From Seismic to Shear: Learning Cumulative Strain for Fault Zoning

A physics‑informed alternative to fault likelihood was developed by teaching a U‑Net to predict cumulative maximum shear strain, γ_{p,max}, from seismic images. Mechanics‑consistent pairs were created by running CDEM simulations of a 45° normal fault in a 48‑layer, 5 km × 2.77 km model with 10 891 particles across 521 time steps; particle‑scale density and velocity updates yielded impedance, vertical reflection coefficients, and band‑limited seismic via 20/40/60 Hz Ricker wavelets, providing inputs matched to γ_{p,max} labels. The U‑Net (MSE loss, AdamW 1×10⁻⁴, 250 epochs) achieved train/val MSE ≈0.0027/0.0056 and test R² ≈0.64. Predicted high‑strain ribbons tracked major fault cores, tips, and splays while preserving finite deformation‑zone thickness and exhibiting tolerance to minor mis‑ties—advantages over categorical masks that compress faults to one‑voxel filaments. Newly learned limitations include under‑resolution of very thin splays and tip‑localized deformation and occasional false positives where strong stratigraphic edges mimic discontinuities; mild cross‑line striping reflects training on 2‑D inline patches. Overall, γ_{p,max} regression is a practical, interpretable path to rank structures by deformation intensity and to derive discrete surfaces by threshold‑and‑link post‑processing, with clear upgrade paths via multi‑task (mask + intensity + orientation) learning and broader 3‑D simulation libraries.