FA-Agent: An extensible multimodal agent framework for first-arrival picking in mountainous seismic data

The 30-second take

• What: FA-Agent combines a multimodal model for visual recognition and decisions with tools and mountainous-data skills, surveying a gather, zooming, then keeping, correcting, or removing initial picks, including on semi-synthetic data with known arrivals and real mountainous noise.
• Why it matters: First-arrival picks still go to humans in low-signal mountain data because energy pickers grab later events and deep pickers need retraining. If an extensible agent can refuse invisible picks, some of that manual loop becomes a cheaper default.
• Who should care: Seismic processors, near-surface velocity-model teams, and groups that cannot retrain a picker for every survey or cannot send confidential gathers to a cloud model.

What the paper actually did

First-arrival traveltimes underpin near-surface velocity modeling and static corrections, but in low signal-to-noise mountainous land data the first arrival is often weak, intermittent, and masked by strong direct waves and noise. Energy-based pickers then pick later arrivals or noise, and deep-learning pickers usually need retraining for a new survey. Erroneous picks bias tomography, so difficult zones are still inspected by hand.

FA-Agent is an agent framework on a multimodal large language model. It combines a model that does visual recognition and decisions, tools that process the data, and skills that carry experience with mountainous data. The framework is open and extensible: new pickers, denoisers, and quality-control tools can be added, and new experience written into skills. It is not restricted to mountains or to one model.

Emulating a processor, the agent surveys the whole gather, examines problem zones in enlarged views, then keeps, corrects, or removes initial picks. On semi-synthetic gathers with known first arrivals and real mountainous noise, it picked accurately where the arrival was clearly visible and removed picks where it was invisible. On randomly sampled gathers from a mountainous survey, it corrected picks onto weak branches and removed unconfirmed ones, including on the most difficult data. Expert-reviewed cases are added to skills automatically, and a locally deployed model can keep confidential data on site.

What makes this disruptive

If the core claim holds, the useful move is not a better neural picker. It is an agent that can refuse to pick. Deleting invisible first arrivals is as important as snapping to a weak branch, because bad picks poison tomography.

The distinctive design is processor emulation: whole-gather survey, zoom, then keep/correct/remove, plus skills that accumulate expert cases and a local-model path for confidential surveys. Extensibility is the product claim; mountains are the first habitat.

This week's editorial ranking placed the preprint eighteenth of twenty, with a score of 56. That is triage for the Disruptive Concepts brief, not a peer-review verdict. Treat FA-Agent as a claimed processing copilot, not as a replacement for a velocity modeler.

Why it matters (outside the lab)

Abundance lens (today's luxuries toward tomorrow's defaults): accurate monitoring of the subsurface is still scarce where the data are ugly. If agent pickers that know when to abstain scale, some mountain statics work can move toward a cheaper default.

Near-term, the preprint is a baseline for processors already fighting land first arrivals. Medium-term, whether skills transfer across surveys without leaking bad habits, and whether local models match the cloud, decide if this is a default component.

No calendar date is implied. Disruptive Concepts, which publishes these explainers free for humans, reads this as a possible brick toward less manual first-arrival triage, not as a promise of automatic imaging.

Limitations & open questions

This is a preprint, not a production processing system. The abstract reports qualitative success on semi-synthetic and real mountainous gathers. We have not seen pick-error histograms or a named baseline table.

The abstract does not report model identity, gather counts, or residual tomography bias after agent edits. Automatic skill updates from expert cases can encode one shop's taste. "High accuracy where visible" leaves the hard invisible case as deletion, which is correct only if invisibility is judged well.

Abundance is not automatic. An agent picker does not demonetize seismic processing on a fixed date. Survey design and velocity inversion still sit between this tool and a default.