SeisEvo: LLM agents evolve seismic algorithms, SNR up 7dB

Seismic data reconstruction has long relied on hand-crafted structural priors and iterative operators—a design space too large for manual trial-and-error. Deep learning methods offer better performance but hide their logic in learned weights, making them hard to inspect or modify. SeisEvo, from Yingjie Xu, Siwei Yu, and Jianwei Ma, takes a different route: instead of optimizing a single reconstruction result, it searches for the algorithm that produces that result. An LLM-driven multi-agent system starts from a classical reconstruction algorithm, edits only user-selected components, and rejects candidates violating physical constraints. The survivors are scored by execution, yielding a standalone, white-box algorithm—no agent or neural network needed at inference time.

Results are striking. For interpolation without added noise, SeisEvo discovered a residual-gated, phase-aligned dip-consistency projection, dubbed Evo-POCS, which improves SNR by 3.49dB on average over classic POCS across missing ratios from 30% to 70%. For simultaneous interpolation and denoising, it discovered a reliability-grouped singular-value shrinkage (Evo-MSSA) that beats classic MSSA by more than 7dB and a stronger rank-reduction baseline by more than 3dB. Both operators retain their gains on unseen data. This is the first study framing seismic reconstruction operator design as a constrained, LLM-driven program evolution task, suggesting agentic evolution can complement deep learning in producing explicit, deployable geophysical processing algorithms.