Faults are important to structural interpretation because their geometry affects fault traps, seal behavior, reservoir compartmentalization, fractured zones and possible fluid pathways. Yet they remain difficult to identify automatically in three-dimensional seismic data. Fault voxels occupy only a small fraction of most volumes, while noise, weak reflections, acquisition footprints and stratigraphic discontinuities can obscure thin or branching structures. Conventional attributes such as coherence and fault likelihood remain physically interpretable and useful, but their outputs often require parameter tuning, threshold selection and manual screening. UMamba3D was developed as an interpretation-support system intended to identify fault candidates for subsequent geological validation, not to replace that final interpretation.

The network has a U-shaped encoder-decoder topology with three downsampling and three corresponding upsampling stages. Convolutional blocks at each level use two 3D convolutions, batch normalization and ReLU activation, while skip connections concatenate encoder and decoder features so that shallow spatial detail is retained. The channel width progresses from 8 to 16, 32 and 64. At the 16 × 16 × 16 bottleneck, a sequence-mixing module reshapes the volumetric features into a spatial sequence, applies depth-wise 1D convolution, layer normalization, channel projection and SiLU activation, and then fuses the result residually with the original bottleneck. The authors deliberately describe this component as “Mamba-inspired”: it does not implement Mamba’s selective scan or input-dependent state-transition matrices, and its narrower purpose is efficient contextual interaction beyond the local receptive field of 3D convolution.

Training combined class-balanced binary cross-entropy, Tversky loss and a local structure constraint. Because background voxels greatly outnumber fault voxels in the public FaultSeg3D benchmark, ordinary binary cross-entropy can be dominated by background classification. Class weighting increases the contribution of sparse foreground voxels, while the Tversky term with alpha set to 0.8 and beta to 0.2 penalizes false positives more strongly than false negatives. The structure term uses local pooling with a radius of one voxel to encourage probability near labeled faults and penalize distant responses. This local tolerance is intended for thin faults, where a one-voxel offset can sharply reduce exact-overlap scores. The training used 200 synthetic cubes and 20 independent validation cubes, each measuring 128 × 128 × 128 voxels; no field labels were used for training in the reported application.

With height and width flip test-time augmentation, the model averaged predictions from the original volume and its H-, W- and combined-axis flips before applying a probability threshold of 0.90. On the 20 validation volumes, it achieved a Dice coefficient of 0.7929, intersection over union of 0.6593, precision of 0.7946 and recall of 0.7949. The precision-oriented F0.5 score was 0.7934, the recall-oriented F2 score was 0.7936, and a distance-tolerant thin-fault figure of merit was 0.9186. Its predicted-positive ratio was 0.0728, close to the 0.0723 ratio in the labels. Without augmentation, Dice fell to 0.7764 and precision to 0.7348 while recall rose to 0.8271, indicating that the four-view ensemble primarily improved false-positive control and structural consistency rather than simply adding more fault voxels.

The threshold analysis confirmed the expected trade-off. Raising the cutoff from 0.90 to 0.92 increased precision from 0.7946 to 0.8125 but reduced recall from 0.7949 to 0.7701; at 0.95, precision reached 0.8474 and recall fell to 0.7114. The investigators selected 0.90 as the main operating point for the augmented result. In retrained comparisons without test-time augmentation, plain 3D U-Net, a parameter-matched convolutional bottleneck and UMamba3D produced nearly identical clean-data Dice scores of 0.7824, 0.7796 and 0.7789. A randomly initialized SwinUNETR baseline achieved higher precision, 0.8441, but lower recall, 0.6215, Dice, 0.7141, F2, 0.6552 and thin-fault figure of merit, 0.7953. The result characterizes SwinUNETR as precision-oriented and under-segmenting under this protocol, not as a universally weak architecture; the study did not test pretraining or substantially larger training sets.

Controlled tests nevertheless showed that the trained UMamba3D checkpoint functionally relied on its sequence-mixing bottleneck. Replacing that module with an identity map after training, without changing the shared weights or data, lowered Dice by 0.0623, precision by 0.0822, F0.5 by 0.0767 and thin-fault figure of merit by 0.0976, while recall changed only slightly. Because the bypassed network was not retrained, this experiment demonstrates functional dependence rather than proving that sequence mixing is superior to every alternative. The more stringent retrained comparison found that the parameter-matched convolutional bottleneck remained essentially equal on clean-data overlap metrics, although UMamba3D had slightly higher recall, F2 and thin-fault agreement.

Robustness was evaluated by adding Gaussian noise equal to 0%, 5%, 10%, 15% and 20% of each input volume’s standard deviation, using identical perturbations, an unchanged threshold of 0.92 and no test-time augmentation. The models were closely matched under clean and mild perturbations, and the convolutional bottleneck was marginally better at 15% noise. At 20%, however, UMamba3D reached Dice of 0.5649, recall of 0.5706, F2 of 0.5668 and thin-fault figure of merit of 0.7441. The corresponding values were 0.5343, 0.4785, 0.4972 and 0.6479 for the convolutional bottleneck, and 0.4225, 0.3360, 0.3639 and 0.4805 for plain U-Net. UMamba3D retained 72.53% of its clean Dice, compared with 68.54% and 54.00%. Its precision was lower at the strongest perturbation, so the observed advantage was primarily preservation of weak or continuous fault responses, not improvement in every metric.

Loss ablations showed complementary effects among the objective’s components. Class-balanced binary cross-entropy alone produced the highest precision, 0.7999, but the lowest recall, 0.7393; adding Dice raised recall to 0.7913. Relative to the directly matched balanced binary cross-entropy plus Tversky objective, adding the local structure term increased Dice from 0.7647 to 0.7685, recall from 0.7643 to 0.7758 and F2 from 0.7643 to 0.7726, but reduced precision from 0.7665 to 0.7636 and thin-fault figure of merit from 0.9279 to 0.9221. The local term therefore behaved as a modest recall-oriented regularizer rather than a universal geometry guarantee. One-factor sensitivity tests produced Dice scores between 0.7644 and 0.7712, indicating stable performance around the selected settings.

UMamba3D contained 355,705 trainable parameters, compared with 351,161 for plain U-Net and 15,702,979 for SwinUNETR. For one 128 × 128 × 128 volume, the UMamba3D and U-Net counts were nearly identical at 32.006 and 31.988 billion multiply-accumulate operations, with peak allocated GPU memory of about 506.5 MB in both cases. Measured inference latency was 17.152 ± 1.353 milliseconds for UMamba3D, 19.452 ± 4.533 milliseconds for U-Net and 139.250 ± 9.883 milliseconds for SwinUNETR. The model was then applied through overlapping sliding windows to a confidential Sichuan Basin volume of approximately 501 × 381 × 551 samples. Its time-slice and arbitrary-line outputs were visually more coherent than eigen coherence, Thin Likelihood and U-Net, preserving branches and reflector-linked fault trends while suppressing fragmented responses. Because the field data were not publicly available and the comparisons were based on visual interpretation rather than ground-truth field scoring, the study positions UMamba3D as a promising interpretation aid requiring labeled multi-survey and multi-seed validation.

Subject of Research: Evaluation of a lightweight structure-constrained neural network for 3D seismic fault segmentation

Article Title: Structure-constrained UMamba3D for seismic fault segmentation and application in the Sichuan Basin

Article References: Li, L., Chen, M., Zhang, W., Li, Y., Wang, X., Wu, W., Wang, L., Lin, P., & Li, C. (2026). Structure-constrained UMamba3D for seismic fault segmentation and application in the Sichuan Basin. Frontiers in Earth Science, 14, Article 1930523. https://doi.org/10.3389/feart.2026.1930523

Image Credits: AI Generated

DOI: 10.3389/feart.2026.1930523

Keywords: seismic fault segmentation, UMamba3D, 3D seismic data, deep learning, Mamba-inspired networks, Sichuan Basin, geophysical interpretation, fault detection, noise robustness, Structure-constrained, seismic, fault