Hello All,
\nWelcome to the Yale/UNC-Chapel Hill – Geophysical Waveform Inversion competition! We've included additional resources below on Full Waveform Inversion (FWI), covering methods, available datasets, and practical tutorials to help you get started.
\nWhat is Full-Waveform Inversion (FWI)?
\nFull-Waveform Inversion (FWI) is a powerful technique used to reconstruct high-resolution subsurface or internal physical properties—such as velocity maps—by analyzing waveform data. Originally developed for geophysical applications, FWI has also gained traction in medical ultrasound imaging for detailed tissue characterization, and in engineering for non-destructive testing of materials. FWI methods are generally categorized into:
\n• Physics-based methods, which rely on wave equations and optimization techniques.
\n• Machine learning (ML) methods, which learn mappings from data.
Challenges in Physics-based FWI
\nTraditional FWI approaches, while grounded in physics, face significant limitations:
\n• Low Accuracy: Incomplete data coverage and sensitivity to noise hinder precision.
\n• Nonlinearity: The optimization problem is highly non-linear, often leading to convergence at local minima.
\n• High Computational Cost: Adjoint-state methods used in classical FWI are computationally intensive, especially for large-scale 2D or 3D models.
\nThese challenges limit the real-world deployment of physics-based FWI across industry sectors.
Purely Data-Driven Machine Learning Models
\nRecent developments in ML have introduced fully supervised models that directly map seismic data to subsurface properties. These methods offer promising alternatives but also face key limitations:
\n• Data Requirements: They need large, high-quality labeled datasets that are often expensive to generate.
\n• Limited Generalization: Models trained on specific data distributions often fail to generalize to new geological settings.
\n• Lack of Interpretability: Without grounding in physics, these models act as “black boxes,” reducing user trust and scientific transparency.
Physics-Guided Machine Learning
\nTo address these issues, physics-guided ML combines physical knowledge with data-driven models. This hybrid approach:
\n• Enhances interpretability and reliability,
\n• Reduces the amount of training data needed,
\n• Improves generalization and computational efficiency.
\nFor an overview of this growing field, see reviews in [1] and [2].
Baseline Models
\nTo help you get started, we’ve listed some well-known ML models for FWI. Many come with public codebases you can build on:
Data Availability
\nThe competition uses data from OpenFWI, a large-scale, open-source benchmark for FWI. It includes multiple velocity structures, seismic waveforms, and metadata. The test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations).
\nWe strongly encourage participants to use OpenFWI for training:
\n• OpenFWI (Dataset)and (Paper)
Online Tutorials
\nThese tutorials offer a quick introduction to solving FWI using the OpenFWI dataset:
They walk you through loading seismic waveform data and velocity maps, visualizing samples, and building basic neural networks.
\nBest,
\nThe Waveform Inversion Team
Hello All,
\nWelcome to the Yale/UNC-Chapel Hill – Geophysical Waveform Inversion competition! We've included additional resources below on Full Waveform Inversion (FWI), covering methods, available datasets, and practical tutorials to help you get started.
\nWhat is Full-Waveform Inversion (FWI)?
\nFull-Waveform Inversion (FWI) is a powerful technique used to reconstruct high-resolution subsurface or internal physical properties—such as velocity maps—by analyzing waveform data. Originally developed for geophysical applications, FWI has also gained traction in medical ultrasound imaging for detailed tissue characterization, and in engineering for non-destructive testing of materials. FWI methods are generally categorized into:
\n• Physics-based methods, which rely on wave equations and optimization techniques.
\n• Machine learning (ML) methods, which learn mappings from data.
Challenges in Physics-based FWI
\nTraditional FWI approaches, while grounded in physics, face significant limitations:
\n• Low Accuracy: Incomplete data coverage and sensitivity to noise hinder precision.
\n• Nonlinearity: The optimization problem is highly non-linear, often leading to convergence at local minima.
\n• High Computational Cost: Adjoint-state methods used in classical FWI are computationally intensive, especially for large-scale 2D or 3D models.
\nThese challenges limit the real-world deployment of physics-based FWI across industry sectors.
Purely Data-Driven Machine Learning Models
\nRecent developments in ML have introduced fully supervised models that directly map seismic data to subsurface properties. These methods offer promising alternatives but also face key limitations:
\n• Data Requirements: They need large, high-quality labeled datasets that are often expensive to generate.
\n• Limited Generalization: Models trained on specific data distributions often fail to generalize to new geological settings.
\n• Lack of Interpretability: Without grounding in physics, these models act as “black boxes,” reducing user trust and scientific transparency.
Physics-Guided Machine Learning
\nTo address these issues, physics-guided ML combines physical knowledge with data-driven models. This hybrid approach:
\n• Enhances interpretability and reliability,
\n• Reduces the amount of training data needed,
\n• Improves generalization and computational efficiency.
\nFor an overview of this growing field, see reviews in [1] and [2].
Baseline Models
\nTo help you get started, we’ve listed some well-known ML models for FWI. Many come with public codebases you can build on:
Data Availability
\nThe competition uses data from OpenFWI, a large-scale, open-source benchmark for FWI. It includes multiple velocity structures, seismic waveforms, and metadata. The test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations).
\nWe strongly encourage participants to use OpenFWI for training:
\n• OpenFWI (Dataset)and (Paper)
Online Tutorials
\nThese tutorials offer a quick introduction to solving FWI using the OpenFWI dataset:
They walk you through loading seismic waveform data and velocity maps, visualizing samples, and building basic neural networks.
\nBest,
\nThe Waveform Inversion Team
OpenFWI is a synthetic dataset. You rightly said:
\n\n\nLimited Generalization: Models trained on specific data distributions often fail to generalize to new geological settings.
\n
This immediately raises the question/suspicion: does the challenge in this competition include a data shift? i.e., does the test set come from the same distribution as the synthetic OpenFWI, i.e., was synthetically generated using the same algorithm, or did you include data shift to try to test generalizability?
\nThis is very important for us to know. I did not see a clear answer to this anywhere- the closest is
\n\nThe test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations).
\n
However, it can still follow the same format of data shape but include a significant data shift.
", "rawMarkdown": "OpenFWI is a synthetic dataset. You rightly said:\n\n> Limited Generalization: Models trained on specific data distributions often fail to generalize to new geological settings. \n\nThis immediately raises the question/suspicion: does the challenge in this competition include a data shift? i.e., does the test set come from the same distribution as the synthetic OpenFWI, i.e., was synthetically generated using the same algorithm, or did you include data shift to try to test generalizability?\nThis is very important for us to know. I did not see a clear answer to this anywhere- the closest is \n\n>The test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations). \n\nHowever, it can still follow the same format of data shape but include a significant data shift.", "votes": 10, "replies": [ { "id": 3174305, "postDate": "2025-04-09T01:10:56.190Z", "rawMarkdown": "", "votes": 2, "isDeleted": true, "replies": [ { "id": 3174321, "postDate": "2025-04-09T01:49:46.160Z", "content": "Thank you for the answer. I will follow up with the next question then. Is the source code and detailed instructions for the creation of the datasets are available? I did not find it in the github repository, I checked the suppelementaries (and the paper of course) and it was very lacking for the purpose of reproducing the dataset creation pipeline. In particular, I did not find the python code (which was 'teanslated' from Matlab) and did not see details about the parameters/parameters range needed to recreate each dataset. If the data is supposed to be available i.e. we are supposed to be able to create more data, I would like to know how if better and more detailed instructions are available somewhere (maybe I missed them). If we are not supposed to be able to ceeate more data, it is fine too.
", "rawMarkdown": "Thank you for the answer. I will follow up with the next question then. Is the source code and detailed instructions for the **creation** of the datasets are available? I did not find it in the github repository, I checked the suppelementaries (and the paper of course) and it was very lacking for the purpose of reproducing the dataset creation pipeline. In particular, I did not find the python code (which was 'teanslated' from Matlab) and did not see details about the parameters/parameters range needed to recreate each dataset. If the data is supposed to be available i.e. we are supposed to be able to create more data, I would like to know how if better and more detailed instructions are available somewhere (maybe I missed them). If we are not supposed to be able to ceeate more data, it is fine too. ", "votes": 4 }, { "id": 3174583, "postDate": "2025-04-09T08:27:08.853Z", "content": "Above there was a comment by the host saying there is no data shift. I see this comment is deleted now. There must be a reason, please explain.
", "rawMarkdown": "Above there was a comment by the host saying there is no data shift. I see this comment is deleted now. There must be a reason, please explain.", "votes": 3 }, { "id": 3175333, "postDate": "2025-04-10T03:34:37.157Z", "content": "\"which was 'teanslated' from Matlab\"
\nIf possible please provide the original Matlab code. Thanks.
", "rawMarkdown": "\"which was 'teanslated' from Matlab\"\n\nIf possible please provide the original Matlab code. Thanks." }, { "id": 3176070, "postDate": "2025-04-10T23:50:02.870Z", "content": "Please refer to the response in the other Discussion:
\nhttps://www.kaggle.com/competitions/waveform-inversion/discussion/572334#3175071
Please refer to the original OpenFWI paper, which has pointed to the MatLab simulation code at: https://csim.kaust.edu.sa/files/SeismicInversion/Chapter.FD/lab.FD2.8/lab.html
", "rawMarkdown": "Please refer to the original OpenFWI paper, which has pointed to the MatLab simulation code at: https://csim.kaust.edu.sa/files/SeismicInversion/Chapter.FD/lab.FD2.8/lab.html " }, { "id": 3176077, "postDate": "2025-04-11T00:17:18.500Z", "content": "Sure, we have access to the matlab code, but since the dataset was created with a python translation, I hoped this python code exist and published somewhere, so we can replicate the same pipeline (like, are we sure the translation is exactly identical to the matlab source? No. we can never be sure).
\n@hanchenwang114
So we are using fake data….. Will my work be pushed forward on real data and someone else take the glory?
", "rawMarkdown": "So we are using fake data..... Will my work be pushed forward on real data and someone else take the glory?", "replies": [ { "id": 3176644, "postDate": "2025-04-11T14:42:10.543Z", "content": "\n\nSo we are using fake data….. Will my work be pushed forward on real data and someone else take the glory?
\n
Thanks for your interest in this problem! The data here is generated using the governing physics, in this case, the wave equation. While it's not real-world data, it's a highly realistic simulation, which is why it's often referred to as full-physics simulation. It closely reflects the actual physical processes involved, making it a strong proxy for real scenarios.
\nFull-waveform inversion is a tough problem to solve. One big reason we use high-quality simulated data is that it gives us a clear ground truth to compare against. This makes it easier to evaluate how well an algorithm works before applying it to real data. In science and engineering, that kind of validation step is critical. It helps ensure the method is solid before putting it into practice.
", "rawMarkdown": "> So we are using fake data..... Will my work be pushed forward on real data and someone else take the glory?\n\nThanks for your interest in this problem! The data here is generated using the governing physics, in this case, the wave equation. While it's not real-world data, it's a highly realistic simulation, which is why it's often referred to as full-physics simulation. It closely reflects the actual physical processes involved, making it a strong proxy for real scenarios.\n\nFull-waveform inversion is a tough problem to solve. One big reason we use high-quality simulated data is that it gives us a clear ground truth to compare against. This makes it easier to evaluate how well an algorithm works before applying it to real data. In science and engineering, that kind of validation step is critical. It helps ensure the method is solid before putting it into practice. ", "votes": 5 }, { "id": 3227720, "postDate": "2025-06-19T08:07:21.987Z", "content": "Thank you for sharing. It's very friendly to beginners.
", "rawMarkdown": "Thank you for sharing. It's very friendly to beginners." } ] } ] }, { "id": 3175741, "postDate": "2025-04-10T14:41:00.770Z", "content": "We might need some folks with experience of this dataset to shed some light on this.
", "rawMarkdown": "We might need some folks with experience of this dataset to shed some light on this.", "votes": 2, "replies": [ { "id": 3176720, "postDate": "2025-04-11T16:18:27.743Z", "content": "Please also take a look at this post, put together and shared by @tpmeli (thank you!). It provides a clear and well-explained overview of the dataset, especially for those who are new to full-waveform inversion problems.
", "rawMarkdown": "Please also take a look at this [post](https://www.kaggle.com/competitions/waveform-inversion/discussion/572747), put together and shared by @tpmeli (thank you!). It provides a clear and well-explained overview of the dataset, especially for those who are new to full-waveform inversion problems.", "votes": 4 } ] }, { "id": 3244661, "postDate": "2025-07-08T11:16:20.030Z", "content": "Thank you for sharing, It is very friendly to beginners.
", "rawMarkdown": "Thank you for sharing, It is very friendly to beginners." }, { "id": 3230156, "postDate": "2025-06-22T14:35:07.637Z", "content": "谢谢你的回答。我将跟进下一个问题。是源代码和详细说明创建的数据集是否可用？我没有在github存储库中找到它，我检查了supplementaries (当然还有论文)，它非常缺乏重现数据集创建管道的目的。特别是，我没有找到python代码 (这是从Matlab “teanslated”)，也没有看到有关重新创建每个数据集所需的参数/参数范围的详细信息。如果数据应该是可用的，即我们应该能够创建更多的数据，我想知道如何更好和更详细的说明是可用的地方 (也许我错过了他们)。
", "rawMarkdown": "谢谢你的回答。我将跟进下一个问题。是源代码和详细说明创建的数据集是否可用？我没有在github存储库中找到它，我检查了supplementaries (当然还有论文)，它非常缺乏重现数据集创建管道的目的。特别是，我没有找到python代码 (这是从Matlab “teanslated”)，也没有看到有关重新创建每个数据集所需的参数/参数范围的详细信息。如果数据应该是可用的，即我们应该能够创建更多的数据，我想知道如何更好和更详细的说明是可用的地方 (也许我错过了他们)。" }, { "id": 3227971, "postDate": "2025-06-19T14:01:07.567Z", "content": "Thank you for sharing. It's the best way to learn something new
", "rawMarkdown": "Thank you for sharing. It's the best way to learn something new" }, { "id": 3226331, "postDate": "2025-06-17T14:01:51.100Z", "content": "Thank you for sharing. It's very friendly to beginners.
", "rawMarkdown": "Thank you for sharing. It's very friendly to beginners." }, { "id": 3225452, "postDate": "2025-06-16T12:35:05.787Z", "content": "Thanks for the detailed info and resources! Excited to begin exploring FWI and learning more through this challenge.
", "rawMarkdown": "Thanks for the detailed info and resources! Excited to begin exploring FWI and learning more through this challenge." }, { "id": 3224788, "postDate": "2025-06-15T13:36:41.543Z", "content": "Thank you for sharing. It's very helpful to me!
", "rawMarkdown": "Thank you for sharing. It's very helpful to me!" }, { "id": 3197768, "postDate": "2025-05-08T15:04:20.043Z", "content": "oo might need some people with experience of this dataset to impart some knowledge
", "rawMarkdown": "oo might need some people with experience of this dataset to impart some knowledge" }, { "id": 3191073, "postDate": "2025-05-01T10:50:05.647Z", "content": "Given that the test data is generally synthetic, I wanted to ask: Are real-world data types typically similar in nature to the types provided in the OpenFWI dataset? This understanding is important for potentially specializing or separating models based on data type.
\nFurthermore, for clarification, could you please let us know if the 'style' data type within OpenFWI is considered the most similar or representative of real-world data?
\nI believe that, due to the synthetic data, the outcomes here might end up being solely for the leaderboard, without real-world use.
\nThank you for your clarification.
Dear competition hosts, I couldn't find any submission code running time limit (either CPU or GPU) requirement in the competition page. Could you please let me know what the running time limit is?
", "rawMarkdown": "Dear competition hosts, I couldn't find any submission code running time limit (either CPU or GPU) requirement in the competition page. Could you please let me know what the running time limit is?", "replies": [ { "id": 3180661, "postDate": "2025-04-17T00:13:21.693Z", "content": "This is not a code competition - there is no time limit. Submissions can be created on any device running for any length of time.
", "rawMarkdown": "This is not a code competition - there is no time limit. Submissions can be created on any device running for any length of time.", "votes": 3 } ] }, { "id": 3178625, "postDate": "2025-04-14T12:00:12.147Z", "content": "Hi all, I am looking to put together a team up for this competition, my friend and I are already a part, so if anyone wants to pitch in please feel free to
", "rawMarkdown": "Hi all, I am looking to put together a team up for this competition, my friend and I are already a part, so if anyone wants to pitch in please feel free to" }, { "id": 3176831, "postDate": "2025-04-11T19:10:50.870Z", "content": "We see multiple sources and receivers in the waveform dataset.
\nwhat could be the distance between sources and receivers.?
I donot have good understanding of problem at hand, but i feel the distance between source-source, sensor-sensor and source-senor might play in determining velocity maps
", "rawMarkdown": "We see multiple sources and receivers in the waveform dataset.\nwhat could be the distance between sources and receivers.?\n\nI donot have good understanding of problem at hand, but i feel the distance between source-source, sensor-sensor and source-senor might play in determining velocity maps", "replies": [ { "id": 3180663, "postDate": "2025-04-17T00:17:25.047Z", "content": "", "rawMarkdown": "[It's a very confusing topic](https://www.kaggle.com/competitions/waveform-inversion/discussion/572867#3180468)", "replies": [ { "id": 3185884, "postDate": "2025-04-24T01:14:38.503Z", "content": "Same here, I understand the topic but I dont know where to start from
", "rawMarkdown": "Same here, I understand the topic but I dont know where to start from" }, { "id": 3205936, "postDate": "2025-05-20T15:12:52.577Z", "content": "Yes , You are right
", "rawMarkdown": "Yes , You are right " } ] } ] }, { "id": 3193338, "postDate": "2025-05-04T08:25:35.597Z", "content": "Please tell me lightly what kind of data set it is.
", "rawMarkdown": "Please tell me lightly what kind of data set it is.", "votes": -3, "isDeleted": true }, { "id": 3220236, "postDate": "2025-06-09T03:38:17.587Z", "content": "Thanks for the information
", "rawMarkdown": "Thanks for the information" } ], "comments": [ { "id": 3174287, "author_name": "greySnow", "author_url": "", "post_date": "2025-04-09T00:11:56.907000", "content": "OpenFWI is a synthetic dataset. You rightly said:
\n\n\nLimited Generalization: Models trained on specific data distributions often fail to generalize to new geological settings.
\n
This immediately raises the question/suspicion: does the challenge in this competition include a data shift? i.e., does the test set come from the same distribution as the synthetic OpenFWI, i.e., was synthetically generated using the same algorithm, or did you include data shift to try to test generalizability?
\nThis is very important for us to know. I did not see a clear answer to this anywhere- the closest is
\n\nThe test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations).
\n
However, it can still follow the same format of data shape but include a significant data shift.
", "votes": 10, "replies": [ { "id": 3174305, "author_name": "", "author_url": "", "post_date": "2025-04-09T01:10:56.190000", "content": "", "votes": 2, "replies": [ { "id": 3174321, "author_name": "greySnow", "author_url": "", "post_date": "2025-04-09T01:49:46.160000", "content": "Thank you for the answer. I will follow up with the next question then. Is the source code and detailed instructions for the creation of the datasets are available? I did not find it in the github repository, I checked the suppelementaries (and the paper of course) and it was very lacking for the purpose of reproducing the dataset creation pipeline. In particular, I did not find the python code (which was 'teanslated' from Matlab) and did not see details about the parameters/parameters range needed to recreate each dataset. If the data is supposed to be available i.e. we are supposed to be able to create more data, I would like to know how if better and more detailed instructions are available somewhere (maybe I missed them). If we are not supposed to be able to ceeate more data, it is fine too.
", "votes": 4, "replies": [] }, { "id": 3174583, "author_name": "greySnow", "author_url": "", "post_date": "2025-04-09T08:27:08.853000", "content": "Above there was a comment by the host saying there is no data shift. I see this comment is deleted now. There must be a reason, please explain.
", "votes": 3, "replies": [] }, { "id": 3175333, "author_name": "Truth Seeker", "author_url": "", "post_date": "2025-04-10T03:34:37.157000", "content": "\"which was 'teanslated' from Matlab\"
\nIf possible please provide the original Matlab code. Thanks.
", "votes": 0, "replies": [] }, { "id": 3176070, "author_name": "Hanchen Wang", "author_url": "", "post_date": "2025-04-10T23:50:02.870000", "content": "Please refer to the response in the other Discussion:
\nhttps://www.kaggle.com/competitions/waveform-inversion/discussion/572334#3175071
Please refer to the original OpenFWI paper, which has pointed to the MatLab simulation code at: https://csim.kaust.edu.sa/files/SeismicInversion/Chapter.FD/lab.FD2.8/lab.html
", "votes": 0, "replies": [] }, { "id": 3176077, "author_name": "greySnow", "author_url": "", "post_date": "2025-04-11T00:17:18.500000", "content": "Sure, we have access to the matlab code, but since the dataset was created with a python translation, I hoped this python code exist and published somewhere, so we can replicate the same pipeline (like, are we sure the translation is exactly identical to the matlab source? No. we can never be sure).
\n@hanchenwang114
So we are using fake data….. Will my work be pushed forward on real data and someone else take the glory?
", "votes": 0, "replies": [ { "id": 3176644, "author_name": "Youzuo Lin", "author_url": "", "post_date": "2025-04-11T14:42:10.543000", "content": "\n\nSo we are using fake data….. Will my work be pushed forward on real data and someone else take the glory?
\n
Thanks for your interest in this problem! The data here is generated using the governing physics, in this case, the wave equation. While it's not real-world data, it's a highly realistic simulation, which is why it's often referred to as full-physics simulation. It closely reflects the actual physical processes involved, making it a strong proxy for real scenarios.
\nFull-waveform inversion is a tough problem to solve. One big reason we use high-quality simulated data is that it gives us a clear ground truth to compare against. This makes it easier to evaluate how well an algorithm works before applying it to real data. In science and engineering, that kind of validation step is critical. It helps ensure the method is solid before putting it into practice.
", "votes": 5, "replies": [] }, { "id": 3227720, "author_name": "Laurie Lei", "author_url": "", "post_date": "2025-06-19T08:07:21.987000", "content": "Thank you for sharing. It's very friendly to beginners.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3175741, "author_name": "AIDLRE001", "author_url": "", "post_date": "2025-04-10T14:41:00.770000", "content": "We might need some folks with experience of this dataset to shed some light on this.
", "votes": 2, "replies": [ { "id": 3176720, "author_name": "Youzuo Lin", "author_url": "", "post_date": "2025-04-11T16:18:27.743000", "content": "Please also take a look at this post, put together and shared by @tpmeli (thank you!). It provides a clear and well-explained overview of the dataset, especially for those who are new to full-waveform inversion problems.
", "votes": 4, "replies": [] } ] }, { "id": 3244661, "author_name": "Zheng Haifei", "author_url": "", "post_date": "2025-07-08T11:16:20.030000", "content": "Thank you for sharing, It is very friendly to beginners.
", "votes": 0, "replies": [] }, { "id": 3230156, "author_name": "Liu Run Lin", "author_url": "", "post_date": "2025-06-22T14:35:07.637000", "content": "谢谢你的回答。我将跟进下一个问题。是源代码和详细说明创建的数据集是否可用？我没有在github存储库中找到它，我检查了supplementaries (当然还有论文)，它非常缺乏重现数据集创建管道的目的。特别是，我没有找到python代码 (这是从Matlab “teanslated”)，也没有看到有关重新创建每个数据集所需的参数/参数范围的详细信息。如果数据应该是可用的，即我们应该能够创建更多的数据，我想知道如何更好和更详细的说明是可用的地方 (也许我错过了他们)。
", "votes": 0, "replies": [] }, { "id": 3227971, "author_name": "Nicola Curtarello", "author_url": "", "post_date": "2025-06-19T14:01:07.567000", "content": "Thank you for sharing. It's the best way to learn something new
", "votes": 0, "replies": [] }, { "id": 3226331, "author_name": "Wangjiayu1125", "author_url": "", "post_date": "2025-06-17T14:01:51.100000", "content": "Thank you for sharing. It's very friendly to beginners.
", "votes": 0, "replies": [] }, { "id": 3225452, "author_name": "Kangkan Kalita", "author_url": "", "post_date": "2025-06-16T12:35:05.787000", "content": "Thanks for the detailed info and resources! Excited to begin exploring FWI and learning more through this challenge.
", "votes": 0, "replies": [] }, { "id": 3224788, "author_name": "Suxing Li", "author_url": "", "post_date": "2025-06-15T13:36:41.543000", "content": "Thank you for sharing. It's very helpful to me!
", "votes": 0, "replies": [] }, { "id": 3197768, "author_name": "SoulShadow8326", "author_url": "", "post_date": "2025-05-08T15:04:20.043000", "content": "oo might need some people with experience of this dataset to impart some knowledge
", "votes": 0, "replies": [] }, { "id": 3191073, "author_name": "Arash karoudi", "author_url": "", "post_date": "2025-05-01T10:50:05.647000", "content": "Given that the test data is generally synthetic, I wanted to ask: Are real-world data types typically similar in nature to the types provided in the OpenFWI dataset? This understanding is important for potentially specializing or separating models based on data type.
\nFurthermore, for clarification, could you please let us know if the 'style' data type within OpenFWI is considered the most similar or representative of real-world data?
\nI believe that, due to the synthetic data, the outcomes here might end up being solely for the leaderboard, without real-world use.
\nThank you for your clarification.
Dear competition hosts, I couldn't find any submission code running time limit (either CPU or GPU) requirement in the competition page. Could you please let me know what the running time limit is?
", "votes": 0, "replies": [ { "id": 3180661, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2025-04-17T00:13:21.693000", "content": "This is not a code competition - there is no time limit. Submissions can be created on any device running for any length of time.
", "votes": 3, "replies": [] } ] }, { "id": 3178625, "author_name": "Dibyojyoti Bhattacharjee", "author_url": "", "post_date": "2025-04-14T12:00:12.147000", "content": "Hi all, I am looking to put together a team up for this competition, my friend and I are already a part, so if anyone wants to pitch in please feel free to
", "votes": 0, "replies": [] }, { "id": 3176831, "author_name": "doteeee", "author_url": "", "post_date": "2025-04-11T19:10:50.870000", "content": "We see multiple sources and receivers in the waveform dataset.
\nwhat could be the distance between sources and receivers.?
I donot have good understanding of problem at hand, but i feel the distance between source-source, sensor-sensor and source-senor might play in determining velocity maps
", "votes": 0, "replies": [ { "id": 3180663, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2025-04-17T00:17:25.047000", "content": "", "votes": 0, "replies": [ { "id": 3185884, "author_name": "Israel Afriyie", "author_url": "", "post_date": "2025-04-24T01:14:38.503000", "content": "Same here, I understand the topic but I dont know where to start from
", "votes": 0, "replies": [] }, { "id": 3205936, "author_name": "Ᏸ𐍂Ꭷ", "author_url": "", "post_date": "2025-05-20T15:12:52.577000", "content": "Yes , You are right
", "votes": 0, "replies": [] } ] } ] }, { "id": 3193338, "author_name": "", "author_url": "", "post_date": "2025-05-04T08:25:35.597000", "content": "Please tell me lightly what kind of data set it is.
", "votes": -3, "replies": [] }, { "id": 3220236, "author_name": "Parth Deshmukh", "author_url": "", "post_date": "2025-06-09T03:38:17.587000", "content": "Thanks for the information
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3174283": "Hello All,\n\nWelcome to the **Yale/UNC-Chapel Hill – Geophysical Waveform Inversion competition**! We've included additional resources below on Full Waveform Inversion (FWI), covering methods, available datasets, and practical tutorials to help you get started.\n\n**What is Full-Waveform Inversion (FWI)?** \nFull-Waveform Inversion (FWI) is a powerful technique used to reconstruct high-resolution subsurface or internal physical properties—such as velocity maps—by analyzing waveform data. Originally developed for geophysical applications, FWI has also gained traction in medical ultrasound imaging for detailed tissue characterization, and in engineering for non-destructive testing of materials. FWI methods are generally categorized into:\n•\t**Physics-based methods**, which rely on wave equations and optimization techniques.\n•\t**Machine learning (ML) methods**, which learn mappings from data.\n\n**Challenges in Physics-based FWI**\nTraditional FWI approaches, while grounded in physics, face significant limitations:\n•\t**Low Accuracy**: Incomplete data coverage and sensitivity to noise hinder precision.\n•\t**Nonlinearity**: The optimization problem is highly non-linear, often leading to convergence at local minima.\n•\t**High Computational Cost**: Adjoint-state methods used in classical FWI are computationally intensive, especially for large-scale 2D or 3D models.\nThese challenges limit the real-world deployment of physics-based FWI across industry sectors.\n\n**Purely Data-Driven Machine Learning Models**\nRecent developments in ML have introduced fully supervised models that directly map seismic data to subsurface properties. These methods offer promising alternatives but also face key limitations:\n•\t**Data Requirements**: They need large, high-quality labeled datasets that are often expensive to generate.\n•\t**Limited Generalization**: Models trained on specific data distributions often fail to generalize to new geological settings.\n•\t**Lack of Interpretability**: Without grounding in physics, these models act as “black boxes,” reducing user trust and scientific transparency.\n\n**Physics-Guided Machine Learning**\nTo address these issues, physics-guided ML combines physical knowledge with data-driven models. This hybrid approach:\n•\tEnhances interpretability and reliability,\n•\tReduces the amount of training data needed,\n•\tImproves generalization and computational efficiency.\nFor an overview of this growing field, see reviews in [[1]](https://ieeexplore.ieee.org/document/10004771) and [[2]](https://arxiv.org/abs/2410.08329).\n\n**Baseline Models**\nTo help you get started, we’ve listed some well-known ML models for FWI. Many come with public codebases you can build on:\n1.\tInversionNet, ([Paper](https://ieeexplore.ieee.org/document/8918045), [Codes](https://openfwi-lanl.github.io/tutorial/#/?id=inversionnet))\n2.\tVelocityGAN, ([Paper](https://ieeexplore.ieee.org/document/9044635), [Codes](https://openfwi-lanl.github.io/tutorial/#/?id=velocitygan)) \n3.\tBigFWI ([Paper](https://www.nature.com/articles/s41598-024-68573-7))\n4.\tDeepONet ([Paper](https://www.sciencedirect.com/science/article/abs/pii/S0045782523004243))\n5. Auto-Linear ([Paper](https://proceedings.mlr.press/v235/feng24a.html))\n\n**Data Availability**\nThe competition uses data from OpenFWI, a large-scale, open-source benchmark for FWI. It includes multiple velocity structures, seismic waveforms, and metadata. The test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations).\nWe strongly encourage participants to use OpenFWI for training:\n•\tOpenFWI ([Dataset](https://smileunc.github.io/projects/openfwi))and ([Paper](https://proceedings.neurips.cc/paper_files/paper/2022/hash/27d3ef263c7cb8d542c4f9815a49b69b-Abstract-Datasets_and_Benchmarks.html))\n\n**Online Tutorials**\nThese tutorials offer a quick introduction to solving FWI using the OpenFWI dataset:\n1.\t[Kaggle_Tutorial](https://www.kaggle.com/code/hanchenwang114/waveform-inversion-kaggle-competition-tutorial)\n2.\t[Medium Article](https://medium.com/better-programming/seismic-data-to-subsurface-models-with-openfwi-bcca0218b4e8)\n\nThey walk you through loading seismic waveform data and velocity maps, visualizing samples, and building basic neural networks.\n\nBest,\nThe Waveform Inversion Team", "3174287": "OpenFWI is a synthetic dataset. You rightly said:\n\n> Limited Generalization: Models trained on specific data distributions often fail to generalize to new geological settings. \n\nThis immediately raises the question/suspicion: does the challenge in this competition include a data shift? i.e., does the test set come from the same distribution as the synthetic OpenFWI, i.e., was synthetically generated using the same algorithm, or did you include data shift to try to test generalizability?\nThis is very important for us to know. I did not see a clear answer to this anywhere- the closest is \n\n>The test data in this competition follows the same format as OpenFWI (e.g., velocity dimensions, source/receiver configurations). \n\nHowever, it can still follow the same format of data shape but include a significant data shift.", "3175741": "We might need some folks with experience of this dataset to shed some light on this.", "3244661": "Thank you for sharing, It is very friendly to beginners.", "3230156": "谢谢你的回答。我将跟进下一个问题。是源代码和详细说明创建的数据集是否可用？我没有在github存储库中找到它，我检查了supplementaries (当然还有论文)，它非常缺乏重现数据集创建管道的目的。特别是，我没有找到python代码 (这是从Matlab “teanslated”)，也没有看到有关重新创建每个数据集所需的参数/参数范围的详细信息。如果数据应该是可用的，即我们应该能够创建更多的数据，我想知道如何更好和更详细的说明是可用的地方 (也许我错过了他们)。", "3227971": "Thank you for sharing. It's the best way to learn something new", "3226331": "Thank you for sharing. It's very friendly to beginners.", "3225452": "Thanks for the detailed info and resources! Excited to begin exploring FWI and learning more through this challenge.", "3224788": "Thank you for sharing. It's very helpful to me!", "3197768": "oo might need some people with experience of this dataset to impart some knowledge", "3191073": "Given that the test data is generally synthetic, I wanted to ask: Are real-world data types typically similar in nature to the types provided in the OpenFWI dataset? This understanding is important for potentially specializing or separating models based on data type.\nFurthermore, for clarification, could you please let us know if the 'style' data type within OpenFWI is considered the most similar or representative of real-world data?\nI believe that, due to the synthetic data, the outcomes here might end up being solely for the leaderboard, without real-world use.\nThank you for your clarification.", "3178841": "Dear competition hosts, I couldn't find any submission code running time limit (either CPU or GPU) requirement in the competition page. Could you please let me know what the running time limit is?", "3178625": "Hi all, I am looking to put together a team up for this competition, my friend and I are already a part, so if anyone wants to pitch in please feel free to", "3176831": "We see multiple sources and receivers in the waveform dataset.\nwhat could be the distance between sources and receivers.?\n\nI donot have good understanding of problem at hand, but i feel the distance between source-source, sensor-sensor and source-senor might play in determining velocity maps", "3193338": "Please tell me lightly what kind of data set it is.", "3220236": "Thanks for the information" } }