AI 中文总结
该研究提出采用几何不变物理编码的笔记本电脑级深度残差全波形反演流程，结合物理条件共形预测校准的集成模型，在合成数据上降误差38%，零样本应用于Marmousi-2测线时可恢复损坏条件下的覆盖率。
AI 中文摘要
直接将地震数据映射到速度模型的神经网络往往会记住其训练时的采集几何，且一旦数据偏离训练分布，其不确定性估计就很少可信。我们描述了一种笔记本电脑级别的流程，可解决这两个问题。该网络从未见过炮集数据，而是将可变采集几何映射为由经典物理算子计算的固定模型空间表示：初始模型、正则化经典反演结果、两个残差梯度图像，以及来自快速行进旅行时的六个照明和波数覆盖图。学习组件并非FWI的替代，而是应用于该物理先验的校准残差校正器。一个由六个异构成员组成的集成模型，带有逐像素方差头，通过物理条件共形预测进行校准，在校准分布上提供有限样本逐像素边际覆盖率。超出该分布后，保留炮的物理审计通过预测分布的样本模拟反演从未使用的炮，并在其数据空间覆盖率峰值处重新缩放区间宽度，且不涉及真实值。在涵盖六个采集家族的1000个合成模型语料库上，该集成模型相对于其经典先验将误差降低了38%，并在从未见过的几何上迁移而无明显性能下降。零样本应用于完整17 km Marmousi-2测线时，其误差从354 m/s降至304 m/s，而原始覆盖率降至0.42；该审计在涵盖噪声、子波误差和炮抽取的11种损坏条件下将覆盖率恢复至0.89-0.91，且其峰值高度可清晰区分物理失配与良性损坏。预算匹配的基于炮集的基线在其训练采集之外表现显著更差。代码和检查点将归档于Zenodo。
英文摘要
Neural networks that map seismic data directly to velocity models tend to memorize the acquisition geometry they were trained on, and their uncertainty estimates are rarely trustworthy once the data drift away from the training distribution. We describe a laptop-scale pipeline that addresses both problems. The network never sees shot gathers. Instead, variable acquisition geometries are mapped into a fixed model-space representation computed by classical physics operators: the starting model, a regularized classical inversion, two misfit-gradient images, and six illumination and wavenumber-coverage maps from fast-marching traveltimes. The learned component is not a replacement for FWI; it is a calibrated residual corrector applied to this physics-derived prior. A six-member heterogeneous ensemble with per-pixel variance heads is calibrated by physics-conditioned conformal prediction, giving finite-sample pixel-wise marginal coverage on the calibration distribution. Beyond that distribution, a held-out-shot physics audit simulates shots the inversion never used through samples of the predictive distribution and rescales interval width where their data-space coverage peaks; no ground truth is involved. On a corpus of 1000 synthetic models spanning six acquisition families, the ensemble reduces error by 38% relative to its classical prior and transfers to never-seen geometries without measurable degradation. Applied zero-shot to the full 17 km Marmousi-2 line, it lowers the error from 354 to 304 m/s while raw coverage collapses to 0.42; the audit restores coverage to 0.89-0.91 across eleven corruption conditions covering noise, wavelet error, and shot decimation, and its peak height cleanly separates physics mismatch from benign corruptions. Budget-matched gather-based baselines underperform substantially off their training acquisition. Code and checkpoints will be archived on Zenodo.
Comments39 pages, 12 figures, 6 supplementary figures. Prepared for submission to Computers and Geosciences
