\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14897182%2F522ca7a451d3f178e0f1b5c86abf4405%2Fdownload.png?generation=1751331349475484&alt=media\")
Hi Everyone, I would like to share my solution as well as the trick I used to boost my score.
\nSomething that stood out for me during the competition was that an almost exact reproduction of the forward modelling simulation code used by the hosts was open-sourced here .
\nThis means that given a set of velocity maps, we can use this code to accurately generate its corresponding seismic data pair. This allows us to create a synthetic dataset of extra data to train our model on.
\nThen comes the next question: What set of velocity maps should we use? One approach is to augment the existing velocity maps. Alternatively, we could train a generator model to create new velocity maps.
\nI pondered over this question for a long time, I wanted a set of velocity maps that had a potential to teach the model useful information that would greatly boost its performance.
\nHmmm……If only I had access to the velocity maps of the test set. That would be great. But wait, I do have access to something similar! The model's predicted test set velocity maps.
\nUsing the forward modelling simulation code, the model's predicted test set velocity maps can be used to reconstruct a synthetic test set that largely resembles the actual test set. Training the model on this synthetic test set greatly enhances performance on the actual test set.
\nMethodology:
\nThis process can be repeated multiple times to further enhance performance. I will refer to each repetition as one \"cycle\".
\nEvery time we get a better model, the predicted test set velocity maps also get closer to the actual test set velocity maps. Hence, the more cycles we run, the more similar our synthetic test set will be to the actual test set. This results in further performance increases every cycle.
\nThe intuition behind this idea can be summarised as follows:
\nEssentially, the forward modelling simulation code is acting as a teacher to the model. This is only possible due how well the forward modelling simulation code can reconstruct the seismic data from the velocity maps.
\nI ran a total of 6 cycles (cycle 1 to cycle 6) and the results are shown below.
\n
I made certain changes to the methodology detailed above in certain cycles mainly because I wanted to try out some ideas that I thought were promising:
\nHowever, in hindsight, I believe that these changes were unnecessary and training the model on only the synthetic test set for all cycles would probably have yielded the same results.
\nKey training parameters adjustments:
\nThis competition has been quite a journey, full of ups and downs.
\nEarly on, I had several ideas I was eager to explore, but unfortunately, most of them didn’t work out as I’d hoped. Despite the setbacks, I kept experimenting until I finally discovered the approach that significantly improved my score.
\nFor a while, I thought I might have a real shot at earning my first gold medal. However, in the final stretch, many teams surged up the leaderboard, leaving me behind.
\nEven so, this has been a valuable learning experience, and I’m determined to come back stronger next time.
\nI’m looking forward to seeing what strategies the top teams used.
\nCAFormer - Full Resolution Improved by @brendanartley
\nwaveform-inversion : Vel to Seis by @jaewook704
\nImproved Vel to Seis by @manatoyo
Inference Notebook
\nModels and Training Codes
\nBest Synthetic Test Set
Hi Everyone, I would like to share my solution as well as the trick I used to boost my score.
\nSomething that stood out for me during the competition was that an almost exact reproduction of the forward modelling simulation code used by the hosts was open-sourced here .
\nThis means that given a set of velocity maps, we can use this code to accurately generate its corresponding seismic data pair. This allows us to create a synthetic dataset of extra data to train our model on.
\nThen comes the next question: What set of velocity maps should we use? One approach is to augment the existing velocity maps. Alternatively, we could train a generator model to create new velocity maps.
\nI pondered over this question for a long time, I wanted a set of velocity maps that had a potential to teach the model useful information that would greatly boost its performance.
\nHmmm……If only I had access to the velocity maps of the test set. That would be great. But wait, I do have access to something similar! The model's predicted test set velocity maps.
\nUsing the forward modelling simulation code, the model's predicted test set velocity maps can be used to reconstruct a synthetic test set that largely resembles the actual test set. Training the model on this synthetic test set greatly enhances performance on the actual test set.
\nMethodology:
\nThis process can be repeated multiple times to further enhance performance. I will refer to each repetition as one \"cycle\".
\nEvery time we get a better model, the predicted test set velocity maps also get closer to the actual test set velocity maps. Hence, the more cycles we run, the more similar our synthetic test set will be to the actual test set. This results in further performance increases every cycle.
\nThe intuition behind this idea can be summarised as follows:
\nEssentially, the forward modelling simulation code is acting as a teacher to the model. This is only possible due how well the forward modelling simulation code can reconstruct the seismic data from the velocity maps.
\nI ran a total of 6 cycles (cycle 1 to cycle 6) and the results are shown below.
\n
I made certain changes to the methodology detailed above in certain cycles mainly because I wanted to try out some ideas that I thought were promising:
\nHowever, in hindsight, I believe that these changes were unnecessary and training the model on only the synthetic test set for all cycles would probably have yielded the same results.
\nKey training parameters adjustments:
\nThis competition has been quite a journey, full of ups and downs.
\nEarly on, I had several ideas I was eager to explore, but unfortunately, most of them didn’t work out as I’d hoped. Despite the setbacks, I kept experimenting until I finally discovered the approach that significantly improved my score.
\nFor a while, I thought I might have a real shot at earning my first gold medal. However, in the final stretch, many teams surged up the leaderboard, leaving me behind.
\nEven so, this has been a valuable learning experience, and I’m determined to come back stronger next time.
\nI’m looking forward to seeing what strategies the top teams used.
\nCAFormer - Full Resolution Improved by @brendanartley
\nwaveform-inversion : Vel to Seis by @jaewook704
\nImproved Vel to Seis by @manatoyo
Inference Notebook
\nModels and Training Codes
\nBest Synthetic Test Set
Thank you for sharing. This idea is great！
", "rawMarkdown": "Thank you for sharing. This idea is great！", "votes": null }, { "id": "3237153", "postDate": "07/01/2025 00:21:57", "content": "Congratulations on your accomplishment you made! Your idea is fascinating to add bagging during the training cycle! Very nice idea that we could learn from!🫡
", "rawMarkdown": "Congratulations on your accomplishment you made! Your idea is fascinating to add bagging during the training cycle! Very nice idea that we could learn from!🫡", "votes": null }, { "id": "3237167", "postDate": "07/01/2025 00:29:34", "content": "Very cool take on pseudo-labelling! And a testament to how neat this competition was -- that just brute force of massive hardware was not the right take here, that an elegant take as yours could perform so well, even amidst so much data.
\nCongrats on the terrific performance and thanks for the write-up! 🙂
", "rawMarkdown": "Very cool take on pseudo-labelling! And a testament to how neat this competition was -- that just brute force of massive hardware was not the right take here, that an elegant take as yours could perform so well, even amidst so much data.\n\nCongrats on the terrific performance and thanks for the write-up! 🙂", "votes": null }, { "id": "3237627", "postDate": "07/01/2025 07:45:30", "content": "I wish i realize this.
\nYour explanation is easy to understand.
\nThanks.
Thank you for sharing. This idea is great！
", "votes": null, "replies": [] }, { "id": 3237153, "author_name": "huasteven", "author_url": "", "post_date": "07/01/2025 00:21:57", "content": "Congratulations on your accomplishment you made! Your idea is fascinating to add bagging during the training cycle! Very nice idea that we could learn from!🫡
", "votes": null, "replies": [] }, { "id": 3237167, "author_name": "radek1", "author_url": "", "post_date": "07/01/2025 00:29:34", "content": "Very cool take on pseudo-labelling! And a testament to how neat this competition was -- that just brute force of massive hardware was not the right take here, that an elegant take as yours could perform so well, even amidst so much data.
\nCongrats on the terrific performance and thanks for the write-up! 🙂
", "votes": null, "replies": [] }, { "id": 3237627, "author_name": "overvalueawareness", "author_url": "", "post_date": "07/01/2025 07:45:30", "content": "I wish i realize this.
\nYour explanation is easy to understand.
\nThanks.