Opening — Why this matters now
Scientific software has a strange tradition: world‑class physics wrapped in workflows that feel frozen in the 1990s. Seismology is no exception. SPECFEM — arguably the gold standard for seismic wave simulation — delivers extraordinary numerical fidelity, but only after users survive a rite of passage involving fragile text files, shell scripts, and MPI incantations.
At the same time, Large Language Models have quietly evolved from chatty copilots into agentic systems — capable of planning, tool use, and multi‑step execution. The obvious question is no longer whether AI can help scientists, but where exactly it should sit in the workflow.
This paper answers that question with refreshing precision: not by rewriting physics, but by wrapping legacy solvers with an agent‑friendly interface. The result is a conversational, intent‑driven seismology pipeline — and a blueprint for modernizing scientific software without breaking it.
Background — From spectral elements to semantic intent
SPECFEM’s dominance comes from its use of the spectral‑element method (SEM): high‑order accuracy, geometric flexibility, and excellent parallel scalability across 2D, 3D regional, and full‑globe simulations. The math is elegant; the workflow is not.
Traditionally, a single experiment requires:
- Manual editing of dozens of interdependent configuration files
- Sequential execution of meshing, decomposition, database generation, and solvers
- Careful MPI and GPU orchestration
- Post‑hoc visualization using external scripts
Graphical interfaces help beginners click faster — but they don’t understand intent. Saying “simulate wave focusing from a buried ridge” still means translating that idea into hundreds of parameters.
Enter LLM‑based agents — systems that can reason, plan, and act. What they lacked, until now, was a standardized way to touch legacy scientific code.
That missing layer is the Model Context Protocol (MCP).
Analysis — What the paper actually builds
The authors implement the first MCP server suite for SPECFEM, covering:
- SPECFEM2D
- SPECFEM3D Cartesian
- SPECFEM3D Globe
Rather than modifying SPECFEM itself, they add a service layer that exposes each step of the simulation lifecycle as agent‑callable tools.
Core architectural idea
| Layer | Role |
|---|---|
| LLM Agent | Interprets scientific intent, plans steps |
| MCP Client | Discovers and calls tools |
| MCP Servers | Translate intent into files + commands |
| SPECFEM Core | Executes physics‑based simulation |
The workflow shifts from file‑driven to intent‑driven:
“Simulate a Mw 9.1 Tohoku earthquake using a global anisotropic Earth model.”
becomes a structured chain of tool calls — automatically generating configuration files, running solvers, and returning visualizations.
Crucially, the system supports human‑in‑the‑loop control. Scientists can intervene, correct assumptions, or steer modeling decisions mid‑execution — retaining authority while offloading drudgery.
Findings — Five case studies, one pattern
Across five increasingly complex case studies — from teaching‑level 2D lensing experiments to GPU‑accelerated global Earth simulations — the results are consistent:
| Dimension | Traditional SPECFEM | Agent‑Driven SPECFEM |
|---|---|---|
| Setup effort | High | Minimal |
| Error recovery | Manual | Agent‑assisted |
| Reproducibility | Fragile | Structured |
| Accessibility | Expert‑only | Broad |
| Scientific control | High | Still high |
Notably, the agent successfully:
- Constructs comparative models with controlled parameter differences
- Integrates user‑provided external meshes
- Configures realistic 3D volcanic simulations
- Executes full‑physics global wave propagation on GPUs
The physics doesn’t change. The interface to physics does.
Implications — This is bigger than seismology
Strip away the geological specifics and a general pattern emerges:
- Legacy solvers are not obsolete — their interfaces are
- AI value comes from orchestration, not numerical replacement
- MCP‑style tool layers may become the lingua franca of scientific software
For businesses and research institutions, the lesson is clear:
ROI from AI doesn’t require reinventing core algorithms. It requires lowering the cognitive and operational cost of using them.
The same architecture could modernize CFD solvers, climate models, molecular dynamics engines, or financial risk simulators — anywhere expertise is locked behind brittle workflows.
Conclusion — Toward autonomous laboratories (carefully)
This paper doesn’t promise self‑discovering AI scientists — and that restraint is its strength. Instead, it delivers something far more actionable: a practical bridge between human intent and industrial‑grade computation.
Agentic AI here is not a replacement for expertise, but a force multiplier — one that respects scientific judgment while erasing unnecessary friction.
If the future of research is autonomous, it will be built one MCP server at a time.
Cognaptus: Automate the Present, Incubate the Future.