Memory-Efficient Deep-Learning-Matching Full Waveform Inversion for Field-Scale Datasets

Summary Classical full waveform inversion (FWI) is sensitive to kinematic inaccuracies in the initial velocity model, converging to local minima due to cycle skipping when low-frequency data are absent. The recently developed deep-learning-matching FWI (DLM-FWI) framework mitigates this by dynamically bridging the phase gap between synthetic and observed data, utilizing a convolutional neural network (CNN) within the inversion loop as a robust kinematic preconditioner. However, because standard deep learning (DL) automatic differentiation (AD) engines scale poorly to large-scale wave equations, deploying DLM-FWI on practical field-scale data has been computationally prohibitive. To overcome this, we introduce a memory-efficient, hybrid computational workflow that interfaces the Flux machine-learning ecosystem with the Julia Devito Inversion framework (JUDI) finite-difference library. By decoupling the network’s AD graph from the physical solver and evaluating the chain rule strictly in the data domain, we avoid storing massive spatiotemporal wavefield histories. After verifying that this decoupled architecture maintains resilience to source-wavelet mismatch and ambient noise on the Marmousi2 synthetic model, we apply the workflow to the Mobil AVO Viking Graben Line 12 marine dataset to evaluate its practical scalability. Compared to classical L-BFGS FWI, the field-scale DLM-FWI results exhibit robust convergence, mitigation of shallow cycle-skipped artifacts, and geologically coherent structural continuity. This study addresses the gap between small-scale synthetic DL studies and field-scale 2D implementations, demonstrating that physics-guided matching strategies can be deployed in practice to improve FWI robustness on complex field datasets.