Quoting the competition's overview :
\n\n\nTraditional physics-based approaches are accurate, but incredibly slow and prone to errors when the signal is weak from noisy data. Pure machine learning solutions are faster, but require vast amounts of labeled data and often fail to generalize to new, unfamiliar \"signal.\"
\nThis competition challenges you to bridge the gap by combining physics and machine learning to advance FWI.
\n
I don't see any computation limits in the rules, or obligation to use ML. Therefore, what prevents a competitor to use \"incredibly slow\" yet \"accurate\" \"traditional physics-based approaches\" to win the competition ?
\nIn other words, do ML solutions really stand a chance ?
\n@hanchenwang114 @yinanfeng @youzuolin
", "messages": [ { "id": "3178957", "postDate": "04/14/2025 20:23:45", "content": "Quoting the competition's overview :
\n\n\nTraditional physics-based approaches are accurate, but incredibly slow and prone to errors when the signal is weak from noisy data. Pure machine learning solutions are faster, but require vast amounts of labeled data and often fail to generalize to new, unfamiliar \"signal.\"
\nThis competition challenges you to bridge the gap by combining physics and machine learning to advance FWI.
\n
I don't see any computation limits in the rules, or obligation to use ML. Therefore, what prevents a competitor to use \"incredibly slow\" yet \"accurate\" \"traditional physics-based approaches\" to win the competition ?
\nIn other words, do ML solutions really stand a chance ?
\n@hanchenwang114 @yinanfeng @youzuolin
", "rawMarkdown": "Quoting the competition's overview :\n\n> Traditional physics-based approaches are accurate, but incredibly slow and prone to errors when the signal is weak from noisy data. Pure machine learning solutions are faster, but require vast amounts of labeled data and often fail to generalize to new, unfamiliar \"signal.\"\n>\n> This competition challenges you to bridge the gap by combining physics and machine learning to advance FWI.\n\nI don't see any computation limits in the rules, or obligation to use ML. Therefore, what prevents a competitor to use \"incredibly slow\" yet \"accurate\" \"traditional physics-based approaches\" to win the competition ?\n\nIn other words, do ML solutions really stand a chance ?\n\n@hanchenwang114 @yinanfeng @youzuolin", "votes": null }, { "id": "3178989", "postDate": "04/14/2025 21:28:24", "content": "People are allowed to use physics-based methods, like FWI, in this competition. However, there are more than 60K test samples and the final score is evaluated by the overall loss of all the test samples. From my experience, getting accurate enough FWI results for all the test samples will take you a lot of time and computational resources, probably a few months to years.
", "rawMarkdown": "People are allowed to use physics-based methods, like FWI, in this competition. However, there are more than 60K test samples and the final score is evaluated by the overall loss of all the test samples. From my experience, getting accurate enough FWI results for all the test samples will take you a lot of time and computational resources, probably a few months to years.", "votes": null }, { "id": "3179001", "postDate": "04/14/2025 21:48:14", "content": "Thanks for the clarification. If it takes that much resources with current analytical methods, then it seems completely fair indeed if someone wins with a non-ML solution. I guess it would be even more revolutionary.
", "rawMarkdown": "Thanks for the clarification. If it takes that much resources with current analytical methods, then it seems completely fair indeed if someone wins with a non-ML solution. I guess it would be even more revolutionary.", "votes": null }, { "id": "3179224", "postDate": "04/15/2025 07:43:43", "content": "If it's a traditional method or a hybrid approach (part AI, part traditional), is there a time limit for using them?
\nSo as long as the test set answers are solved and uploaded before the competition ends, it's okay, right?
\nIt seems like this isn't a code competition, so there's no strict time limit.
You are right. People can use traditional methods or hybrid approaches to beat the leader board. However, I believe they also need to design algorithms with good code to do so.
", "rawMarkdown": "You are right. People can use traditional methods or hybrid approaches to beat the leader board. However, I believe they also need to design algorithms with good code to do so.", "votes": null }, { "id": "3184472", "postDate": "04/22/2025 06:12:06", "content": "Very well said.
", "rawMarkdown": "Very well said.", "votes": null } ], "comments": [ { "id": 3178989, "author_name": "hanchenwang114", "author_url": "", "post_date": "04/14/2025 21:28:24", "content": "People are allowed to use physics-based methods, like FWI, in this competition. However, there are more than 60K test samples and the final score is evaluated by the overall loss of all the test samples. From my experience, getting accurate enough FWI results for all the test samples will take you a lot of time and computational resources, probably a few months to years.
", "votes": null, "replies": [ { "id": 3179001, "author_name": "gguillard", "author_url": "", "post_date": "04/14/2025 21:48:14", "content": "Thanks for the clarification. If it takes that much resources with current analytical methods, then it seems completely fair indeed if someone wins with a non-ML solution. I guess it would be even more revolutionary.
", "votes": null, "replies": [ { "id": 3184472, "author_name": "tpmeli", "author_url": "", "post_date": "04/22/2025 06:12:06", "content": "Very well said.
", "votes": null, "replies": [] } ] }, { "id": 3179224, "author_name": "atom1231", "author_url": "", "post_date": "04/15/2025 07:43:43", "content": "If it's a traditional method or a hybrid approach (part AI, part traditional), is there a time limit for using them?
\nSo as long as the test set answers are solved and uploaded before the competition ends, it's okay, right?
\nIt seems like this isn't a code competition, so there's no strict time limit.
You are right. People can use traditional methods or hybrid approaches to beat the leader board. However, I believe they also need to design algorithms with good code to do so.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3178957": "Quoting the competition's overview :\n\n> Traditional physics-based approaches are accurate, but incredibly slow and prone to errors when the signal is weak from noisy data. Pure machine learning solutions are faster, but require vast amounts of labeled data and often fail to generalize to new, unfamiliar \"signal.\"\n>\n> This competition challenges you to bridge the gap by combining physics and machine learning to advance FWI.\n\nI don't see any computation limits in the rules, or obligation to use ML. Therefore, what prevents a competitor to use \"incredibly slow\" yet \"accurate\" \"traditional physics-based approaches\" to win the competition ?\n\nIn other words, do ML solutions really stand a chance ?\n\n@hanchenwang114 @yinanfeng @youzuolin", "3178989": "People are allowed to use physics-based methods, like FWI, in this competition. However, there are more than 60K test samples and the final score is evaluated by the overall loss of all the test samples. From my experience, getting accurate enough FWI results for all the test samples will take you a lot of time and computational resources, probably a few months to years.", "3179001": "Thanks for the clarification. If it takes that much resources with current analytical methods, then it seems completely fair indeed if someone wins with a non-ML solution. I guess it would be even more revolutionary.", "3179224": "If it's a traditional method or a hybrid approach (part AI, part traditional), is there a time limit for using them?\nSo as long as the test set answers are solved and uploaded before the competition ends, it's okay, right?\nIt seems like this isn't a code competition, so there's no strict time limit.", "3179878": "You are right. People can use traditional methods or hybrid approaches to beat the leader board. However, I believe they also need to design algorithms with good code to do so.", "3184472": "Very well said." }, "source": "meta" }