth’s deep structure.

Current methodologies for simulating seismic wave propagation across the entire globe face significant hurdles due to Earth’s inherent complexity and vast scale. Existing computational approaches, while effective for localized studies, often struggle to accurately represent the planet’s heterogeneous structure – varying material properties, intricate fault lines, and three-dimensional velocity variations – without becoming prohibitively expensive in terms of processing time and resources. This limitation directly impacts the interpretation of seismic data; subtle signals that could reveal details about Earth’s core, mantle, or deep crustal processes may be obscured or misinterpreted due to simplifications made in the simulation. Consequently, refining these simulations is not merely a technological pursuit, but a necessary step toward unlocking a more complete understanding of the planet’s dynamic interior and the geological processes shaping its surface.

The pursuit of increasingly realistic and computationally feasible Earth models is fueling innovation in numerical methods and computational strategies. Researchers are actively developing algorithms that balance the need for high-fidelity simulations – capturing the intricate details of seismic wave behavior – with the practical limitations of available computing power. This drive has led to explorations of techniques like adaptive mesh refinement, which concentrates computational resources in areas of high complexity, and the utilization of massively parallel computing architectures to distribute the workload across numerous processors. Furthermore, investigations into novel wave propagation algorithms, such as spectral element methods and finite-difference schemes optimized for heterogeneous media, are continuously refining the ability to accurately simulate Earth’s response to seismic events. These advancements aren’t merely about faster processing; they represent a fundamental shift towards more insightful interpretations of seismic data and a deeper understanding of our planet’s hidden interior.

Simulations of the s362ani model (Case 5) demonstrate the global propagation of seismic surface waves over time.
Simulations of the s362ani model (Case 5) demonstrate the global propagation of seismic surface waves over time.

SPECFEM: A Toolkit for the Inevitably Imperfect

The SPECFEM suite comprises three primary simulation tools designed for seismic wave propagation modeling: SPECFEM2D, SPECFEM3D_Cartesian, and SPECFEM3D_Globe. SPECFEM2D is optimized for two-dimensional simulations, offering computational efficiency for simplified scenarios. SPECFEM3D_Cartesian facilitates three-dimensional simulations on Cartesian grids, suitable for localized, high-resolution studies. SPECFEM3D_Globe enables full-wave simulations on a spherical Earth model, accommodating global and regional seismology research. These tools collectively support a range of geometries, from simple layered models to complex, realistic Earth structures, and are designed to handle various source types and boundary conditions relevant to earthquake and seismic exploration studies.

The Spectral-Element Method (SEM) employed within the SPECFEM suite is a finite element technique that combines the accuracy of spectral methods with the geometric flexibility of finite elements. SEM represents the solution using high-order polynomials within each element, allowing for efficient and accurate discretization of the wave equation, typically expressed as $ \nabla^2 u – \frac{1}{c^2} \frac{\partial^2 u}{\partial t^2} = 0 $, where $u$ represents the wavefield and $c$ the wave speed. This approach minimizes numerical dispersion, a common issue in finite difference methods, and enables the use of relatively coarse meshes to achieve a desired level of accuracy. The method’s efficiency stems from the use of Gauss-Lobatto-Legendre (GLL) quadrature for numerical integration, reducing computational cost without significant loss of precision.

The SPECFEM suite is architected with a modular design, enabling researchers to customize simulations based on both scientific objectives and available computational resources. This modularity extends to multiple aspects of the simulation process, including the ability to select different element types, mesh resolutions, source mechanisms, and boundary conditions. Furthermore, individual modules can be independently updated or replaced without requiring modifications to the entire codebase. This flexibility allows for efficient use of high-performance computing (HPC) systems by scaling specific computationally intensive modules while maintaining lower resource demands for others, and supports research requiring varying levels of simulation complexity and fidelity.

This wavefield simulation of a salt dome accurately models fluid-solid coupling between water and the elastic subsurface using a user-defined mesh, as evidenced by receiver networks at the sea surface (AA), seabed (BB), and within the subsurface (CC).
This wavefield simulation of a salt dome accurately models fluid-solid coupling between water and the elastic subsurface using a user-defined mesh, as evidenced by receiver networks at the sea surface (AA), seabed (BB), and within the subsurface (CC).

Extracting Every Cycle: Accelerating the Inevitable

Modern computational seismology relies on high-performance computing to manage the demands of large-scale simulations. The SPECFEM software package leverages Graphics Processing Units (GPUs) to accelerate computationally intensive tasks, notably the solution of finite difference equations governing wave propagation. This GPU acceleration results in a substantial reduction in runtime compared to traditional CPU-based simulations; benchmarks demonstrate speedups ranging from 20x to 100x for comparable problem sizes. The efficiency gains are achieved through parallel processing, where the numerous cores within a GPU are utilized to simultaneously compute wavefields across different spatial locations within the simulation domain. This allows for the modeling of more complex Earth structures and the exploration of a wider range of seismic scenarios within a reasonable timeframe.

Convolutional Perfectly Matched Layers (C-PML) are implemented as an absorbing boundary condition within the simulation domain to minimize spurious reflections from the computational grid edges. Traditional absorbing boundaries often introduce reflections due to numerical dispersion; C-PML addresses this by employing a convolutional approach to the implementation of the absorbing layer. This technique effectively stretches the computational mesh near the boundaries, gradually damping outgoing wave energy and simulating an infinitely extended medium. The convolutional implementation improves computational efficiency and stability compared to conventional finite-difference PML implementations, leading to more accurate wave propagation modeling by reducing artificial reflections that would otherwise contaminate the simulation results.

Accurate seismic wave propagation modeling requires a realistic representation of Earth’s internal structure. The s362ani model is a globally averaged, three-dimensional Earth model that provides velocity and density variations as a function of depth and lateral position. This model incorporates data from numerous seismic observations and is discretized into layers with varying material properties. Utilizing such a detailed model, as opposed to simplified, layered structures, allows for more accurate predictions of wave travel times, amplitudes, and phases. The s362ani model specifically provides $P$-wave and $S$-wave velocity, density, and anisotropy information, enabling simulations that better reflect actual seismic wave behavior within the Earth.

The system integrates a user interface and agent logic (theclineplugin) with SPECFEM MCP servers and core executables to facilitate interaction and computation.
The system integrates a user interface and agent logic (theclineplugin) with SPECFEM MCP servers and core executables to facilitate interaction and computation.

The Illusion of Control: Towards Autonomous Seismic Research

Predicting how seismic waves propagate through the Earth is fundamental to understanding earthquakes and the planet’s internal structure. Researchers leverage the SPECFEM suite, a powerful collection of open-source simulation tools, in conjunction with forward modeling techniques to accomplish this. By defining specific Earth models – detailing variations in density, composition, and physical state – and characterizing earthquake sources with parameters like location, magnitude, and rupture mechanism, SPECFEM can computationally simulate the resulting wavefield. This allows scientists to predict ground motion at various locations, offering insights into potential hazards and facilitating the development of more accurate seismic monitoring systems. The fidelity of these predictions relies heavily on the accuracy of the input Earth models and source parameters, prompting ongoing efforts to refine these crucial components of the simulation process.

Automated exploration of complex seismic scenarios is now possible through the integration of Large Language Model (LLM)-based Agents with established geophysical modeling suites. These agents leverage the Model Context Protocol (MCP) to interact directly with software like SPECFEM, effectively translating high-level instructions into actionable simulation parameters. This capability allows for automated parameter sweeps – systematically varying inputs like earthquake location, magnitude, or Earth material properties – and subsequent analysis of the resulting seismic wavefields. Rather than requiring manual intervention for each simulation run, the LLM agent can independently design experiments, execute simulations, and interpret outputs, dramatically accelerating the pace of seismic research and enabling investigations of previously intractable problems. The system essentially creates a closed-loop process where the agent learns from each simulation, refining its approach and maximizing the information gained from each computational cycle.

Recent advancements detail the successful coupling of Large Language Models with the SPECFEM seismic modeling suite, facilitated by Model Context Protocol servers. This integration allows for an unprecedented level of automation in complex geophysical simulations, extending from localized two-dimensional models to comprehensive, global-scale analyses. By utilizing natural language as an interface, researchers can now direct and interpret simulations without the need for specialized coding or intricate scripting. The system effectively translates human instructions into actionable parameters within SPECFEM, enabling automated parameter sweeps, scenario testing, and data analysis – ultimately accelerating the pace of seismic research and broadening access to sophisticated modeling capabilities.

The agent utilizes the Model Context Protocol to dynamically discover and invoke external servers in response to high-level natural language instructions, enabling access to compute resources, databases, and file systems via a JSON-RPC client-server interface.
The agent utilizes the Model Context Protocol to dynamically discover and invoke external servers in response to high-level natural language instructions, enabling access to compute resources, databases, and file systems via a JSON-RPC client-server interface.

The pursuit of automating seismic modeling, as detailed in this work, feels predictably ambitious. It’s a classic case of applying sophisticated tools to problems that, while complex, will inevitably reveal new, unforeseen complications. Andrey Kolmogorov observed, “The most important things are often the ones that can’t be measured.” This rings true; the elegance of the Model Context Protocol (MCP) and the LLM agent integration is likely to be challenged by the messy reality of production-level simulations. The system may successfully navigate predefined scenarios, but the moment researchers attempt genuinely novel modeling – pushing beyond existing datasets and assumptions – the limitations will become apparent. It’s a beautifully engineered solution, destined to become, like so many before it, a stepping stone towards the next iteration of complexity.

The Inevitable Friction

The integration of Large Language Models into SPECFEM, as demonstrated, is less a breakthrough and more a deferral of inevitable entropy. The current architecture, reliant on the Model Context Protocol, neatly sidesteps the messy reality that geophysical models are, by their nature, incomplete representations of chaotic systems. Any system described with natural language is, by definition, open to interpretation – and therefore, to error. The elegance of natural language control will be tested when production workflows inevitably expose edge cases unanticipated by the training data. Anything ‘self-healing’ simply hasn’t broken yet.

Future work will not focus on improving the agent’s ‘understanding’ of geophysics – that’s a category error. The true challenge lies in robust error detection and, crucially, a system for documenting why a simulation failed – a task which is, let’s be honest, collective self-delusion. If a bug is reproducible, the system is stable; the goal isn’t to eliminate errors, but to contain them.

The field will likely shift toward a focus on verifiable simulation provenance. Not ‘explainable AI’, but auditable simulation logs. The question isn’t what the agent thinks it’s doing, but what it did, and why it did it, traceable back to the initial model parameters and data inputs. Expect a proliferation of tools for forensic analysis of simulation failures, because the real work starts when the pretty pictures stop appearing.


Original article: https://arxiv.org/pdf/2512.14429.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-17 13:39