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

Abstract

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 require retraining for a new survey. Because erroneous picks bias first-arrival tomography, many difficult zones are still inspected and corrected manually. We present FA-Agent, an agent framework for first-arrival picking built on a multimodal large language model. It combines a model that performs visual recognition and makes decisions, tools that process the data, and skills that carry experience of processing mountainous data. The framework is open and extensible: new pickers, denoising methods, and quality-control tools can be added, and new experience can be written into the skills. It is restricted neither to mountainous data nor to a particular model. Emulating a seismic data processor, the agent first surveys the whole gather to find zones needing attention, then examines them in enlarged views, and finally keeps, corrects, or removes the initial picks. On semi-synthetic gathers with known first arrivals and real mountainous noise, the agent picked the first arrival with high accuracy where it was clearly visible and removed the picks where it was invisible and could not be judged. Applied to randomly sampled gathers from a mountainous survey, it corrected picks onto weak first-arrival branches and removed picks where no first arrival could be confirmed, and it performed well on the most difficult data. Cases reviewed by experts are added to the skills automatically, so the agent improves with use, and a locally deployed model keeps confidential data on site. FA-Agent can thus become an essential component of future automated seismic data processing workflows.