Checkpointing-Assisted Automatic Differentiation Full Waveform Inversion

Full waveform inversion (FWI) is a powerful technique for high-resolution subsurface imaging in seismic exploration. The emergence of automatic differentiation full waveform inversion (ADFWI) further enhances this process by enabling more accurate and efficient gradient computation through automatic differentiation, simplifying the implementation of complex workflows and reducing human error. However, the memory requirements of ADFWI are drastically higher than those of traditional FWI, as the entire computation graph must be retained for ADFWI, necessitating the storage of numerous intermediate states. To address this challenge, we propose checkpointing-assisted and disk-checkpointed strategies that reduce memory usage by selectively saving and recomputing intermediate states during the backward pass. This paper discusses the impact of different checkpointing strategies on memory usage, runtime performance, and inversion accuracy. We analyze the trade-offs of varying the number of checkpoints and find that the relationship between memory usage and runtime is nonlinear, following a U-shaped curve. Additionally, the inversion accuracy decreases as the number of checkpoints increases. Field data applications in the Chicxulub Crater confirm that the method is robust, achieving memory-efficient inversions with geologically consistent results under hardware constraints. Experimental results highlight the importance of carefully balancing memory efficiency, computational overhead, and accuracy when selecting the optimal checkpointing strategy. This study concludes that a systematic trade-off analysis is essential for determining the best parameters in large-scale or complex-media inversion scenarios.