Mountainous first arrivals go to a multimodal agent: FA-Agent reads the gather and deletes picks it cannot judge
Synopsis
First arrivals in mountainous seismic data are often weak, intermittent, and masked by strong direct waves and noise, so energy-based pickers jump to later arrivals or noise, and deep-learning pickers often need retraining for a new survey. FA-Agent combines a multimodal large language model, deterministic processing tools, and skill files that carry processing experience into an extensible framework: the agent surveys the whole gather to mark doubtful zones, then examines them enlarged and decides per segment whether to keep, correct, or remove each pick. On semi-synthetic gathers with known first arrivals and real mountainous noise, it picked clearly visible first arrivals accurately and removed picks where the first arrival was invisible instead of picking later events.
Interpretation
First-arrival picking is split in two: tools produce candidate picks, and a multimodal large language model reads the gather image like a processor and decides window by window whether a pick should be kept, corrected, or removed. Earlier automatic pickers mainly predict arrival times, and whether a pick in a difficult zone is trustworthy is usually decided later, during quality control and manual review; deep-learning pickers moved to a survey with different acquisition, waveforms, and noise usually need retraining or fine-tuning with labeled data from that survey. On semi-synthetic gathers whose first arrivals are known and whose noise comes from real mountainous gathers, the agent picked accurately where the first arrival was clearly visible and removed picks where the first arrival was invisible and could not be judged, rather than picking later events.
The agent first views the whole gather without the pick curve, identifies the earliest coherent wavefront and its branches, then overlays the initial pick and marks the few windows where the two disagree or the wavefront is uncertain. The initial pick combines an envelope energy ratio with an AIC onset score and a lateral continuity constraint that links picks across the gather; in mountainous data that continuity constraint can carry the curve onto a stronger later event or through zones where no first arrival is visible. Each window is examined enlarged first without and then with the initial pick, with enough neighboring traces to show the local trend; the agent divides the window into segments and decides for each whether to keep, correct with anchors on the visible wavefront, or remove, then reviews the assembled picks on the whole gather and resolves jumps at segment joints.
Processing experience lives in skill files rather than model weights, and gathers an expert has reviewed are written automatically into the skill case library, so the agent accumulates cases with use. Experience is loaded as reusable instructions and reference cases, so expert review criteria and reviewed examples can be supplied and revised without changing model weights. The skills were developed on mountainous field data in consultation with a geophysicist with many years of experience processing mountainous seismic data; the agent processed 1000 randomly sampled gathers, and gathers 0813, 0831, and 0001 presented here lie outside its case library, with 0813 and 0831 in the most difficult class A.
Perspective
The demonstrated scope is semi-synthetic tests with known first arrivals and noise taken from real mountainous gathers, plus three field examples (0813, 0831, 0001) drawn from 1000 randomly sampled gathers of one three-dimensional mountainous survey, two of them in the most difficult class A. The intended readers are engineers and researchers working on land seismic processing and first-arrival tomography; the field data are confidential and cannot be released, and the complete framework code with example test data is planned for release after initial review.
True first arrivals in the field data are unknown, and expert picks of weak or obscured arrivals are themselves subjective, so the objective comparison comes only from the semi-synthetic gathers. Correcting one gather took about 30–90 s on a single core of a local AMD CPU depending on difficulty; a commercial large language model also consumes tokens mainly for image input, adding cost, and sending confidential data to an external service may not be acceptable.