[Paper] Seismology modeling agent: A smart assistant for geophysical researchers

Source: arXiv - 2512.14429v1

Overview

The paper presents Seismology Modeling Agent, a conversational AI assistant that wraps the popular open‑source seismic wave simulator SPECFEM with a suite of Large Language Model (LLM)‑driven tools. By turning a traditionally file‑centric, command‑line workflow into an intent‑driven chat interface, the authors dramatically lower the barrier for geophysical researchers and make large‑scale seismic simulations more reproducible and developer‑friendly.

Key Contributions

• Model Context Protocol (MCP) server suite for SPECFEM – a modular back‑end that exposes every stage of a SPECFEM run (parameter creation, mesh generation, partitioning, solver launch, post‑processing) as independent, LLM‑callable services.
• Intent‑driven conversational workflow – users can describe what they want (e.g., “run a 3‑D globe simulation for a magnitude‑6 earthquake in the Pacific”) and the agent translates that into the exact sequence of SPECFEM commands and file edits.
• Hybrid automation modes – supports fully autonomous execution as well as a “human‑in‑the‑loop” mode where the researcher can intervene, approve, or tweak intermediate steps.
• Cross‑platform support – works with SPECFEM2D, SPECFEM3D Cartesian, and SPECFEM3D Globe, covering the most common use‑cases in computational seismology.
• Open‑source release – complete code, Docker images, and example notebooks are publicly available, enabling rapid adoption and community extensions.

Methodology

1. Decompose SPECFEM workflow – The authors identified 12 atomic operations (e.g., generate source‑time function, create mesh, partition domain) and wrapped each as a microservice exposing a JSON‑based API.
2. Model Context Protocol (MCP) – A lightweight protocol that lets an LLM keep track of the current simulation state, request the appropriate service, and receive structured responses (e.g., file paths, status codes).
3. LLM orchestration layer – A prompt‑engineered LLM (GPT‑4‑style) acts as the “brain” of the assistant. It parses natural‑language user intents, decides which MCP calls are needed, and assembles the resulting command‑line invocations.
4. Human‑in‑the‑loop UI – A simple web chat (React + Flask) displays the LLM’s plan, lets users approve or edit parameters, and streams log output from the underlying SPECFEM processes.
5. Validation – The pipeline was exercised on three benchmark scenarios (2‑D fault rupture, 3‑D Cartesian basin, 3‑D global mantle) and compared against manually‑run baselines to verify numerical fidelity.

Results & Findings

• Speedup: 3–4× faster to obtain a usable simulation, mainly because the assistant eliminates repetitive file editing and job‑script tuning.
• Reproducibility: All steps are logged in a machine‑readable “session transcript,” enabling exact replay of any experiment.
• User satisfaction: In informal developer surveys, participants rated the chat interface as “intuitive” (4.6/5) and reported a steep reduction in onboarding time for new graduate students.

Practical Implications

• For developers: The MCP pattern can be reused for other scientific codes (e.g., CFD, climate models), turning any CLI‑heavy tool into an LLM‑driven service without rewriting the core solver.
• For geophysical firms: Faster prototyping of earthquake scenarios means quicker risk assessments and the ability to run “what‑if” studies on demand.
• For cloud/edge deployment: Because each step is containerized, the entire workflow can be orchestrated on Kubernetes or serverless platforms, opening the door to on‑the‑fly seismic simulations for real‑time monitoring systems.
• Educational impact: New students can focus on physics and interpretation rather than mastering dozens of configuration files, accelerating research cycles in academia and industry labs.

Limitations & Future Work

• LLM dependence – The current prototype relies on a proprietary GPT‑4‑style model; performance may vary with open‑source alternatives.
• Scope of automation – Only the most common SPECFEM options are covered; exotic material models or custom source functions still require manual intervention.
• Scalability testing – Benchmarks were run on modest HPC nodes; future work will evaluate the assistant on large‑scale clusters (thousands of cores) and assess latency of LLM calls.
• Extending MCP – The authors plan to open the MCP specification to the broader scientific software community, enabling plug‑and‑play integration with other geophysical toolchains.

Authors

• Yukun Ren
• Siwei Yu
• Kai Chen
• Jianwei Ma

Paper Information

• arXiv ID: 2512.14429v1
• Categories: cs.AI, cs.SE
• Published: December 16, 2025
• PDF: Download PDF