{
  "id": 572399,
  "title": "[LB 249.7] OpenFWI Models  -- [LB  130.5] Ensemble",
  "url": "/competitions/waveform-inversion/discussion/572399",
  "author_name": "",
  "post_date": "2025-04-09T10:51:27.209759800Z",
  "votes": 11,
  "comment_count": 5,
  "views": 0,
  "content": "<h1><a href=\"https://arxiv.org/abs/2111.02926\" target=\"_blank\">Paper - OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion</a></h1>\n<blockquote>\n  <p><strong>Full waveform inversion (FWI)</strong> is widely used in <strong>geophysics</strong> to reconstruct <strong>high-resolution velocity maps</strong> from <strong>seismic data</strong>. The recent success of <strong>data-driven FWI methods</strong> results in a rapidly increasing demand for <strong>open datasets</strong> to serve the geophysics community. We present <strong>OpenFWI</strong>, a collection of <strong>large-scale multi-structural benchmark datasets</strong>, to facilitate <strong>diversified, rigorous, and reproducible research</strong> on FWI.</p>\n  <p>In particular, <strong>OpenFWI</strong> consists of <strong>12 datasets (2.1TB in total)</strong> synthesized from multiple sources. It encompasses <strong>diverse domains in geophysics</strong> (<strong>interface</strong>, <strong>fault</strong>, <strong>CO₂ reservoir</strong>, etc.), covers different <strong>geological subsurface structures</strong> (<strong>flat</strong>, <strong>curve</strong>, etc.), and contains various amounts of <strong>data samples (2K - 67K)</strong>. It also includes a dataset for <strong>3D FWI</strong>.</p>\n  <p>Moreover, we use <strong>OpenFWI</strong> to perform <strong>benchmarking</strong> over <strong>four deep learning methods</strong>, covering both <strong>supervised</strong> and <strong>unsupervised learning</strong> regimes. Along with the benchmarks, we implement additional experiments, including <strong>physics-driven methods</strong>, <strong>complexity analysis</strong>, <strong>generalization study</strong>, <strong>uncertainty quantification</strong>, and so on, to sharpen our understanding of <strong>datasets and methods</strong>. The studies either provide <strong>valuable insights</strong> into the datasets and the performance, or uncover their <strong>current limitations</strong>.</p>\n  <p>We hope <strong>OpenFWI</strong> supports <strong>prospective research on FWI</strong> and inspires future <strong>open-source efforts on AI for science</strong>.</p>\n</blockquote>\n<hr>\n<blockquote>\n  <h1><a href=\"https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models\" target=\"_blank\">OpenFWI Models - Kaggle Dataset</a></h1>\n  <h1><a href=\"https://www.kaggle.com/datasets/seshurajup/waveform-inversion-exps\" target=\"_blank\">OpenFWI Model outputs - Kaggle Dataset</a></h1>\n  <h1><a href=\"https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698\" target=\"_blank\">OpenFWI Model submissions - Kaggle Notebook</a></h1>\n  <table>\n  <thead>\n  <tr>\n  <th>Model Type</th>\n  <th>Model Name</th>\n  <th>Loss</th>\n  <th>LB Score</th>\n  <th>Local CV Score</th>\n  <th>Dataset</th>\n  <th>MAE</th>\n  <th>RMSE</th>\n  <th>SSIM</th>\n  </tr>\n  </thead>\n  <tbody>\n  <tr>\n  <td><strong>Ensemble</strong></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  </tr>\n  <tr>\n  <td>Ensemble</td>\n  <td>Best Models</td>\n  <td>-</td>\n  <td><strong>130.5</strong></td>\n  <td></td>\n  <td>All</td>\n  <td>-</td>\n  <td>-</td>\n  <td>-</td>\n  </tr>\n  <tr>\n  <td><strong>InversionNet / VelocityGAN</strong></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>249.7</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.0946</td>\n  <td>0.1553</td>\n  <td>0.7552</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>258.3</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1268</td>\n  <td>0.2618</td>\n  <td>0.7111</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>259.1</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.0925</td>\n  <td>0.1600</td>\n  <td>0.7476</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffb_l2</td>\n  <td>L2</td>\n  <td><strong>261.0</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.1106</td>\n  <td>0.1723</td>\n  <td>0.7186</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffb_l1</td>\n  <td>L1</td>\n  <td><strong>262.8</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.1055</td>\n  <td>0.1741</td>\n  <td>0.7208</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cvb_l2</td>\n  <td>L2</td>\n  <td><strong>269.1</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1624</td>\n  <td>0.2801</td>\n  <td>0.6661</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>272.0</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1583</td>\n  <td>0.2336</td>\n  <td>0.6033</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cvb_l1</td>\n  <td>L1</td>\n  <td><strong>272.6</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1497</td>\n  <td>0.2891</td>\n  <td>0.6727</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>278.8</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1428</td>\n  <td>0.2611</td>\n  <td>0.6962</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>279.4</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1571</td>\n  <td>0.2427</td>\n  <td>0.5996</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfb_l2</td>\n  <td>L2</td>\n  <td><strong>285.8</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1669</td>\n  <td>0.2412</td>\n  <td>0.6053</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfb_l1</td>\n  <td>L1</td>\n  <td><strong>288.5</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1646</td>\n  <td>0.2477</td>\n  <td>0.6163</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>304.8</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0510</td>\n  <td>0.0976</td>\n  <td>0.8758</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cva_l2</td>\n  <td>L2</td>\n  <td><strong>308.4</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0690</td>\n  <td>0.1022</td>\n  <td>0.8223</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>310.4</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0482</td>\n  <td>0.1034</td>\n  <td>0.8624</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>319.4</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0216</td>\n  <td>0.0505</td>\n  <td>0.9687</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cva_l1</td>\n  <td>L1</td>\n  <td><strong>324.6</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0685</td>\n  <td>0.1273</td>\n  <td>0.8074</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>328.3</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0319</td>\n  <td>0.0531</td>\n  <td>0.9798</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>338.8</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0258</td>\n  <td>0.0606</td>\n  <td>0.9613</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfa_l2</td>\n  <td>L2</td>\n  <td><strong>340.2</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0280</td>\n  <td>0.0602</td>\n  <td>0.9592</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fvb_l2</td>\n  <td>L2</td>\n  <td><strong>348.1</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0417</td>\n  <td>0.0909</td>\n  <td>0.9402</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fvb_l1</td>\n  <td>L1</td>\n  <td><strong>359.0</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0351</td>\n  <td>0.0876</td>\n  <td>0.9461</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffa_l2</td>\n  <td>L2</td>\n  <td><strong>364.5</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0174</td>\n  <td>0.0362</td>\n  <td>0.9798</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffa_l1</td>\n  <td>L1</td>\n  <td><strong>366.2</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0172</td>\n  <td>0.0426</td>\n  <td>0.9766</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfa_l1</td>\n  <td>L1</td>\n  <td><strong>378.7</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0260</td>\n  <td>0.0650</td>\n  <td>0.9566</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fva_l1</td>\n  <td>L1</td>\n  <td><strong>379.3</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0131</td>\n  <td>0.0211</td>\n  <td>0.9895</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>381.5</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0612</td>\n  <td>0.1000</td>\n  <td>0.8883</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fva_l2</td>\n  <td>L2</td>\n  <td><strong>387.9</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0111</td>\n  <td>0.0180</td>\n  <td>0.9887</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>390.2</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0868</td>\n  <td>0.1485</td>\n  <td>0.9313</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>sta_l2</td>\n  <td>L2</td>\n  <td><strong>391.5</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0610</td>\n  <td>0.0989</td>\n  <td>0.8910</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>391.7</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0645</td>\n  <td>0.1025</td>\n  <td>0.8882</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>399.0</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0329</td>\n  <td>0.0807</td>\n  <td>0.9524</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>stb_l2</td>\n  <td>L2</td>\n  <td><strong>405.1</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0586</td>\n  <td>0.0893</td>\n  <td>0.7599</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>sta_l1_new</td>\n  <td>L1</td>\n  <td><strong>408.7</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0625</td>\n  <td>0.1024</td>\n  <td>0.8859</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>414.2</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0697</td>\n  <td>0.1108</td>\n  <td>0.6953</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>418.7</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0649</td>\n  <td>0.0979</td>\n  <td>0.7249</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>stb_l1</td>\n  <td>L1</td>\n  <td><strong>429.7</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0689</td>\n  <td>0.1614</td>\n  <td>0.6314</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>459.7</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0118</td>\n  <td>0.0178</td>\n  <td>0.9916</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>461.2</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0328</td>\n  <td>0.0787</td>\n  <td>0.9556</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>474.8</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0065</td>\n  <td>0.0783</td>\n  <td>0.9453</td>\n  </tr>\n  </tbody>\n  </table>\n</blockquote>\n<hr>\n<blockquote>\n  <h1>What is best way to build CV? or author CV split choice ? <a href=\"https://github.com/lanl/OpenFWI/tree/48754806b7b4c5877259c6b958a87f4513fcdc0b/split_files\" target=\"_blank\">CV split files</a></h1>\n</blockquote>\n<hr>\n<blockquote>\n  <h1>Correlation</h1>\n  <p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F761268%2Fee75964c6fd3ee5443d26c876d4dd079%2FScreenshot%202025-04-09%20at%204.46.21PM.png?generation=1744197431043341&amp;alt=media\" alt=\"\"></p>\n</blockquote>",
  "messages": [
    {
      "id": "3174682",
      "postDate": "04/09/2025 10:51:27",
      "content": "<h1><a href=\"https://arxiv.org/abs/2111.02926\" target=\"_blank\">Paper - OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion</a></h1>\n<blockquote>\n  <p><strong>Full waveform inversion (FWI)</strong> is widely used in <strong>geophysics</strong> to reconstruct <strong>high-resolution velocity maps</strong> from <strong>seismic data</strong>. The recent success of <strong>data-driven FWI methods</strong> results in a rapidly increasing demand for <strong>open datasets</strong> to serve the geophysics community. We present <strong>OpenFWI</strong>, a collection of <strong>large-scale multi-structural benchmark datasets</strong>, to facilitate <strong>diversified, rigorous, and reproducible research</strong> on FWI.</p>\n  <p>In particular, <strong>OpenFWI</strong> consists of <strong>12 datasets (2.1TB in total)</strong> synthesized from multiple sources. It encompasses <strong>diverse domains in geophysics</strong> (<strong>interface</strong>, <strong>fault</strong>, <strong>CO₂ reservoir</strong>, etc.), covers different <strong>geological subsurface structures</strong> (<strong>flat</strong>, <strong>curve</strong>, etc.), and contains various amounts of <strong>data samples (2K - 67K)</strong>. It also includes a dataset for <strong>3D FWI</strong>.</p>\n  <p>Moreover, we use <strong>OpenFWI</strong> to perform <strong>benchmarking</strong> over <strong>four deep learning methods</strong>, covering both <strong>supervised</strong> and <strong>unsupervised learning</strong> regimes. Along with the benchmarks, we implement additional experiments, including <strong>physics-driven methods</strong>, <strong>complexity analysis</strong>, <strong>generalization study</strong>, <strong>uncertainty quantification</strong>, and so on, to sharpen our understanding of <strong>datasets and methods</strong>. The studies either provide <strong>valuable insights</strong> into the datasets and the performance, or uncover their <strong>current limitations</strong>.</p>\n  <p>We hope <strong>OpenFWI</strong> supports <strong>prospective research on FWI</strong> and inspires future <strong>open-source efforts on AI for science</strong>.</p>\n</blockquote>\n<hr>\n<blockquote>\n  <h1><a href=\"https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models\" target=\"_blank\">OpenFWI Models - Kaggle Dataset</a></h1>\n  <h1><a href=\"https://www.kaggle.com/datasets/seshurajup/waveform-inversion-exps\" target=\"_blank\">OpenFWI Model outputs - Kaggle Dataset</a></h1>\n  <h1><a href=\"https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698\" target=\"_blank\">OpenFWI Model submissions - Kaggle Notebook</a></h1>\n  <table>\n  <thead>\n  <tr>\n  <th>Model Type</th>\n  <th>Model Name</th>\n  <th>Loss</th>\n  <th>LB Score</th>\n  <th>Local CV Score</th>\n  <th>Dataset</th>\n  <th>MAE</th>\n  <th>RMSE</th>\n  <th>SSIM</th>\n  </tr>\n  </thead>\n  <tbody>\n  <tr>\n  <td><strong>Ensemble</strong></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  </tr>\n  <tr>\n  <td>Ensemble</td>\n  <td>Best Models</td>\n  <td>-</td>\n  <td><strong>130.5</strong></td>\n  <td></td>\n  <td>All</td>\n  <td>-</td>\n  <td>-</td>\n  <td>-</td>\n  </tr>\n  <tr>\n  <td><strong>InversionNet / VelocityGAN</strong></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  <td></td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>249.7</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.0946</td>\n  <td>0.1553</td>\n  <td>0.7552</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>258.3</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1268</td>\n  <td>0.2618</td>\n  <td>0.7111</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>259.1</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.0925</td>\n  <td>0.1600</td>\n  <td>0.7476</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffb_l2</td>\n  <td>L2</td>\n  <td><strong>261.0</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.1106</td>\n  <td>0.1723</td>\n  <td>0.7186</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffb_l1</td>\n  <td>L1</td>\n  <td><strong>262.8</strong></td>\n  <td></td>\n  <td>FlatFault-B</td>\n  <td>0.1055</td>\n  <td>0.1741</td>\n  <td>0.7208</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cvb_l2</td>\n  <td>L2</td>\n  <td><strong>269.1</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1624</td>\n  <td>0.2801</td>\n  <td>0.6661</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>272.0</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1583</td>\n  <td>0.2336</td>\n  <td>0.6033</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cvb_l1</td>\n  <td>L1</td>\n  <td><strong>272.6</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1497</td>\n  <td>0.2891</td>\n  <td>0.6727</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>278.8</strong></td>\n  <td></td>\n  <td>CurveVel-B</td>\n  <td>0.1428</td>\n  <td>0.2611</td>\n  <td>0.6962</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>279.4</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1571</td>\n  <td>0.2427</td>\n  <td>0.5996</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfb_l2</td>\n  <td>L2</td>\n  <td><strong>285.8</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1669</td>\n  <td>0.2412</td>\n  <td>0.6053</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfb_l1</td>\n  <td>L1</td>\n  <td><strong>288.5</strong></td>\n  <td></td>\n  <td>CurveFault-B</td>\n  <td>0.1646</td>\n  <td>0.2477</td>\n  <td>0.6163</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>304.8</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0510</td>\n  <td>0.0976</td>\n  <td>0.8758</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cva_l2</td>\n  <td>L2</td>\n  <td><strong>308.4</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0690</td>\n  <td>0.1022</td>\n  <td>0.8223</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvevel_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>310.4</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0482</td>\n  <td>0.1034</td>\n  <td>0.8624</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>319.4</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0216</td>\n  <td>0.0505</td>\n  <td>0.9687</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cva_l1</td>\n  <td>L1</td>\n  <td><strong>324.6</strong></td>\n  <td></td>\n  <td>CurveVel-A</td>\n  <td>0.0685</td>\n  <td>0.1273</td>\n  <td>0.8074</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>328.3</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0319</td>\n  <td>0.0531</td>\n  <td>0.9798</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>curvefault_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>338.8</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0258</td>\n  <td>0.0606</td>\n  <td>0.9613</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfa_l2</td>\n  <td>L2</td>\n  <td><strong>340.2</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0280</td>\n  <td>0.0602</td>\n  <td>0.9592</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fvb_l2</td>\n  <td>L2</td>\n  <td><strong>348.1</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0417</td>\n  <td>0.0909</td>\n  <td>0.9402</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fvb_l1</td>\n  <td>L1</td>\n  <td><strong>359.0</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0351</td>\n  <td>0.0876</td>\n  <td>0.9461</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffa_l2</td>\n  <td>L2</td>\n  <td><strong>364.5</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0174</td>\n  <td>0.0362</td>\n  <td>0.9798</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>ffa_l1</td>\n  <td>L1</td>\n  <td><strong>366.2</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0172</td>\n  <td>0.0426</td>\n  <td>0.9766</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>cfa_l1</td>\n  <td>L1</td>\n  <td><strong>378.7</strong></td>\n  <td></td>\n  <td>CurveFault-A</td>\n  <td>0.0260</td>\n  <td>0.0650</td>\n  <td>0.9566</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fva_l1</td>\n  <td>L1</td>\n  <td><strong>379.3</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0131</td>\n  <td>0.0211</td>\n  <td>0.9895</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>381.5</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0612</td>\n  <td>0.1000</td>\n  <td>0.8883</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>fva_l2</td>\n  <td>L2</td>\n  <td><strong>387.9</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0111</td>\n  <td>0.0180</td>\n  <td>0.9887</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatfault_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>390.2</strong></td>\n  <td></td>\n  <td>FlatFault-A</td>\n  <td>0.0868</td>\n  <td>0.1485</td>\n  <td>0.9313</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>sta_l2</td>\n  <td>L2</td>\n  <td><strong>391.5</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0610</td>\n  <td>0.0989</td>\n  <td>0.8910</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>391.7</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0645</td>\n  <td>0.1025</td>\n  <td>0.8882</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>399.0</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0329</td>\n  <td>0.0807</td>\n  <td>0.9524</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>stb_l2</td>\n  <td>L2</td>\n  <td><strong>405.1</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0586</td>\n  <td>0.0893</td>\n  <td>0.7599</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>sta_l1_new</td>\n  <td>L1</td>\n  <td><strong>408.7</strong></td>\n  <td></td>\n  <td>Style-A</td>\n  <td>0.0625</td>\n  <td>0.1024</td>\n  <td>0.8859</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_b_l1_480</td>\n  <td>L1</td>\n  <td><strong>414.2</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0697</td>\n  <td>0.1108</td>\n  <td>0.6953</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>style_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>418.7</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0649</td>\n  <td>0.0979</td>\n  <td>0.7249</td>\n  </tr>\n  <tr>\n  <td>InversionNet</td>\n  <td>stb_l1</td>\n  <td>L1</td>\n  <td><strong>429.7</strong></td>\n  <td></td>\n  <td>Style-B</td>\n  <td>0.0689</td>\n  <td>0.1614</td>\n  <td>0.6314</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_a_l1_480</td>\n  <td>L1</td>\n  <td><strong>459.7</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0118</td>\n  <td>0.0178</td>\n  <td>0.9916</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_b_l2_480</td>\n  <td>L2</td>\n  <td><strong>461.2</strong></td>\n  <td></td>\n  <td>FlatVel-B</td>\n  <td>0.0328</td>\n  <td>0.0787</td>\n  <td>0.9556</td>\n  </tr>\n  <tr>\n  <td>VelocityGAN</td>\n  <td>flatvel_a_l2_480</td>\n  <td>L2</td>\n  <td><strong>474.8</strong></td>\n  <td></td>\n  <td>FlatVel-A</td>\n  <td>0.0065</td>\n  <td>0.0783</td>\n  <td>0.9453</td>\n  </tr>\n  </tbody>\n  </table>\n</blockquote>\n<hr>\n<blockquote>\n  <h1>What is best way to build CV? or author CV split choice ? <a href=\"https://github.com/lanl/OpenFWI/tree/48754806b7b4c5877259c6b958a87f4513fcdc0b/split_files\" target=\"_blank\">CV split files</a></h1>\n</blockquote>\n<hr>\n<blockquote>\n  <h1>Correlation</h1>\n  <p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F761268%2Fee75964c6fd3ee5443d26c876d4dd079%2FScreenshot%202025-04-09%20at%204.46.21PM.png?generation=1744197431043341&amp;alt=media\" alt=\"\"></p>\n</blockquote>",
      "rawMarkdown": "# [Paper - OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion](https://arxiv.org/abs/2111.02926)\n> **Full waveform inversion (FWI)** is widely used in **geophysics** to reconstruct **high-resolution velocity maps** from **seismic data**. The recent success of **data-driven FWI methods** results in a rapidly increasing demand for **open datasets** to serve the geophysics community. We present **OpenFWI**, a collection of **large-scale multi-structural benchmark datasets**, to facilitate **diversified, rigorous, and reproducible research** on FWI.\n\n> In particular, **OpenFWI** consists of **12 datasets (2.1TB in total)** synthesized from multiple sources. It encompasses **diverse domains in geophysics** (**interface**, **fault**, **CO₂ reservoir**, etc.), covers different **geological subsurface structures** (**flat**, **curve**, etc.), and contains various amounts of **data samples (2K - 67K)**. It also includes a dataset for **3D FWI**.\n\n> Moreover, we use **OpenFWI** to perform **benchmarking** over **four deep learning methods**, covering both **supervised** and **unsupervised learning** regimes. Along with the benchmarks, we implement additional experiments, including **physics-driven methods**, **complexity analysis**, **generalization study**, **uncertainty quantification**, and so on, to sharpen our understanding of **datasets and methods**. The studies either provide **valuable insights** into the datasets and the performance, or uncover their **current limitations**.\n\n> We hope **OpenFWI** supports **prospective research on FWI** and inspires future **open-source efforts on AI for science**.\n\n---\n> # [OpenFWI Models - Kaggle Dataset](https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models)\n> # [OpenFWI Model outputs - Kaggle Dataset](https://www.kaggle.com/datasets/seshurajup/waveform-inversion-exps)\n> # [OpenFWI Model submissions - Kaggle Notebook](https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698)\n\n> | Model Type   | Model Name             | Loss | LB Score | Local CV Score | Dataset        | MAE | RMSE   | SSIM   |\n|--------------|------------------------|------|----------|----------|----------------|--------|--------|--------|\n|**Ensemble** |\n|  Ensemble  | Best Models     |  -   | **130.5**  |          | All    | - | - | - |\n|**InversionNet / VelocityGAN** |\n| VelocityGAN  | flatfault_b_l2_480     | L2   | **249.7**  |          | FlatFault-B    | 0.0946 | 0.1553 | 0.7552 |\n| VelocityGAN  | curvevel_b_l1_480      | L1   |  **258.3**    |          | CurveVel-B     | 0.1268 | 0.2618 | 0.7111 |\n| VelocityGAN  | flatfault_b_l1_480     | L1   |     **259.1**     |          | FlatFault-B    | 0.0925 | 0.1600 | 0.7476 |\n| InversionNet | ffb_l2                 | L2   |     **261.0**     |          | FlatFault-B    | 0.1106  | 0.1723 | 0.7186 |\n| InversionNet | ffb_l1                 | L1   |      **262.8**    |          | FlatFault-B    | 0.1055 | 0.1741 | 0.7208 |\n| InversionNet | cvb_l2                 | L2   |    **269.1**      |          | CurveVel-B     | 0.1624 | 0.2801 | 0.6661 |\n| VelocityGAN  | curvefault_b_l2_480    | L2   | **272.0** |          | CurveFault-B   | 0.1583 | 0.2336 | 0.6033 |\n| InversionNet | cvb_l1                 | L1   |    **272.6**      |          | CurveVel-B     | 0.1497 | 0.2891 | 0.6727 |\n| VelocityGAN  | curvevel_b_l2_480      | L2   | **278.8** |          | CurveVel-B     | 0.1428 | 0.2611 | 0.6962 |\n| VelocityGAN  | curvefault_b_l1_480    | L1   |  **279.4** |          | CurveFault-B   | 0.1571 | 0.2427 | 0.5996 |\n| InversionNet | cfb_l2                 | L2   |    **285.8**      |          | CurveFault-B   | 0.1669 | 0.2412 | 0.6053 |\n| InversionNet | cfb_l1                 | L1   |    **288.5**      |          | CurveFault-B   | 0.1646 | 0.2477 | 0.6163 |\n| VelocityGAN  | curvevel_a_l2_480      | L2   | **304.8** |          | CurveVel-A     | 0.0510 | 0.0976 | 0.8758 |\n| InversionNet | cva_l2                 | L2   |      **308.4**    |          | CurveVel-A     | 0.0690 | 0.1022 | 0.8223 |\n| VelocityGAN  | curvevel_a_l1_480      | L1   | **310.4** |          | CurveVel-A     | 0.0482 | 0.1034 | 0.8624 |\n| VelocityGAN  | curvefault_a_l2_480    | L2   | **319.4** |          | CurveFault-A   | 0.0216 | 0.0505 | 0.9687 |\n| InversionNet | cva_l1                 | L1   | **324.6**    |          | CurveVel-A     | 0.0685 | 0.1273 | 0.8074 |\n| VelocityGAN  | flatfault_a_l2_480     | L2   | **328.3** |          | FlatFault-A    | 0.0319 | 0.0531 | 0.9798 |\n| VelocityGAN  | curvefault_a_l1_480    | L1   | **338.8** |          | CurveFault-A   | 0.0258 | 0.0606 | 0.9613 |\n| InversionNet | cfa_l2                 | L2   |    **340.2**      |          | CurveFault-A   | 0.0280 | 0.0602 | 0.9592 |\n| InversionNet | fvb_l2                 | L2   |   **348.1**       |          | FlatVel-B      | 0.0417 | 0.0909 | 0.9402 |\n| InversionNet | fvb_l1                 | L1   |     **359.0**     |          | FlatVel-B      | 0.0351 | 0.0876 | 0.9461 |\n| InversionNet | ffa_l2                 | L2   |     **364.5**   |          | FlatFault-A    | 0.0174 | 0.0362 | 0.9798 |\n| InversionNet | ffa_l1                 | L1   |    **366.2**      |          | FlatFault-A    | 0.0172 | 0.0426 | 0.9766 |\n| InversionNet | cfa_l1                 | L1   |    **378.7**      |          | CurveFault-A   | 0.0260 | 0.0650 | 0.9566 |\n| InversionNet | fva_l1                 | L1   | **379.3**  |          | FlatVel-A      | 0.0131 | 0.0211 | 0.9895 |\n| VelocityGAN  | style_a_l1_480         | L1   | **381.5** |          | Style-A        | 0.0612 | 0.1000 | 0.8883 |\n| InversionNet | fva_l2                 | L2   |   **387.9**       |          | FlatVel-A      | 0.0111 | 0.0180 | 0.9887 |\n| VelocityGAN  | flatfault_a_l1_480     | L1   |   **390.2**      |          | FlatFault-A    | 0.0868 | 0.1485 | 0.9313 |\n| InversionNet | sta_l2                 | L2   |   **391.5**       |          | Style-A        | 0.0610 | 0.0989 | 0.8910 |\n| VelocityGAN  | style_a_l2_480         | L2   |   **391.7**   |          | Style-A        | 0.0645 | 0.1025 | 0.8882 |\n| VelocityGAN  | flatvel_b_l1_480       | L1   | **399.0** |          | FlatVel-B      | 0.0329 | 0.0807 | 0.9524 |\n| InversionNet | stb_l2                 | L2   |    **405.1**      |          | Style-B        | 0.0586 | 0.0893 | 0.7599 |\n| InversionNet | sta_l1_new             | L1   |    **408.7**      |          | Style-A        | 0.0625 | 0.1024 | 0.8859 |\n| VelocityGAN  | style_b_l1_480         | L1   | **414.2** |          | Style-B        | 0.0697 | 0.1108 | 0.6953 |\n| VelocityGAN  | style_b_l2_480         | L2   | **418.7** |          | Style-B        | 0.0649 | 0.0979 | 0.7249 |\n| InversionNet | stb_l1                 | L1   |     **429.7**     |          | Style-B        | 0.0689 | 0.1614 | 0.6314 |\n| VelocityGAN  | flatvel_a_l1_480       | L1   | **459.7** |          | FlatVel-A      | 0.0118 | 0.0178 | 0.9916 |\n| VelocityGAN  | flatvel_b_l2_480       | L2   | **461.2** |          | FlatVel-B      | 0.0328 | 0.0787 | 0.9556 |\n| VelocityGAN  | flatvel_a_l2_480       | L2   | **474.8** |          | FlatVel-A      | 0.0065 | 0.0783 | 0.9453 |\n\n---\n\n> # What is best way to build CV? or author CV split choice ? [CV split files](https://github.com/lanl/OpenFWI/tree/48754806b7b4c5877259c6b958a87f4513fcdc0b/split_files)\n\n---\n\n> # Correlation\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F761268%2Fee75964c6fd3ee5443d26c876d4dd079%2FScreenshot%202025-04-09%20at%204.46.21PM.png?generation=1744197431043341&alt=media)",
      "votes": null
    },
    {
      "id": "3175965",
      "postDate": "04/10/2025 19:01:10",
      "content": "<p>Hello,</p>\n<p>You may use your own cross validation split choices. Every file in the same subset should follow a similar data distribution. However, the Fault Family files have different layers that can be reflected by their file names. Please refer to the original OpenFWI paper for naming: <a href=\"https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html\" target=\"_blank\">https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html</a></p>\n<p>Best,<br>\nThe Waveform Inversion Team</p>",
      "rawMarkdown": "Hello,\n\nYou may use your own cross validation split choices. Every file in the same subset should follow a similar data distribution. However, the Fault Family files have different layers that can be reflected by their file names. Please refer to the original OpenFWI paper for naming: https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html\n\nBest,\nThe Waveform Inversion Team",
      "votes": null
    },
    {
      "id": "3178525",
      "postDate": "04/14/2025 09:45:48",
      "content": "<p>Do you manually download all the models and submit them one by one to check their scores on the LB? If yes, is there a more convenient way to do it—like maybe a Kaggle dataset that already includes the models, so you don't have to download everything manually?</p>",
      "rawMarkdown": "Do you manually download all the models and submit them one by one to check their scores on the LB? If yes, is there a more convenient way to do it—like maybe a Kaggle dataset that already includes the models, so you don't have to download everything manually?",
      "votes": null
    },
    {
      "id": "3178537",
      "postDate": "04/14/2025 10:06:48",
      "content": "<p><a href=\"https://www.kaggle.com/karnakbaevarthur\" target=\"_blank\">@karnakbaevarthur</a> in topic, i already shared dataset having all models</p>\n<blockquote>\n  <h1><a href=\"https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models\" target=\"_blank\">OpenFWI Models - Kaggle Dataset</a></h1>\n</blockquote>",
      "rawMarkdown": "karnakbaevarthur in topic, i already shared dataset having all models\n\n> # [OpenFWI Models - Kaggle Dataset](https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models)",
      "votes": null
    },
    {
      "id": "3181999",
      "postDate": "04/18/2025 16:19:13",
      "content": "<table>\n<thead>\n<tr>\n<th><strong>Feature</strong></th>\n<th><strong>Focus</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>Datasets</strong></td>\n<td><strong>CurveVel-A</strong> &gt; <strong>CurveVel-B</strong> &gt; <strong>CurveFault-A</strong> &gt; <strong>FlatVel-A</strong> &gt; <strong>FlatFault-A</strong></td>\n</tr>\n<tr>\n<td><strong>A or B</strong></td>\n<td><strong>B</strong> &gt; <strong>A</strong></td>\n</tr>\n<tr>\n<td><strong>Vel vs Fault</strong></td>\n<td><strong>Fault</strong> &gt; <strong>Vel</strong></td>\n</tr>\n<tr>\n<td><strong>Vel or Flat or Style</strong></td>\n<td><strong>Flat</strong> &gt; <strong>Vel</strong> &gt; <strong>Style</strong></td>\n</tr>\n<tr>\n<td><strong>L2 or L1</strong></td>\n<td><strong>Avg(L2+L1)</strong> (latest papers) &gt; <strong>L2</strong> &gt; <strong>L1</strong></td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "| **Feature**                   | **Focus**                                                                                                                                                     |\n|------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| **Datasets**  | **CurveVel-A** > **CurveVel-B** > **CurveFault-A** > **FlatVel-A** > **FlatFault-A**  |\n| **A or B**               | **B** > **A** |\n| **Vel vs Fault**      | **Fault** > **Vel** |\n| **Vel or Flat or Style** | **Flat** > **Vel** > **Style** |\n| **L2 or L1**          | **Avg(L2+L1)** (latest papers) > **L2** > **L1** |",
      "votes": null
    },
    {
      "id": "3189924",
      "postDate": "04/30/2025 00:52:33",
      "content": "<p><a href=\"https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698\" target=\"_blank\">OpenFWI Model submissions - Kaggle Notebook</a></p>\n<blockquote>\n  <p>LB improved from individual best <strong>249.7</strong> to <strong>130.5</strong></p>\n</blockquote>",
      "rawMarkdown": "[OpenFWI Model submissions - Kaggle Notebook](https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698)\n\n> LB improved from individual best **249.7** to **130.5**",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3175965,
      "author_name": "hanchenwang114",
      "author_url": "",
      "post_date": "04/10/2025 19:01:10",
      "content": "<p>Hello,</p>\n<p>You may use your own cross validation split choices. Every file in the same subset should follow a similar data distribution. However, the Fault Family files have different layers that can be reflected by their file names. Please refer to the original OpenFWI paper for naming: <a href=\"https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html\" target=\"_blank\">https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html</a></p>\n<p>Best,<br>\nThe Waveform Inversion Team</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3178525,
      "author_name": "karnakbaevarthur",
      "author_url": "",
      "post_date": "04/14/2025 09:45:48",
      "content": "<p>Do you manually download all the models and submit them one by one to check their scores on the LB? If yes, is there a more convenient way to do it—like maybe a Kaggle dataset that already includes the models, so you don't have to download everything manually?</p>",
      "votes": null,
      "replies": [
        {
          "id": 3178537,
          "author_name": "seshurajup",
          "author_url": "",
          "post_date": "04/14/2025 10:06:48",
          "content": "<p><a href=\"https://www.kaggle.com/karnakbaevarthur\" target=\"_blank\">@karnakbaevarthur</a> in topic, i already shared dataset having all models</p>\n<blockquote>\n  <h1><a href=\"https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models\" target=\"_blank\">OpenFWI Models - Kaggle Dataset</a></h1>\n</blockquote>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 3181999,
      "author_name": "seshurajup",
      "author_url": "",
      "post_date": "04/18/2025 16:19:13",
      "content": "<table>\n<thead>\n<tr>\n<th><strong>Feature</strong></th>\n<th><strong>Focus</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td><strong>Datasets</strong></td>\n<td><strong>CurveVel-A</strong> &gt; <strong>CurveVel-B</strong> &gt; <strong>CurveFault-A</strong> &gt; <strong>FlatVel-A</strong> &gt; <strong>FlatFault-A</strong></td>\n</tr>\n<tr>\n<td><strong>A or B</strong></td>\n<td><strong>B</strong> &gt; <strong>A</strong></td>\n</tr>\n<tr>\n<td><strong>Vel vs Fault</strong></td>\n<td><strong>Fault</strong> &gt; <strong>Vel</strong></td>\n</tr>\n<tr>\n<td><strong>Vel or Flat or Style</strong></td>\n<td><strong>Flat</strong> &gt; <strong>Vel</strong> &gt; <strong>Style</strong></td>\n</tr>\n<tr>\n<td><strong>L2 or L1</strong></td>\n<td><strong>Avg(L2+L1)</strong> (latest papers) &gt; <strong>L2</strong> &gt; <strong>L1</strong></td>\n</tr>\n</tbody>\n</table>",
      "votes": null,
      "replies": []
    },
    {
      "id": 3189924,
      "author_name": "seshurajup",
      "author_url": "",
      "post_date": "04/30/2025 00:52:33",
      "content": "<p><a href=\"https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698\" target=\"_blank\">OpenFWI Model submissions - Kaggle Notebook</a></p>\n<blockquote>\n  <p>LB improved from individual best <strong>249.7</strong> to <strong>130.5</strong></p>\n</blockquote>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3174682": "# [Paper - OpenFWI: Large-Scale Multi-Structural Benchmark Datasets for Seismic Full Waveform Inversion](https://arxiv.org/abs/2111.02926)\n> **Full waveform inversion (FWI)** is widely used in **geophysics** to reconstruct **high-resolution velocity maps** from **seismic data**. The recent success of **data-driven FWI methods** results in a rapidly increasing demand for **open datasets** to serve the geophysics community. We present **OpenFWI**, a collection of **large-scale multi-structural benchmark datasets**, to facilitate **diversified, rigorous, and reproducible research** on FWI.\n\n> In particular, **OpenFWI** consists of **12 datasets (2.1TB in total)** synthesized from multiple sources. It encompasses **diverse domains in geophysics** (**interface**, **fault**, **CO₂ reservoir**, etc.), covers different **geological subsurface structures** (**flat**, **curve**, etc.), and contains various amounts of **data samples (2K - 67K)**. It also includes a dataset for **3D FWI**.\n\n> Moreover, we use **OpenFWI** to perform **benchmarking** over **four deep learning methods**, covering both **supervised** and **unsupervised learning** regimes. Along with the benchmarks, we implement additional experiments, including **physics-driven methods**, **complexity analysis**, **generalization study**, **uncertainty quantification**, and so on, to sharpen our understanding of **datasets and methods**. The studies either provide **valuable insights** into the datasets and the performance, or uncover their **current limitations**.\n\n> We hope **OpenFWI** supports **prospective research on FWI** and inspires future **open-source efforts on AI for science**.\n\n---\n> # [OpenFWI Models - Kaggle Dataset](https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models)\n> # [OpenFWI Model outputs - Kaggle Dataset](https://www.kaggle.com/datasets/seshurajup/waveform-inversion-exps)\n> # [OpenFWI Model submissions - Kaggle Notebook](https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698)\n\n> | Model Type   | Model Name             | Loss | LB Score | Local CV Score | Dataset        | MAE | RMSE   | SSIM   |\n|--------------|------------------------|------|----------|----------|----------------|--------|--------|--------|\n|**Ensemble** |\n|  Ensemble  | Best Models     |  -   | **130.5**  |          | All    | - | - | - |\n|**InversionNet / VelocityGAN** |\n| VelocityGAN  | flatfault_b_l2_480     | L2   | **249.7**  |          | FlatFault-B    | 0.0946 | 0.1553 | 0.7552 |\n| VelocityGAN  | curvevel_b_l1_480      | L1   |  **258.3**    |          | CurveVel-B     | 0.1268 | 0.2618 | 0.7111 |\n| VelocityGAN  | flatfault_b_l1_480     | L1   |     **259.1**     |          | FlatFault-B    | 0.0925 | 0.1600 | 0.7476 |\n| InversionNet | ffb_l2                 | L2   |     **261.0**     |          | FlatFault-B    | 0.1106  | 0.1723 | 0.7186 |\n| InversionNet | ffb_l1                 | L1   |      **262.8**    |          | FlatFault-B    | 0.1055 | 0.1741 | 0.7208 |\n| InversionNet | cvb_l2                 | L2   |    **269.1**      |          | CurveVel-B     | 0.1624 | 0.2801 | 0.6661 |\n| VelocityGAN  | curvefault_b_l2_480    | L2   | **272.0** |          | CurveFault-B   | 0.1583 | 0.2336 | 0.6033 |\n| InversionNet | cvb_l1                 | L1   |    **272.6**      |          | CurveVel-B     | 0.1497 | 0.2891 | 0.6727 |\n| VelocityGAN  | curvevel_b_l2_480      | L2   | **278.8** |          | CurveVel-B     | 0.1428 | 0.2611 | 0.6962 |\n| VelocityGAN  | curvefault_b_l1_480    | L1   |  **279.4** |          | CurveFault-B   | 0.1571 | 0.2427 | 0.5996 |\n| InversionNet | cfb_l2                 | L2   |    **285.8**      |          | CurveFault-B   | 0.1669 | 0.2412 | 0.6053 |\n| InversionNet | cfb_l1                 | L1   |    **288.5**      |          | CurveFault-B   | 0.1646 | 0.2477 | 0.6163 |\n| VelocityGAN  | curvevel_a_l2_480      | L2   | **304.8** |          | CurveVel-A     | 0.0510 | 0.0976 | 0.8758 |\n| InversionNet | cva_l2                 | L2   |      **308.4**    |          | CurveVel-A     | 0.0690 | 0.1022 | 0.8223 |\n| VelocityGAN  | curvevel_a_l1_480      | L1   | **310.4** |          | CurveVel-A     | 0.0482 | 0.1034 | 0.8624 |\n| VelocityGAN  | curvefault_a_l2_480    | L2   | **319.4** |          | CurveFault-A   | 0.0216 | 0.0505 | 0.9687 |\n| InversionNet | cva_l1                 | L1   | **324.6**    |          | CurveVel-A     | 0.0685 | 0.1273 | 0.8074 |\n| VelocityGAN  | flatfault_a_l2_480     | L2   | **328.3** |          | FlatFault-A    | 0.0319 | 0.0531 | 0.9798 |\n| VelocityGAN  | curvefault_a_l1_480    | L1   | **338.8** |          | CurveFault-A   | 0.0258 | 0.0606 | 0.9613 |\n| InversionNet | cfa_l2                 | L2   |    **340.2**      |          | CurveFault-A   | 0.0280 | 0.0602 | 0.9592 |\n| InversionNet | fvb_l2                 | L2   |   **348.1**       |          | FlatVel-B      | 0.0417 | 0.0909 | 0.9402 |\n| InversionNet | fvb_l1                 | L1   |     **359.0**     |          | FlatVel-B      | 0.0351 | 0.0876 | 0.9461 |\n| InversionNet | ffa_l2                 | L2   |     **364.5**   |          | FlatFault-A    | 0.0174 | 0.0362 | 0.9798 |\n| InversionNet | ffa_l1                 | L1   |    **366.2**      |          | FlatFault-A    | 0.0172 | 0.0426 | 0.9766 |\n| InversionNet | cfa_l1                 | L1   |    **378.7**      |          | CurveFault-A   | 0.0260 | 0.0650 | 0.9566 |\n| InversionNet | fva_l1                 | L1   | **379.3**  |          | FlatVel-A      | 0.0131 | 0.0211 | 0.9895 |\n| VelocityGAN  | style_a_l1_480         | L1   | **381.5** |          | Style-A        | 0.0612 | 0.1000 | 0.8883 |\n| InversionNet | fva_l2                 | L2   |   **387.9**       |          | FlatVel-A      | 0.0111 | 0.0180 | 0.9887 |\n| VelocityGAN  | flatfault_a_l1_480     | L1   |   **390.2**      |          | FlatFault-A    | 0.0868 | 0.1485 | 0.9313 |\n| InversionNet | sta_l2                 | L2   |   **391.5**       |          | Style-A        | 0.0610 | 0.0989 | 0.8910 |\n| VelocityGAN  | style_a_l2_480         | L2   |   **391.7**   |          | Style-A        | 0.0645 | 0.1025 | 0.8882 |\n| VelocityGAN  | flatvel_b_l1_480       | L1   | **399.0** |          | FlatVel-B      | 0.0329 | 0.0807 | 0.9524 |\n| InversionNet | stb_l2                 | L2   |    **405.1**      |          | Style-B        | 0.0586 | 0.0893 | 0.7599 |\n| InversionNet | sta_l1_new             | L1   |    **408.7**      |          | Style-A        | 0.0625 | 0.1024 | 0.8859 |\n| VelocityGAN  | style_b_l1_480         | L1   | **414.2** |          | Style-B        | 0.0697 | 0.1108 | 0.6953 |\n| VelocityGAN  | style_b_l2_480         | L2   | **418.7** |          | Style-B        | 0.0649 | 0.0979 | 0.7249 |\n| InversionNet | stb_l1                 | L1   |     **429.7**     |          | Style-B        | 0.0689 | 0.1614 | 0.6314 |\n| VelocityGAN  | flatvel_a_l1_480       | L1   | **459.7** |          | FlatVel-A      | 0.0118 | 0.0178 | 0.9916 |\n| VelocityGAN  | flatvel_b_l2_480       | L2   | **461.2** |          | FlatVel-B      | 0.0328 | 0.0787 | 0.9556 |\n| VelocityGAN  | flatvel_a_l2_480       | L2   | **474.8** |          | FlatVel-A      | 0.0065 | 0.0783 | 0.9453 |\n\n---\n\n> # What is best way to build CV? or author CV split choice ? [CV split files](https://github.com/lanl/OpenFWI/tree/48754806b7b4c5877259c6b958a87f4513fcdc0b/split_files)\n\n---\n\n> # Correlation\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F761268%2Fee75964c6fd3ee5443d26c876d4dd079%2FScreenshot%202025-04-09%20at%204.46.21PM.png?generation=1744197431043341&alt=media)",
    "3175965": "Hello,\n\nYou may use your own cross validation split choices. Every file in the same subset should follow a similar data distribution. However, the Fault Family files have different layers that can be reflected by their file names. Please refer to the original OpenFWI paper for naming: https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html\n\nBest,\nThe Waveform Inversion Team",
    "3178525": "Do you manually download all the models and submit them one by one to check their scores on the LB? If yes, is there a more convenient way to do it—like maybe a Kaggle dataset that already includes the models, so you don't have to download everything manually?",
    "3178537": "karnakbaevarthur in topic, i already shared dataset having all models\n\n> # [OpenFWI Models - Kaggle Dataset](https://www.kaggle.com/datasets/seshurajup/waveform-inversion-models)",
    "3181999": "| **Feature**                   | **Focus**                                                                                                                                                     |\n|------------------------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| **Datasets**  | **CurveVel-A** > **CurveVel-B** > **CurveFault-A** > **FlatVel-A** > **FlatFault-A**  |\n| **A or B**               | **B** > **A** |\n| **Vel vs Fault**      | **Fault** > **Vel** |\n| **Vel or Flat or Style** | **Flat** > **Vel** > **Style** |\n| **L2 or L1**          | **Avg(L2+L1)** (latest papers) > **L2** > **L1** |",
    "3189924": "[OpenFWI Model submissions - Kaggle Notebook](https://www.kaggle.com/code/seshurajup/waveform-inversion-model-experiments?scriptVersionId=236924698)\n\n> LB improved from individual best **249.7** to **130.5**"
  },
  "source": "meta"
}