In this notebook, I tried to create a loss function for PINN using Deepwave, but the resulting seismis data is not numerically exact. Has someone managed to configure Deepwave correctly, to get numerically accurate results?
", "messages": [ { "id": 3221106, "postDate": "2025-06-10T12:13:03.583Z", "content": "In this notebook, I tried to create a loss function for PINN using Deepwave, but the resulting seismis data is not numerically exact. Has someone managed to configure Deepwave correctly, to get numerically accurate results?
", "rawMarkdown": "In [this notebook](https://www.kaggle.com/code/fpeccia/deepwave-forward-propagation-loss-function), I tried to create a loss function for PINN using Deepwave, but the resulting seismis data is not numerically exact. Has someone managed to configure Deepwave correctly, to get numerically accurate results?", "votes": 2 }, { "id": 3221872, "postDate": "2025-06-11T14:49:31.877Z", "content": "After experimenting with various improvements to Deepwave's seismic wave data inference, I found none to be effective. In my tests, the model exhibits errors of approximately 0.08 MAE and 0.03 MSE.
\nWhile Deepwave's fast inference is appealing, it's not sufficiently reliable. As a result, my current approach is to either use https://www.kaggle.com/code/manatoyo/improved-vel-to-seis (though it's slow and can't perform backpropagation, it's quite accurate) or employ Deepwave itself (allowing faster inference at the cost of reduced accuracy). Another option is to implement the notebook's code in PyTorch (which runs faster than the notebook implementation and achieves remarkably good accuracy—about 0.00001 MAE—but remains several orders of magnitude slower than Deepwave).
I think the gradient flow from reconstruction of the seismic data through this forward modeling eqn will me much more informative than raw mae.
\nby the way how are you guys using this forward modelling ??
I think there are two usage of backpropagation-able forward modelling.
\nThe first one is, to use it as a reconstruction loss function when you train a inverse model. The second one is, to use it when you refine your predicted velocity map especially for test data(the true velocity map is unknown).
Actually, I was going for the second option. In the last days I managed to get a code working to actually refine my predictions, and although it works well on the validation data, I dont see any changes when pushing a new submission. I can't seem to break the 28.2 barrier no matters what. Finetuning helps on local validation score, but doesnt change LB. Refining with wave propagation does the same…. I don't know what else to try.
", "rawMarkdown": "Actually, I was going for the second option. In the last days I managed to get a code working to actually refine my predictions, and although it works well on the validation data, I dont see any changes when pushing a new submission. I can't seem to break the 28.2 barrier no matters what. Finetuning helps on local validation score, but doesnt change LB. Refining with wave propagation does the same.... I don't know what else to try." } ] } ] } ], "comments": [ { "id": 3221872, "author_name": "Harui-ig", "author_url": "", "post_date": "2025-06-11T14:49:31.877000", "content": "After experimenting with various improvements to Deepwave's seismic wave data inference, I found none to be effective. In my tests, the model exhibits errors of approximately 0.08 MAE and 0.03 MSE.
\nWhile Deepwave's fast inference is appealing, it's not sufficiently reliable. As a result, my current approach is to either use https://www.kaggle.com/code/manatoyo/improved-vel-to-seis (though it's slow and can't perform backpropagation, it's quite accurate) or employ Deepwave itself (allowing faster inference at the cost of reduced accuracy). Another option is to implement the notebook's code in PyTorch (which runs faster than the notebook implementation and achieves remarkably good accuracy—about 0.00001 MAE—but remains several orders of magnitude slower than Deepwave).
I think the gradient flow from reconstruction of the seismic data through this forward modeling eqn will me much more informative than raw mae.
\nby the way how are you guys using this forward modelling ??
I think there are two usage of backpropagation-able forward modelling.
\nThe first one is, to use it as a reconstruction loss function when you train a inverse model. The second one is, to use it when you refine your predicted velocity map especially for test data(the true velocity map is unknown).
Actually, I was going for the second option. In the last days I managed to get a code working to actually refine my predictions, and although it works well on the validation data, I dont see any changes when pushing a new submission. I can't seem to break the 28.2 barrier no matters what. Finetuning helps on local validation score, but doesnt change LB. Refining with wave propagation does the same…. I don't know what else to try.
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3221106": "In [this notebook](https://www.kaggle.com/code/fpeccia/deepwave-forward-propagation-loss-function), I tried to create a loss function for PINN using Deepwave, but the resulting seismis data is not numerically exact. Has someone managed to configure Deepwave correctly, to get numerically accurate results?", "3221872": "After experimenting with various improvements to Deepwave's seismic wave data inference, I found none to be effective. In my tests, the model exhibits errors of approximately 0.08 MAE and 0.03 MSE.\nWhile Deepwave's fast inference is appealing, it's not sufficiently reliable. As a result, my current approach is to either use https://www.kaggle.com/code/manatoyo/improved-vel-to-seis (though it's slow and can't perform backpropagation, it's quite accurate) or employ Deepwave itself (allowing faster inference at the cost of reduced accuracy). Another option is to implement the notebook's code in PyTorch (which runs faster than the notebook implementation and achieves remarkably good accuracy—about 0.00001 MAE—but remains several orders of magnitude slower than Deepwave).\n" } }