SeismoLabV3+: Deep Learning for Seismic Inversion

• SeismoLabV3+ is a seismic-velocity inversion model that adapts DeepLabV3+ for dense regression, leveraging multi-channel shot gathers to image subsurface structures.
• It integrates a ResNeXt50_32x4d backbone, ASPP, and a regression head to preserve geological context and local interface details in the inversion process.
• Benchmark tests on the ThinkOnward 2025 dataset show that SeismoLabV3+ achieves lower MAPE than U-Net variants, underscoring its incremental but effective performance gain.

SeismoLabV3+ is a seismic-velocity inversion model that adapts DeepLabV3+ from semantic segmentation to dense regression for subsurface imaging. In the formulation reported in "Seismic Velocity Inversion from Multi-Source Shot Gathers Using Deep Segmentation Networks: Benchmarking U-Net Variants and SeismoLabV3+" (Hasan, 7 Sep 2025), the model ingests a five-channel seismic shot-gather tensor and predicts a continuous-valued velocity map, with the stated objective of reconstructing subsurface structure from multi-source seismic wave data. Within that benchmark, SeismoLabV3+ is presented not as a new network family from first principles, but as a task-adapted DeepLabV3+ variant whose defining changes include a ResNeXt50_32x4d backbone, five-channel input handling, and a single-channel linear regression head.

1. Definition and problem setting

SeismoLabV3+ is defined in the literature as an optimized DeepLabV3+ variant for seismic velocity inversion, benchmarked against U-Net, U-Net++, and a baseline DeepLabV3+ on the ThinkOnward 2025 Speed Structure Challenge dataset (Hasan, 7 Sep 2025). The inversion task is framed as dense prediction: the input is explicitly $\mathbf{X}\in \mathbb{R}^{C\times H\times W}$ with $C=5$, $H=300$, and $W=1259$, while the target is a velocity image $\mathbf{Y}\in\mathbb{R}^{H\times W}$.

This formulation is significant because it recasts a classical inverse problem into the architecture space of modern encoder–decoder vision models. Traditional physics-driven methods such as Full Waveform Inversion are described as computationally demanding, sensitive to initialization, and limited by seismic bandwidth, whereas the benchmark evaluates whether dense neural predictors can recover high-resolution velocity structure from multi-shot gathers. In that framing, SeismoLabV3+ is aligned with a data-driven surrogate paradigm rather than direct numerical inversion.

The paper emphasizes that the task requires inference of a full subsurface field from multi-source responses. That requirement makes both local interface preservation and broader geological context central design constraints. SeismoLabV3+ is therefore positioned at the intersection of geophysical inversion and segmentation-derived multi-scale representation learning rather than as a purely generic image-to-image regressor.

2. Data representation and benchmark dataset

The benchmark uses the official ThinkOnward 2025 Speed Structure Challenge dataset, consisting of 2,000 labeled training samples and 150 unlabeled test samples (Hasan, 7 Sep 2025). For each sample, five files named like receiver_data_src_<i>.npy correspond to sources at 1, 75, 150, 225, and 300, each with shape $300\times1259$, and the target file vp_model.npy stores the ground-truth velocity model with the same spatial size.

The five source gathers are stacked into a 5-channel tensor, and each channel is normalized to $[0,1]$. This representation preserves the multi-source acquisition structure explicitly instead of collapsing the observations into a single channel. A plausible implication is that the network can learn source-dependent correlations and shared structural cues across acquisition geometries, although the paper does not provide an explicit feature-level ablation of that effect.

For internal validation, the 2,000 labeled samples are split into 60% training, 20% validation, and 20% test with a fixed random seed of 42. For external comparison, the official hidden test set is used through leaderboard scoring. This dual evaluation protocol matters because it separates model selection on a visible split from final assessment on unavailable labels, reducing the interpretive ambiguity that can arise from reporting only internal validation.

3. Architectural design and task adaptation

SeismoLabV3+ retains the overall encoder–decoder structure of DeepLabV3+, including a backbone feature extractor, Atrous Spatial Pyramid Pooling (ASPP), and a lightweight decoder that fuses high-level and low-level features to sharpen boundaries (Hasan, 7 Sep 2025). Its contribution lies in adapting that structure to seismic regression rather than semantic labeling.

Component	SeismoLabV3+ specification
Backbone	ResNeXt50_32x4d with ImageNet pretraining
Input adaptation	First convolution modified for five seismic channels
Output head	Single-channel linear regression head
ASPP	Dilation rates $(12,24,36)$, dropout probability 0.5
Decoder	Refinement path retained for high-/low-level feature fusion
Encoder output stride	16

The shift from an ordinary ResNet-50 backbone to ResNeXt50_32x4d is one of the model’s principal modifications. The authors also report exploring ResNet-34, ResNet-50, ResNeXt-50 (32×4d), and EfficientNet-B0, as well as ImageNet, SSL, and SWSL initialization schemes, with ImageNet-pretrained ResNeXt50_32x4d selected as the best-performing configuration. The paper does not provide a full ablation table, but its narrative attributes the final gain to the combined effect of the stronger backbone and the regression-specific redesign.

The architectural rationale is explicitly comparative. Relative to U-Net, SeismoLabV3+ is described as more context-aware because atrous convolutions and ASPP expand the receptive field without excessive downsampling. Relative to U-Net++, it trades denser skip topology and deep supervision for stronger multi-scale context aggregation. The paper’s interpretation is that this balance is better aligned with seismic velocity inversion, where broad geological context, localized interfaces, and long-range dependencies all matter.

4. Preprocessing, training protocol, and evaluation criterion

Training and preprocessing are kept consistent across all benchmarked models to maintain comparability (Hasan, 7 Sep 2025). After channel stacking and normalization, the data are converted to PyTorch tensors. Padding is then used to satisfy architecture-specific divisibility constraints: U-Net++ inputs are padded to $320\times1280$, while U-Net and SeismoLabV3+ are padded to $304\times1264$. Ground-truth velocity maps are padded or resized in the same way to preserve spatial alignment.

The implementation stack is PyTorch with segmentation-models-pytorch, trained on NVIDIA A100 GPUs with automatic mixed precision enabled. Optimization uses AdamW with weight decay $C=5$0, learning rate $C=5$1, and a ReduceLROnPlateau scheduler. All models are trained for 50 epochs with batch size 4. These choices define the benchmark protocol rather than an isolated model specification; their importance lies in ensuring that performance differences are attributed to architecture and task adaptation rather than to substantially different optimization regimes.

The sole reported quantitative metric is MAPE, which is also treated as the official evaluation criterion because hidden test labels are not available. The paper gives the metric as

$C=5$2

where $C=5$3 is the ground-truth value, $C=5$4 is the predicted value, and $C=5$5 is the number of data points. Lower MAPE indicates better inversion accuracy. The use of a single metric makes the benchmark operationally direct, though it also means that no complementary error measures are reported for boundary sharpness, structural similarity, or calibration.

5. Benchmark performance and qualitative behavior

SeismoLabV3+ achieves the best reported MAPE on both the internal split and the official hidden test set (Hasan, 7 Sep 2025).

Model	Internal 60/20/20 MAPE	Hidden test MAPE
U-Net	0.03084	0.033172
U-Net++	0.03049	0.032766
DeepLabV3+	0.03038	0.031762
SeismoLabV3+	0.03025	0.031246

The numerical pattern is consistent across both evaluation settings. SeismoLabV3+ outperforms the baseline DeepLabV3+, and both DeepLab-based models outperform U-Net and U-Net++. The paper characterizes the gain over DeepLabV3+ as modest rather than dramatic, which is important for interpretation: the results support incremental improvement through task-specific refinement rather than a categorical architectural discontinuity.

The hidden test scores are slightly worse than internal validation by about 0.001–0.003 MAPE, which the authors describe as a normal generalization gap. This point matters because it suggests that internal ranking and leaderboard ranking are directionally aligned. The paper’s broader conclusion is that multi-scale context and regression-oriented adaptation are more consequential than increasingly elaborate skip-connection schemes for this dataset and task formulation.

Qualitative analysis reinforces that interpretation. In a random test sample, the ground-truth velocity map is described as transitioning from low velocity on the left to high velocity on the right, with thin quasi-horizontal bands and laminations. SeismoLabV3+ captures the broad gradient and the low-velocity wedge near $C=5$6, but the predicted map is smoother than the ground truth and loses many fine stratified details and sharp layer boundaries. The paper reads this as typical of regression models that recover macro-structure more reliably than high-frequency detail.

6. Interpretation, scope, and relation to naming ambiguity

The paper’s explanation for SeismoLabV3+ centers on the structure of seismic velocity inversion as an ill-posed, band-limited problem that requires integrating information over large distances in the shot gathers (Hasan, 7 Sep 2025). In that account, U-Net-style skip connections preserve local detail but do not explicitly encode multi-scale contextual reasoning to the same degree as ASPP-based architectures. SeismoLabV3+ therefore derives its advantage from balancing global context, boundary refinement, and domain-specific input/output adaptation.

This framing has a methodological implication. SeismoLabV3+ is not presented as a replacement for physics-based inversion in the strict mechanistic sense; rather, it is described as evidence that deep encoder–decoder segmentation models, when adapted carefully, can serve as practical surrogates for computationally expensive physics-based methods when the aim is fast, data-driven estimation of velocity fields. The paper’s emphasis on hidden-set consistency and macro-structural recovery supports that narrower interpretation.

A separate issue is terminological ambiguity around the name. Another paper, "Earthquake Simulation" (Chawla, 13 Apr 2025), describes a simple, educational earthquake simulation system written in C that is only conceptually related to a tool like SeismoLabV3+ at a high level. Its simulator visualizes fault patterns and surface stress or pressure accumulation on a 2D grid using randomized updates and color-coded terminal output; it is explicitly not presented as a numerical geophysics engine, a data assimilation framework, or a real-world forecasting system. The comparison clarifies a common misconception: SeismoLabV3+, as described in (Hasan, 7 Sep 2025), is a deep regression model for seismic velocity inversion from multi-source shot gathers, whereas the earlier C-based simulator is a pedagogical visualization tool with a fundamentally different scope.

That distinction also helps delimit what SeismoLabV3+ does not claim. The benchmark does not report physically realistic wave propagation inside the model, parameter-count analyses, exhaustive ablations, or alternative geophysical objectives beyond MAPE-minimizing velocity reconstruction. Its contribution is specific: adapting DeepLabV3+ into a five-channel, single-output regression architecture for the ThinkOnward 2025 Speed Structure task and showing that this adaptation yields the best reported benchmark performance among the compared encoder–decoder models.