OpenFWI: Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion
New1. Overview
Full waveform inversion (FWI) reconstructs subsurface velocity structure from surface-recorded seismic waveforms, and it is the shared technical foundation for carbon sequestration, reservoir characterization, subsurface energy exploration, and earthquake early warning. This line of work has long been dominated by physics-driven formulations, whose costs are high computational consumption, cycle skipping, and ill-posedness. Deep learning rewrote the pipeline into a data-driven inverse mapping, and the burden shifted onto the coverage of the training data. That is why the community needs a public, structurally diverse benchmark with a unified evaluation protocol.
OpenFWI was built by a team at Los Alamos National Laboratory and is among the larger and more structurally varied benchmarks of this kind. The project homepage defines it as a large-scale multi-structural benchmark dataset collection for deep-learning-driven seismic FWI, containing twelve datasets together with baseline results for four deep learning methods.
The datasets form four families. The Vel family supplies flat and curved layer velocity maps that characterize interface imaging. The Fault family introduces fault-induced discontinuities for fault detection and reservoir localization. The Style family generates velocity maps by style transfer from natural images to raise geological diversity. The Kimberlina family is built around a geologic carbon storage reservoir, containing Kimberlina-CO2, which describes the spatial and temporal migration of a supercritical CO2 plume, plus 3D Kimberlina-V1, the first large-scale public 3D FWI dataset.
As of November 2026 the GitHub repository has 182 stars, 34 forks, and 5 open issues. The code is released under BSD 3-Clause and the data under CC BY-NC-SA 4.0. The paper was accepted as a spotlight in the NeurIPS 2022 Datasets and Benchmarks track.
2. How It Works
The governing equation for acoustic wave propagation in an isotropic medium with constant density is the Laplacian of pressure minus the second time derivative scaled by the inverse squared velocity, equal to the source term. Here c is the velocity map, p the pressure field, and s the source term. Given velocity c, the forward operation that computes the pressure field p is written as p = f(c), where f is a highly nonlinear forward mapping. Data-driven FWI trains a neural network to learn the inverse mapping c = f⁻¹(p). The project focuses on active-source imaging, so the source term s is given.
Four baseline methods are supplied. InversionNet models the inversion process with a fully convolutional network, where encoder and decoder are trained in a supervised scheme that maps multi-source two-dimensional seismic data to two-dimensional velocity maps. VelocityGAN solves the problem with a generative adversarial network: the generator keeps the encoder-decoder structure, the discriminator is a CNN that separates real from synthesized velocity maps, and network-based deep transfer learning is used to improve robustness and generalization. UPFWI connects forward modeling and a CNN in a loop to achieve unsupervised learning without ground truth velocity maps; the CNN predicts a velocity map from the seismic data, the prediction is fed into differentiable forward modeling to reconstruct the seismic data, and the loop closes by comparing input and reconstruction. InversionNet3D extends InversionNet to three dimensions, adopting group convolution in the encoder and a partially reversible architecture built on invertible layers with additive coupling to reduce the memory footprint and improve computational efficiency.
Velocity map synthesis is the first stage of data generation, simulating interfaces, layers, and faults from different priors, and forward modeling then produces the seismic data. The dataset parameters are recorded in dataset_config.json, covering minimum and maximum velocity, grid size, and frequency, and they are also consumed by the normalization and forward modeling modules of UPFWI.
The Vel, Fault, and Style families share one parameter set: 10 m grid spacing, a 0.7 by 0.7 km² spatial extent, a velocity range of 1500 to 4500 m/s, 140 m source spacing, 0.7 km source line length, 10 m receiver spacing, 0.7 km receiver line length, a 0.001 s time step, and 1 s of recorded time. The configuration file also records five sources, 70 receivers, 1000 time samples, a 15 Hz dominant frequency, a 70-point grid, and 500 samples per file.
Kimberlina-CO2 differs: its extent is 1.4 by 4 km², source spacing is 400 m with a 3.6 km source line, receiver spacing is 40 m with a 4 km receiver line, the time step is 0.002 s, and 2.5 s are recorded. 3D Kimberlina-V1 spans 3.5 by 4 by 4 km³, with 800 m source spacing, 100 m receiver spacing, a 0.001 s time step, and 5 s of recorded time.
Complexity is characterized with three standard metrics: Shannon entropy, spatial information that captures average boundary magnitude, and gradient sparsity index that measures the share of non-smooth pixels. These are used to relate geological complexity to inversion performance.
3. Scope and Assumptions
Practical prerequisites are substantial. A user must be able to absorb data volumes between 43 GB and 1.4 TB, plus training resources that fit a two-dimensional 70 by 70 grid or a three-dimensional 350 by 400 by 400 one. The repository environment is dated: the README asks for PyTorch 1.7.1 aligned with torchvision 0.8.2, alongside scikit-learn, numpy, and matplotlib.
Licensing constrains use. The data carries CC BY-NC-SA 4.0, including a non-commercial clause and a share-alike clause, so commercial pipelines and redistribution scenarios need separate review. The code itself is BSD 3-Clause and can be used freely.
Evaluation results are tied to the paper protocol. Changing the loss function, the train-validation split, or the normalization range makes the official numbers non-comparable. UPFWI and InversionNet3D do not report results on every dataset, leaving gaps on Kimberlina-CO2 and the 3D set.
Scope boundaries are explicit in the sources. The official benchmarks cover active-source imaging, while passive seismic work appears only as a preliminary event-picking test in the supplementary material. The paper itself names four open challenges: inversion for complex velocity maps, generalization of data-driven methods, computational efficiency, and passive seismic imaging.
A further limit is code availability. The repository README states that InversionNet and VelocityGAN are fully supported across ten datasets, while the UPFWI and InversionNet3D code is pending approval. Reproducing all four baselines therefore requires filling that gap independently or waiting for release.
Hardware and inference-side constraints matter as well. All training durations reported in the paper are measured on NVIDIA Tesla P100, with UPFWI and InversionNet3D occupying 32 GPUs and the remaining methods using a single card, so these figures cannot be carried over to other hardware without recalibration. Inference has its own boundary: the InversionNet architecture fixes the input and output at a two-dimensional 70 by 70 grid, so transferring to a different grid size requires redesigning the network or adding an interpolation step. Kimberlina-CO2 has a different input and output shape from the other three families, and the paper notes that the model architecture must be adjusted slightly for it.
Data consistency is another boundary. Normalization bounds are specified per dataset and are not shared across datasets. Velocity maps are mapped uniformly into the 1500 to 4500 m/s range, while seismic data receives a log and min-max normalization against each dataset's own bounds. Applying one dataset's statistics to another breaks training and inference consistency outright.
4. Getting Started
Data comes first. The project homepage lists all twelve datasets by family, with name, size, training and validation sample counts, input and output shapes, and physical meaning. Downloads are hosted on Google Drive, while 3D Kimberlina-V1 was released on the EDX site in June 2023 together with a data paper in Geophysical Data Journal.
After unpacking, verify the shapes. The Vel, Fault, and Style families ship as .npy files, each holding a batch of 500 samples, where data{i}.npy holds the i-th seismic batch and model{i}.npy the matching velocity maps. Seismic data has shape (500, 5, 1000, 70) and velocity maps have shape (500, 1, 70, 70). Each Kimberlina-CO2 file holds a single sample, with velocity map shape (401, 141) and seismic data shape (9, 1251, 101).
The code side is organized by role. dataset.py defines FWIDataset, a torch.utils.data subclass that loads samples into the dataloader according to a text manifest. network.py holds the InversionNet and VelocityGAN architectures. train.py and gan_train.py run the two training loops. transforms.py and utils.py provide normalization, metric logging, and loss functions, and vis.py handles visualization of velocity maps and seismic data. scheduler.py and pytorch_ssim.py cover learning-rate scheduling and the SSIM metric; the latter is borrowed from the pytorch-ssim project and must be kept next to rainbow256.npy.
Splits are expressed through manifests. The split_files/ directory provides 36 manifest files covering the flatvel, curvevel, flatfault, curvefault, style, kaggle_tutorial, and tutorial series, each with a training and a validation list. Every line holds a directory path plus a filename, and the loader parses them line by line.
To train InversionNet from scratch on FlatVel-A with an ℓ1 loss, run python train.py -ds flatvel-a -n YOUR_DIRECTORY -m InversionNet -g2v 0 --tensorboard -t flatvel_a_train.txt -v flatvel_a_val.txt. The -ds flag selects the dataset, -n names the directory for checkpoints and logs, -g2v zeroes the ℓ2 loss coefficient, and -t and -v assign the training and validation manifests. Add -r CHECKPOINT.PTH to resume from a checkpoint. Testing runs python test.py -ds flatvel-a -n YOUR_DIRECTORY -m InversionNet -v flatvel_a_val.txt -r CHECKPOINT.PTH --vis -vb 2 -vsa 3, where --vis enables visualization and -vb and -vsa control how many maps are shown. VelocityGAN swaps in gan_train.py and keeps the same arguments for testing.
Loading applies transforms first. Seismic data goes through a LogTransform plus MinMaxNormalize composition and velocity maps through MinMaxNormalize, bringing both into the range [−1, 1]. Before visualization, tonumpy_denormalize() in transforms.py restores the physical scale.
A Google Colab tutorial provides an end-to-end walkthrough on a toy dataset for readers new to data-driven FWI.
5. Comparison with Alternatives
Against physics-driven FWI, the paper compares accuracy and computational cost directly. Data-driven methods win on inversion accuracy by a large margin and are faster when the ratio of training to test samples is below 62. That ratio marks a practical dividing line: scarce samples favor data-driven approaches, while abundant samples amortize the cost of forward solving.
Against real-data studies, multiscale methods couple two-dimensional data-driven inversion with field seismic surveys, learning from synthetic data and transferring to measurements. OpenFWI is the synthetic base and contains no measured data itself.
Against unsupervised and physics-guided methods, UPFWI represents the label-free route. It avoids ground truth but is computationally expensive and, on most datasets, less accurate than the two supervised baselines. Supplementary experiments compare physics-guided data-driven inversion with purely data-driven methods, discussing adaptive data augmentation and physical consistency constraints.
Against generative and diffusion-prior methods, diffusion priors can guide inversion recovery in a plug-and-play manner, and learned diffusion priors can regularize the FWI optimization landscape. WaveDiffusion jointly generates seismic data and velocity maps in a shared latent space, while VelocityGPT explores autoregressive layer-by-layer velocity generation with uncertainty quantification. These approaches seek generative priors outside the OpenFWI setting and complement, rather than overlap with, its large-scale paired supervision.
Against large-scale model routes, work published in 2026 reports results directly on the OpenFWI benchmark with a billion-parameter model. The route differs clearly from the four original baselines: inversion is reframed as discrete sequence generation rather than convolutional regression.
For positioning within this site, OpenFWI is a structured data benchmark for inverse problems. It reads well beside surrogate modeling and reduced order and uncertainty and Bayesian inversion; for engineering context, seismic structural engineering AI and reservoir simulation complete the chain from inversion algorithm to field application.
6. Benchmarks and Evidence
The twelve datasets total roughly 2.1 TB, with sample counts spanning 2K to 67K.
The Vel family has FlatVel-A, FlatVel-B, CurveVel-A, and CurveVel-B at 43 GB each, 24K training and 6K validation samples, input (5, 1000, 70), and output (1, 70, 70). The A versions are the simple cases and the B versions the difficult ones. The Fault family has FlatFault-A, FlatFault-B, CurveFault-A, and CurveFault-B at 77 GB each with 48K training and 6K validation samples, identical shapes, and a physical reading of flat or curved layers plus one fault. The Style family has Style-A and Style-B at 95 GB each with 60K training and 7K validation samples, generated by style transfer from natural images; the A version is smoothed by a Gaussian filter with fewer events in the seismic data, and the B version is high-resolution. The Kimberlina family has Kimberlina-CO2 at 96 GB with 15K training and 4430 validation samples, input (9, 1251, 101) and output (1, 401, 141), plus 3D Kimberlina-V1 at 1.4 TB with 1664 training and 163 validation samples, input (25, 5001, 40, 40) and output (1, 350, 400, 400).
Three metrics are used: MAE, RMSE, and SSIM. InversionNet and VelocityGAN report results for both ℓ1 and ℓ2 losses, while UPFWI combines ℓ1, ℓ2, and perceptual losses.
On the relatively simple FlatVel-A, InversionNet with ℓ1 reaches MAE 0.0131, RMSE 0.0211, and SSIM 0.9895, and with ℓ2 reaches MAE 0.0111, RMSE 0.0180, and SSIM 0.9887. VelocityGAN with ℓ1 reaches MAE 0.0118, RMSE 0.0178, and SSIM 0.9916, while UPFWI reaches MAE 0.0621, RMSE 0.1233, and SSIM 0.9565. On the harder CurveVel-B, InversionNet with ℓ1 falls to MAE 0.1497 and SSIM 0.6727, VelocityGAN with ℓ1 gives MAE 0.1268 and SSIM 0.7111, and UPFWI gives MAE 0.1777 and SSIM 0.6614. On the hardest fault case CurveFault-B, InversionNet with ℓ1 gives MAE 0.1646 and SSIM 0.6163, and UPFWI drops to SSIM 0.3941.
On Kimberlina-CO2, the paper reports InversionNet with ℓ1 at MAE 0.0061, RMSE 0.0374, and SSIM 0.9872, and VelocityGAN at MAE 0.0122, RMSE 0.0574, and SSIM 0.9716 with ℓ1, and MAE 0.0119, RMSE 0.0387, and SSIM 0.9527 with ℓ2. UPFWI was not evaluated on this dataset because of its computational cost.
In three dimensions, InversionNet3D reports on 3D Kimberlina-V1 over three selected channel groups, with MAE of 0.0111, 0.0108, and 0.0161 and corresponding SSIM of 0.9773, 0.9776, and 0.9356.
Training cost in the paper is measured on NVIDIA Tesla P100 GPUs. InversionNet needs 2 hours for the Vel family, 4 hours for the Fault family, 5.5 hours for Style, 3.5 hours for Kimberlina-CO2, and 5.5 hours for the 3D set. VelocityGAN needs 8.6, 16, and 30 hours on the first three families and 32 hours on Kimberlina-CO2, with the 3D set not applicable. UPFWI needs 30, 60, and 60 hours, with the last two datasets not evaluated. UPFWI and InversionNet3D occupy 32 GPUs during training while the other methods use a single GPU.
The generalization study runs pairwise tests across the ten datasets of the Vel, Fault, and Style families, taking the best VelocityGAN model trained on each dataset and testing it on the other nine, producing a 10 by 10 SSIM matrix. Within a family, entries in the lower triangle are consistently larger than those in the upper triangle, so generalizing from a harder dataset to an easier one is more feasible than the reverse. Across families, with the source fixed, generalization degrades as target complexity rises; A-version nodes always have higher degree than B-version nodes, and Style-B has neither incoming nor outgoing edges to other families, making it a difficult generalization case.
For uncertainty quantification, a case study on CurveVel-A with InversionNet shows higher uncertainty at boundaries than elsewhere, implying that boundary predictions are more sensitive. Uncertainty rises gradually as noise level increases, and the uncertainty of cross-dataset use is substantially higher than same-dataset training and testing, indicating that domain shift inflates uncertainty.
7. Role in Physics-AI Workflows
As a supervised regression benchmark, OpenFWI offers paired seismic data and velocity maps in the standard supervised form, so any downstream network can be attached directly. This complements the representation-learning route collected under physics foundation models: one relies on large-scale synthetic pretraining, the other on rigorous in-domain evaluation.
As a complexity covariate, the negative relation between three complexity metrics and SSIM yields an actionable rule. Estimate the complexity of the target velocity map first, then judge whether a given baseline's error magnitude is acceptable. That is closer to engineering judgment than reading an average accuracy.
As a generalization stress field, the ten-dataset pairwise matrix supplies a ready-made protocol. Field data differs from training data in distribution, so cross-domain generalization decides whether a method survives deployment, and OpenFWI turns that hard problem into a reproducible benchmark.
As a testbed for physics-guided versus purely data-driven methods, UPFWI represents the route that wires differentiable forward modeling into the training loop and is directly comparable with the supervised baselines on the same data. Pairing it with industrial solvers collected under numerics and solvers allows the accuracy and cost trade-off of hybrid schemes to be assessed.
As a source of uncertainty, the InversionNet visualization shows that boundary regions deserve extra scrutiny, which resonates with the ill-posed inversion questions framed in physics-informed neural network primers: both point toward constraining data-driven inversion with physical priors.
As a bridge to 3D and realistic benchmarks, 3D Kimberlina-V1 is the first large-scale public 3D FWI dataset, generated by multiple institutions and released under the support of the U.S. Department of Energy SMART Initiative, and it opens the door to 3D technique research. The large-model route that followed transfers a model trained purely on 2D synthetic data in a zero-shot manner to Marmousi, 2D SEG/EAGE Salt and Overthrust, 2004 BP, Sigsbee, and SEAM Phase I, and OpenFWI is one of the training data sources behind that result.
8. Limitations and Failure Modes
Cross-domain generalization failure is the primary risk. One of the paper's central findings is that data-driven FWI generalizes poorly and that degradation correlates with subsurface structural complexity. The ten-dataset pairwise matrix shows that applying a model trained on one structural family to a structurally different scene can push SSIM below 0.6, and that Style-B, a high-complexity case, is nearly unconnected to the other families. Deploying a model trained on synthetic data directly onto field data should be expected to degrade significantly.
Accuracy ceilings appear on complex structures. Across CurveVel-B, CurveFault-B, and Style-B, the three baselines generally land between 0.6 and 0.76 in SSIM, and UPFWI falls to 0.3941 on CurveFault-B. For reservoir description or velocity model updating that demands high-fidelity imaging, these baselines fall well short.
Noise sensitivity is documented by a robustness experiment: training on clean data and testing on noisy seismic data degrades as noise grows, and InversionNet is the most sensitive of the models. Real acquisitions usually have lower signal-to-noise than synthetic data, so this compounds the transfer risk.
Computational cost is unbalanced. UPFWI and InversionNet3D occupy 32 GPUs and take markedly longer than the supervised baselines, and the paper lists computational efficiency among the four open challenges. The 3D route is also bounded by memory, with InversionNet3D controlling cost by downsampling channels, which the authors note may cost information.
Coverage gaps exist. UPFWI reports nothing on Kimberlina-CO2 or the 3D datasets, and passive seismic work appears only as a preliminary event-picking test in the supplementary material. Scenarios such as earthquake early warning, which depend on passive sources or accumulation over many events, are not directly supported by this benchmark.
Data volume sets a hard floor. A single set starts at 43 GB and the 3D dataset reaches 1.4 TB, and full training requires substantial storage and memory. FWIDataset preloads samples into memory when preload=True, so capacity planning must account for that.
Licensing constraints apply. The CC BY-NC-SA 4.0 data terms, covering non-commercial use and share-alike, directly constrain commercial data pipeline design.
The environment is dated. The repository asks for PyTorch 1.7.1 and torchvision 0.8.2 to align exactly, a combination from 2021 that needs extra compatibility handling on newer CUDA and driver stacks.
9. Version and Source Notes
The repository is lanl/OpenFWI, created on 4 August 2022 with the most recent push on 7 April 2025, default branch main, primary language Python, not archived. The GitHub API reports BSD 3-Clause, consistent with the LICENSE file whose copyright line reads Copyright (c) 2022, Los Alamos National Laboratory. The README opens with a paraphrase of the same three-clause BSD terms. The data license is not in the code repository and is taken from the CC BY-NC-SA 4.0 statement on the project homepage data page.
The paper carries arXiv identifier 2111.02926, first published 4 November 2021, currently at v6 and updated 24 June 2023, with nine authors in the order Chengyuan Deng, Shihang Feng, Hanchen Wang, Xitong Zhang, Peng Jin, Yinan Feng, Qili Zeng, Yinpeng Chen, and Youzuo Lin; the author comment records acceptance to the NeurIPS 2022 Datasets and Benchmarks track. The BibTeX on the documentation site uses the title OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion, while the repository README cites OpenFWI: Benchmark Seismic Datasets for Machine Learning-Based Full Waveform Inversion. The two titles differ, and the official arXiv title is the one to use when citing.
Baseline counts and code availability answer different questions and should not be merged. The project consistently says four deep learning methods, while the README documents full support for InversionNet and VelocityGAN across ten datasets, with UPFWI and InversionNet3D pending approval.
Dataset counts also carry a time dimension. The data page lists twelve datasets, eleven 2D plus one 3D. The update log records the sequence: on 13 June 2022 the first official release covered eleven 2D datasets, on 15 August 2022 code and pretrained models were released, on 17 September 2022 the paper was accepted to NeurIPS, on 16 November 2022 a benchmark update corrected some errors and aligned with the pretrained model results, and on 22 June 2023 3D Kimberlina-V1 was released on EDX. The phrases twelve datasets and eleven 2D datasets therefore refer to different points in time.
Official benchmark values differ slightly between sources. On the homepage benchmark page, UPFWI SSIM on FlatVel-A is recorded as 0.9563 while the paper table gives 0.9565, and the homepage InversionNet 2D table omits the Kimberlina-CO2 row while the paper table reports MAE 0.0061, RMSE 0.0374, and SSIM 0.9872 for it. Any specific number should state whether it comes from the homepage benchmark page or the paper table.
The large-model baseline belongs to a later time. The benchmark page adds a 2D Benchmark by Large Model (with Latent GD) group citing a 2026 paper, with an MAE mean of 0.0136 across all ten 2D datasets, including 0.0043 on FlatVel-A, 0.0169 on CurveVel-B, and 0.0282 on Style-B. These numbers belong to a 2026 addition and are distinct from the four original baselines in the paper.
Training time carries a hardware premise. The table in the paper is based on NVIDIA Tesla P100 and distinguishes single-GPU from 32-GPU configurations, so durations scale with hardware and the card count must be preserved when comparing methods.
Team composition appears on the team page, listing nine members: five from LANL plus collaborators from Rutgers University, Michigan State University, the Pennsylvania State University, and Microsoft Research, with Qili Zeng now at Qualcomm. The contact address is openfwi@gmail.com, and a Google group receives updates.
10. References
- Deng C., Feng S., Wang H., Zhang X., Jin P., Feng Y., Zeng Q., Chen Y., Lin Y. OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion. arXiv:2111.02926 (2021, v6 updated 2023-06-24).
- OpenFWI project homepage and data page. https://openfwi-lanl.github.io/ and https://openfwi-lanl.github.io/docs/data.html
- lanl/OpenFWI code repository. https://github.com/lanl/OpenFWI
- OpenFWI official benchmark results page. https://openfwi-lanl.github.io/docs/benchmark.html
- OpenFWI update log page. https://openfwi-lanl.github.io/docs/update.html
- OpenFWI papers and dataset list page. https://openfwi-lanl.github.io/docs/pub.html
- OpenFWI team page. https://openfwi-lanl.github.io/docs/team.html
- OpenFWI tutorial and Colab demo. https://openfwi-lanl.github.io/tutorial/
- Wu Y., Lin Y. InversionNet: A Real-Time and Accurate Full Waveform Inversion with CNNs and continuous CRFs. arXiv:1811.07875 (2018).
- Zhang Z., Wu Y., Zhou Z., Lin Y. VelocityGAN: Subsurface Velocity Image Estimation Using Conditional Adversarial Networks.
- Zeng Q., Feng S., Wohlberg B., Lin Y. InversionNet3D: Efficient and Scalable Learning for 3D Full Waveform Inversion. arXiv:2103.14158 (2021).
- Jin P., Zhang X., Chen Y., Huang S. X., Liu Z., Lin Y. Unsupervised Learning of Full-Waveform Inversion: Connecting CNN and Partial Differential Equation in a Loop. arXiv:2110.07584 (2021).
- Feng S., Lin Y., Wohlberg B. Multiscale Data-driven Seismic Full-waveform Inversion with Field Data Study. arXiv:2103.04007 (2021).
- Feng Y., Chen Y., Jin P., Liu Z., Lin Y. An Intriguing Property of Geophysics Inversion.
- Lin Y., Theiler J., Wohlberg B. Physics-Guided Data-Driven Seismic Inversion: Recent Progress and Future Opportunities in Full Waveform Inversion.
- Feng Y., Jin P., Guo Y., Chen Y., Lin Y. Improving Full Waveform Inversion in Large Model Era. arXiv:2603.00377 (2026).
- Feng S., Jin P., Wang H., Zhang X., Chen Y., Lin Y. Unsupervised Learning of Full-Waveform Inversion related work and the OpenFWI evaluation protocol.
- Rojas-Gómez R., Yang J., Lin Y., Theiler J., Wohlberg B. Physics-Consistent Data-Driven Waveform Inversion With Adaptive Data Augmentation.
dataset_config.jsonin the lanl/OpenFWI repository, per-dataset generation and normalization parameters.split_files/directory in the lanl/OpenFWI repository, 36 training and validation sample manifests.- Po-Hsun Su. pytorch-ssim, SSIM metric implementation borrowed by OpenFWI.
- U.S. Department of Energy SMART Initiative, generation and release support for 3D Kimberlina-V1.