{
  "id": 584426,
  "title": "anyone try to use Physics Driven Neural Network for Interval",
  "url": "/competitions/waveform-inversion/discussion/584426",
  "author_name": "Bill Fan",
  "post_date": "2025-06-13T08:52:37.759000",
  "votes": -1,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I find it difficult to improve the model's performance when using a U-Net. I have studied some papers on Physics-Driven Neural Networks for seismic inversion, but I struggle due to my limited knowledge of Geophysical Waveform Inversion.  this is my note <a href=\"https://www.kaggle.com/code/billfan88/physics-driven-neural-network-for-interval\" target=\"_blank\">https://www.kaggle.com/code/billfan88/physics-driven-neural-network-for-interval</a>. most code is generated by AI 😂</p>",
  "messages": [
    {
      "id": 3224262,
      "postDate": "2025-06-14T15:35:42.790Z",
      "content": "<p>here is a shortcut:<br>\n1) train set =(x,y). Train a model m, e.g, unet from public code<br>\n2) evaluate on kaggle test m(xtest) = ypredict<br>\n3) since we have code of forward wave modeling (check the public code or the matlab code from paper), we can:<br>\nxpredict = forwardwave(ypredict)<br>\n4) fine-tune  train set =(x,y) + (xpredict, ypredict)<br>\n5) repeat until converge </p>\n<hr>\n<p>how it works:<br>\nour target is (xtest, ytest). we are creating their very close neighbors  (xpredict, predict) for training.<br>\nit works only if the forward wave modeling is correct and test/train samples are \"quite close\", else results will diverge.</p>\n<p>you can treat this as creating test augmentation data.</p>\n<p>another analogy is sentence pair samples in language translation model where forward-backward translation is used as augmentation</p>\n<hr>\n<p>on a side note:</p>\n<ul>\n<li>how to create augmentation from train set.</li>\n<li>apply scale, rotate, shift, etc on velocity images.<br>\n(or goggle for github openFWI velocity GAN generator)</li>\n<li>use  forward wave modeling on augmented velocity images to get augmented semsic images</li>\n</ul>",
      "rawMarkdown": "here is a shortcut:\n1) train set =(x,y). Train a model m, e.g, unet from public code\n2) evaluate on kaggle test m(xtest) = ypredict\n3) since we have code of forward wave modeling (check the public code or the matlab code from paper), we can:\nxpredict = forwardwave(ypredict)\n4) fine-tune  train set =(x,y) + (xpredict, ypredict)\n5) repeat until converge \n\n-----\nhow it works:\nour target is (xtest, ytest). we are creating their very close neighbors  (xpredict, predict) for training.\nit works only if the forward wave modeling is correct and test/train samples are \"quite close\", else results will diverge.\n\nyou can treat this as creating test augmentation data.\n\nanother analogy is sentence pair samples in language translation model where forward-backward translation is used as augmentation\n\n\n----\n\non a side note:\n- how to create augmentation from train set.\n- apply scale, rotate, shift, etc on velocity images.\n(or goggle for github openFWI velocity GAN generator)\n- use  forward wave modeling on augmented velocity images to get augmented semsic images\n",
      "votes": 2,
      "replies": [
        {
          "id": 3224281,
          "postDate": "2025-06-14T16:14:10.100Z",
          "content": "<p>You tried this shortcut or just speculate?</p>",
          "rawMarkdown": "You tried this shortcut or just speculate?",
          "replies": [
            {
              "id": 3227358,
              "postDate": "2025-06-18T22:26:07.090Z",
              "content": "<p><a href=\"https://www.kaggle.com/shlomoron\" target=\"_blank\">@shlomoron</a> did you find the way perfectly reconstruct seis from velocity map? That’s the key and I don’t know why host didn’t open their python code, if their python code produce the same with their matlab code , what”s the  detailed parameters they use?</p>",
              "rawMarkdown": "@shlomoron did you find the way perfectly reconstruct seis from velocity map? That’s the key and I don’t know why host didn’t open their python code, if their python code produce the same with their matlab code , what”s the  detailed parameters they use?"
            }
          ]
        },
        {
          "id": 3237026,
          "postDate": "2025-06-30T20:30:16.053Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 3223405,
      "postDate": "2025-06-13T08:52:37.760Z",
      "content": "<p>I find it difficult to improve the model's performance when using a U-Net. I have studied some papers on Physics-Driven Neural Networks for seismic inversion, but I struggle due to my limited knowledge of Geophysical Waveform Inversion.  this is my note <a href=\"https://www.kaggle.com/code/billfan88/physics-driven-neural-network-for-interval\" target=\"_blank\">https://www.kaggle.com/code/billfan88/physics-driven-neural-network-for-interval</a>. most code is generated by AI 😂</p>",
      "rawMarkdown": "I find it difficult to improve the model's performance when using a U-Net. I have studied some papers on Physics-Driven Neural Networks for seismic inversion, but I struggle due to my limited knowledge of Geophysical Waveform Inversion.  this is my note https://www.kaggle.com/code/billfan88/physics-driven-neural-network-for-interval. most code is generated by AI 😂",
      "votes": -1
    }
  ],
  "comments": [
    {
      "id": 3224262,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "2025-06-14T15:35:42.790000",
      "content": "<p>here is a shortcut:<br>\n1) train set =(x,y). Train a model m, e.g, unet from public code<br>\n2) evaluate on kaggle test m(xtest) = ypredict<br>\n3) since we have code of forward wave modeling (check the public code or the matlab code from paper), we can:<br>\nxpredict = forwardwave(ypredict)<br>\n4) fine-tune  train set =(x,y) + (xpredict, ypredict)<br>\n5) repeat until converge </p>\n<hr>\n<p>how it works:<br>\nour target is (xtest, ytest). we are creating their very close neighbors  (xpredict, predict) for training.<br>\nit works only if the forward wave modeling is correct and test/train samples are \"quite close\", else results will diverge.</p>\n<p>you can treat this as creating test augmentation data.</p>\n<p>another analogy is sentence pair samples in language translation model where forward-backward translation is used as augmentation</p>\n<hr>\n<p>on a side note:</p>\n<ul>\n<li>how to create augmentation from train set.</li>\n<li>apply scale, rotate, shift, etc on velocity images.<br>\n(or goggle for github openFWI velocity GAN generator)</li>\n<li>use  forward wave modeling on augmented velocity images to get augmented semsic images</li>\n</ul>",
      "votes": 2,
      "replies": [
        {
          "id": 3224281,
          "author_name": "greySnow",
          "author_url": "",
          "post_date": "2025-06-14T16:14:10.100000",
          "content": "<p>You tried this shortcut or just speculate?</p>",
          "votes": 0,
          "replies": [
            {
              "id": 3227358,
              "author_name": "lhwcv",
              "author_url": "",
              "post_date": "2025-06-18T22:26:07.090000",
              "content": "<p><a href=\"https://www.kaggle.com/shlomoron\" target=\"_blank\">@shlomoron</a> did you find the way perfectly reconstruct seis from velocity map? That’s the key and I don’t know why host didn’t open their python code, if their python code produce the same with their matlab code , what”s the  detailed parameters they use?</p>",
              "votes": 0,
              "replies": []
            }
          ]
        },
        {
          "id": 3237026,
          "author_name": "",
          "author_url": "",
          "post_date": "2025-06-30T20:30:16.053000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3224262": "here is a shortcut:\n1) train set =(x,y). Train a model m, e.g, unet from public code\n2) evaluate on kaggle test m(xtest) = ypredict\n3) since we have code of forward wave modeling (check the public code or the matlab code from paper), we can:\nxpredict = forwardwave(ypredict)\n4) fine-tune  train set =(x,y) + (xpredict, ypredict)\n5) repeat until converge \n\n-----\nhow it works:\nour target is (xtest, ytest). we are creating their very close neighbors  (xpredict, predict) for training.\nit works only if the forward wave modeling is correct and test/train samples are \"quite close\", else results will diverge.\n\nyou can treat this as creating test augmentation data.\n\nanother analogy is sentence pair samples in language translation model where forward-backward translation is used as augmentation\n\n\n----\n\non a side note:\n- how to create augmentation from train set.\n- apply scale, rotate, shift, etc on velocity images.\n(or goggle for github openFWI velocity GAN generator)\n- use  forward wave modeling on augmented velocity images to get augmented semsic images\n",
    "3223405": "I find it difficult to improve the model's performance when using a U-Net. I have studied some papers on Physics-Driven Neural Networks for seismic inversion, but I struggle due to my limited knowledge of Geophysical Waveform Inversion.  this is my note https://www.kaggle.com/code/billfan88/physics-driven-neural-network-for-interval. most code is generated by AI 😂"
  }
}