Hello everyone,
\n\nI am one of the Los Alamos researchers on the organizing team. We are a small group of machine learning scientists and geophysicists, and we have been using and developing ML methods on seismic data for the last 3 years.
\n\nOur published work on the topic is in the following papers:\nhttps://doi.org/10.1002/2017GL074677\nhttps://doi.org/10.1002/2017GL076708\nhttps://rdcu.be/bdG8Y\nhttps://rdcu.be/bdG9r
\n\nThe first 3 papers will explain our initial work on data similar to the one you have for the challenge. The 4th paper will show initial results on applying the methods developed on lab data on field data with success, on a particular type of earthquakes known as slow earthquakes.
\n\nThe data for this challenge comes from a classic laboratory earthquake experiment, that has been studied in depth as a tabletop analog of seismogenic faults for decades. A number of physical laws widely used by the geoscience community have been derived from this earthquake machine.
\n\nFor this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly.
\n\nThe data from this classic earthquake machine had never been studied using ML until our early efforts 3 years ago, and much remains to be explored and discovered.
\n\nMany thanks to Kaggle for hosting the competition pro bono, we are looking forward to seeing what the community can discover in our data!
\n\nBertrand Rouet-Leduc\nhttps://twitter.com/bertrandrl
", "messages": [ { "id": 455430, "postDate": "2019-01-13T21:08:03.540Z", "content": "Hello everyone,
\n\nI am one of the Los Alamos researchers on the organizing team. We are a small group of machine learning scientists and geophysicists, and we have been using and developing ML methods on seismic data for the last 3 years.
\n\nOur published work on the topic is in the following papers:\nhttps://doi.org/10.1002/2017GL074677\nhttps://doi.org/10.1002/2017GL076708\nhttps://rdcu.be/bdG8Y\nhttps://rdcu.be/bdG9r
\n\nThe first 3 papers will explain our initial work on data similar to the one you have for the challenge. The 4th paper will show initial results on applying the methods developed on lab data on field data with success, on a particular type of earthquakes known as slow earthquakes.
\n\nThe data for this challenge comes from a classic laboratory earthquake experiment, that has been studied in depth as a tabletop analog of seismogenic faults for decades. A number of physical laws widely used by the geoscience community have been derived from this earthquake machine.
\n\nFor this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly.
\n\nThe data from this classic earthquake machine had never been studied using ML until our early efforts 3 years ago, and much remains to be explored and discovered.
\n\nMany thanks to Kaggle for hosting the competition pro bono, we are looking forward to seeing what the community can discover in our data!
\n\nBertrand Rouet-Leduc\nhttps://twitter.com/bertrandrl
", "rawMarkdown": "Hello everyone, \n\nI am one of the Los Alamos researchers on the organizing team. We are a small group of machine learning scientists and geophysicists, and we have been using and developing ML methods on seismic data for the last 3 years.\n\nOur published work on the topic is in the following papers:\nhttps://doi.org/10.1002/2017GL074677\nhttps://doi.org/10.1002/2017GL076708\nhttps://rdcu.be/bdG8Y\nhttps://rdcu.be/bdG9r\n\nThe first 3 papers will explain our initial work on data similar to the one you have for the challenge. The 4th paper will show initial results on applying the methods developed on lab data on field data with success, on a particular type of earthquakes known as slow earthquakes.\n\nThe data for this challenge comes from a classic laboratory earthquake experiment, that has been studied in depth as a tabletop analog of seismogenic faults for decades. A number of physical laws widely used by the geoscience community have been derived from this earthquake machine.\n\nFor this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly. \n\nThe data from this classic earthquake machine had never been studied using ML until our early efforts 3 years ago, and much remains to be explored and discovered.\n\nMany thanks to Kaggle for hosting the competition *pro bono*, we are looking forward to seeing what the community can discover in our data!\n\nBertrand Rouet-Leduc\nhttps://twitter.com/bertrandrl\n\n\n\n\n\n", "votes": 109 }, { "id": 455466, "postDate": "2019-01-14T00:58:58.703Z", "content": "Hi Bertrand, thank you and your team for hosting this meaningful kaggle competition and sharing your work with us.\nI briefly looked through the above mentioned papers. All the model performances were evaluated using R2, effectively MSE. However in this competition, we use MAE as the performance metric. Could you share with us some insights regarding why we change the metric.
\n\nSincerely
", "rawMarkdown": "Hi Bertrand, thank you and your team for hosting this meaningful kaggle competition and sharing your work with us.\nI briefly looked through the above mentioned papers. All the model performances were evaluated using R2, effectively MSE. However in this competition, we use MAE as the performance metric. Could you share with us some insights regarding why we change the metric.\n\nSincerely", "votes": 22, "replies": [ { "id": 455649, "postDate": "2019-01-14T09:44:11.623Z", "content": "I would like to know this too :)
", "rawMarkdown": "I would like to know this too :)" }, { "id": 456509, "postDate": "2019-01-16T00:02:39.963Z", "content": "Hi Eliott, thanks for participating in our competition. Machine learning is in its infancy in earth science and seismology and there is no benchmark or commonly used metric. We chose to go with MAE for the competition so that the metric is a physical quantity, time.
", "rawMarkdown": "Hi Eliott, thanks for participating in our competition. Machine learning is in its infancy in earth science and seismology and there is no benchmark or commonly used metric. We chose to go with MAE for the competition so that the metric is a physical quantity, time. ", "votes": 19 }, { "id": 522930, "postDate": "2019-04-25T08:46:13.167Z", "content": "\n\n\nso that the metric is a physical quantity, time.
\n
Root mean squared error would also be time.
", "rawMarkdown": "> so that the metric is a physical quantity, time. \n\nRoot mean squared error would also be time.", "votes": 2 } ] }, { "id": 460846, "postDate": "2019-01-24T14:33:19.437Z", "content": "I have received the dataset mentioned in the paper(p4581 and p2394 dataset). I could see the acoustic data data similar to this competition, but I could not find the time to failure data.\nIs the time to failure data mentioned in the article publicly available?\nThanks.
", "rawMarkdown": "I have received the dataset mentioned in the paper(p4581 and p2394 dataset). I could see the acoustic data data similar to this competition, but I could not find the time to failure data.\nIs the time to failure data mentioned in the article publicly available?\nThanks.", "votes": 3, "replies": [ { "id": 461814, "postDate": "2019-01-27T03:52:11.263Z", "content": "I was looking at those two dataset as well. But up to my current understanding, they are of little use to our competition.
\n\nP4581, yes, was recorded using the same device that recorded our data. But the dataset was used for different purposes.
\n\nI could find two papers using P4581, one is to estimate friction, one is to identify abnormal event.
\n\nIn our competition, of course, you can imagine it as estimating both friction and identifying abnormal events, then using the results to estimate fault slip time. If you are building some complicated pipeline like this, the dataset might be useful to you..
\n\nIf the host can help check my understanding, will be grateful.
\n\nI do have features to pick up abnormal events anyway, but they show little/no contributions to my current model. They can't help improve the systematic estimation error occurs for the 3 lengthiest fault slips (time to failure starting with >13) in the training set. I'm afraid because of this, those who are implementing CNN/RNN would hardly find any surprises as well.
", "rawMarkdown": "I was looking at those two dataset as well. But up to my current understanding, they are of little use to our competition.\n\nP4581, yes, was recorded using the same device that recorded our data. But the dataset was used for different purposes.\n\nI could find two papers using P4581, one is to estimate friction, one is to identify abnormal event.\n\nIn our competition, of course, you can imagine it as estimating both friction and identifying abnormal events, then using the results to estimate fault slip time. If you are building some complicated pipeline like this, the dataset might be useful to you..\n\nIf the host can help check my understanding, will be grateful.\n\n\n\n*I do have features to pick up abnormal events anyway, but they show little/no contributions to my current model. They can't help improve the systematic estimation error occurs for the 3 lengthiest fault slips (time to failure starting with >13) in the training set. I'm afraid because of this, those who are implementing CNN/RNN would hardly find any surprises as well.*\n\n", "votes": 3 }, { "id": 466166, "postDate": "2019-02-04T19:48:16.647Z", "rawMarkdown": "", "isDeleted": true }, { "id": 466167, "postDate": "2019-02-04T19:48:56.533Z", "content": "Hi Elliott, p4581 is an experiment that has the same setup in terms of machine, material and recording apparatus. The stress conditions applied on the artificial faults were different however, resulting in very different earthquake cycles. Time to failure is identified from the friction and therefore using p4581 to train a model to predict friction could indeed help for the competition, since the properties of the seismic data scale with the stress conditions, as is shown is the second paper referenced above.
", "rawMarkdown": "Hi Elliott, p4581 is an experiment that has the same setup in terms of machine, material and recording apparatus. The stress conditions applied on the artificial faults were different however, resulting in very different earthquake cycles. Time to failure is identified from the friction and therefore using p4581 to train a model to predict friction could indeed help for the competition, since the properties of the seismic data scale with the stress conditions, as is shown is the second paper referenced above.", "votes": 2 }, { "id": 467547, "postDate": "2019-02-07T09:58:22.903Z", "content": "Thank you Bertrand, this explains it, crystal clear!
", "rawMarkdown": "Thank you Bertrand, this explains it, crystal clear!", "votes": 1 }, { "id": 500688, "postDate": "2019-03-26T12:23:55.513Z", "content": "interesting. about experiment p4581 - I could not find the stress conditions or friction values in the files, only the acoustic signal. Are they at all available?
", "rawMarkdown": "interesting. about experiment p4581 - I could not find the stress conditions or friction values in the files, only the acoustic signal. Are they at all available?" } ] }, { "id": 464841, "postDate": "2019-02-01T16:03:12.020Z", "content": "Bertrand, thank you for hosting this competition and congratulations on your Nature Geoscience publications. I had actually hoped, when I saw the announcement for this competition, that you would have spatio-temporal data (kind of like weather forecasting). Time-to-failure 'only' data makes this more like an exercise in technical stock-price forecasting (i.e., based on price time-series only, rather than news and financials). I'm guessing you will find a lot of artifacts from overfitting; my financial analogue would be Elliott Wave Theory. Anyway, I think I'll sit this one out, but if you ever post spatio-temporal tremors data, I would be keen to try my hand.
", "rawMarkdown": "Bertrand, thank you for hosting this competition and congratulations on your Nature Geoscience publications. I had actually hoped, when I saw the announcement for this competition, that you would have spatio-temporal data (kind of like weather forecasting). Time-to-failure 'only' data makes this more like an exercise in technical stock-price forecasting (i.e., based on price time-series only, rather than news and financials). I'm guessing you will find a lot of artifacts from overfitting; my financial analogue would be Elliott Wave Theory. Anyway, I think I'll sit this one out, but if you ever post spatio-temporal tremors data, I would be keen to try my hand.", "votes": 1 }, { "id": 486538, "postDate": "2019-03-08T23:46:36.207Z", "content": "Hi Bertrand,
\n\nThis is an exciting challenge. I've been reviewing the various papers you've shared, and a few others from other sources that came up on USGS. I have a few basic questions (I apologize if I'm way off in my understanding here):
\n\nSome of the above papers refer to \"repeating earthquakes\" with \"continuous seismic signals\", while the data shared is for \"aperiodic earthquakes\". Is there any literature that has used the classic laboratory model to create these \"aperiodic earthquakes\"? It would be good to get some literature on these specific type of earthquakes to find a good way to frame the ML question, before applying a model.
In one of your responses below, you've mentioned using another data set to help determine friction that can be helpful in applying to this competition - can you please elaborate what you mean by that? Did you mean estimating the friction as a function of the acoustic signal?
Is the failure observed for this competition's dataset, a.k.a. \"aperiodic earthquakes\" due to stick-slip failure? If not, what is the technical term for the type of failure that triggers these aperiodic earthquakes?
Thanks!
", "rawMarkdown": "Hi Bertrand,\n\nThis is an exciting challenge. I've been reviewing the various papers you've shared, and a few others from other sources that came up on USGS. I have a few basic questions (I apologize if I'm way off in my understanding here):\n\n1. Some of the above papers refer to \"repeating earthquakes\" with \"continuous seismic signals\", while the data shared is for \"aperiodic earthquakes\". Is there any literature that has used the classic laboratory model to create these \"aperiodic earthquakes\"? It would be good to get some literature on these specific type of earthquakes to find a good way to frame the ML question, before applying a model.\n\n2. In one of your responses below, you've mentioned using another data set to help determine friction that can be helpful in applying to this competition - can you please elaborate what you mean by that? Did you mean estimating the friction as a function of the acoustic signal?\n\n3. Is the failure observed for this competition's dataset, a.k.a. \"aperiodic earthquakes\" due to stick-slip failure? If not, what is the technical term for the type of failure that triggers these aperiodic earthquakes?\n\nThanks!", "votes": 2 }, { "id": 469542, "postDate": "2019-02-11T11:55:04.487Z", "content": "You may be interested in my air voltage signals for quake forecast :-)\nhttps://www.facebook.com/dysondyson
", "rawMarkdown": "You may be interested in my air voltage signals for quake forecast :-)\nhttps://www.facebook.com/dysondyson", "votes": -2 }, { "id": 524782, "postDate": "2019-04-29T14:12:29.507Z", "content": "What is the expected degree of precision?
", "rawMarkdown": "What is the expected degree of precision? " }, { "id": 515364, "postDate": "2019-04-12T13:53:57.047Z", "content": "Hi Bertrand,
\n\nThanks for the info. I wonder whether I can get information about the physical model of the experiments.
", "rawMarkdown": "Hi Bertrand,\n\nThanks for the info. I wonder whether I can get information about the physical model of the experiments. " }, { "id": 497347, "postDate": "2019-03-23T12:09:25.940Z", "content": "Does the signal data provided contain displacement amplitudes or acceleration?
", "rawMarkdown": "Does the signal data provided contain displacement amplitudes or acceleration?" }, { "id": 478766, "postDate": "2019-02-26T15:51:35.620Z", "content": "Great Challenge ... lot of possiblities and it will change a little bit of house price estimation, image recognition or malware classification ... Thanks a lot :)
", "rawMarkdown": "Great Challenge ... lot of possiblities and it will change a little bit of house price estimation, image recognition or malware classification ... Thanks a lot :)" }, { "id": 477084, "postDate": "2019-02-23T20:42:16.970Z", "content": "Bertrand, I enjoyed reading the articles you posted and am looking forward to participating in this competition.
", "rawMarkdown": "Bertrand, I enjoyed reading the articles you posted and am looking forward to participating in this competition." }, { "id": 464981, "postDate": "2019-02-01T23:52:27.643Z", "content": "Hi Bertrand,\nwhat is the difference between the dataset provided in \"LANL Earthquake Prediction\"-Kaggle Challenge\nand the dataset which was used in the following paper:\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2017GL074677 ???
\n\nSame testbench? What is different about this experiment?
\n\n\"For this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly.\"\nCan you provide more details, please? Where is the difference between this experiment and the experiment in the paper above?
\n\nThanks in Advance
", "rawMarkdown": "Hi Bertrand,\nwhat is the difference between the dataset provided in \"LANL Earthquake Prediction\"-Kaggle Challenge\nand the dataset which was used in the following paper:\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2017GL074677 ???\n\nSame testbench? What is different about this experiment?\n\n\"For this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly.\"\nCan you provide more details, please? Where is the difference between this experiment and the experiment in the paper above?\n\nThanks in Advance", "replies": [ { "id": 466157, "postDate": "2019-02-04T19:35:57.383Z", "content": "Hi JohnDillinger, the experimental data comes from the same machine with the same fault material. The main difference is that the stress applied to the fault is lower, leading to more unstable and random lab earthquakes. An other difference is that the device used to record seismic data is much more sensitive, it is a piezoceramic sensor instead of an accelerometer.
", "rawMarkdown": "Hi JohnDillinger, the experimental data comes from the same machine with the same fault material. The main difference is that the stress applied to the fault is lower, leading to more unstable and random lab earthquakes. An other difference is that the device used to record seismic data is much more sensitive, it is a piezoceramic sensor instead of an accelerometer.", "votes": 1 } ] }, { "id": 464577, "postDate": "2019-02-01T05:24:26.550Z", "content": "a curiosity Bertrand, we know this prediction on a test set (after training whatever ensemble model will appear) would be the \"prediction of the century\"....nobody has ever predicted real (non laboratory) earthquakes ...so what brings you to believe that this important exercise will work here : ) .... and let me say that I do believe you and kaggle are doing the right thing : )
", "rawMarkdown": "a curiosity Bertrand, we know this prediction on a test set (after training whatever ensemble model will appear) would be the \"prediction of the century\"....nobody has ever predicted real (non laboratory) earthquakes ...so what brings you to believe that this important exercise will work here : ) .... and let me say that I do believe you and kaggle are doing the right thing : ) " }, { "id": 460918, "postDate": "2019-01-24T18:02:07.037Z", "content": "Thank you and yout team for setting up such challenge. I will dive into those articles to get a better vision of the problem.\nRegards.
", "rawMarkdown": "Thank you and yout team for setting up such challenge. I will dive into those articles to get a better vision of the problem.\nRegards. " }, { "id": 500808, "postDate": "2019-03-26T15:23:32.070Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 455466, "author_name": "Elliot", "author_url": "", "post_date": "2019-01-14T00:58:58.703000", "content": "Hi Bertrand, thank you and your team for hosting this meaningful kaggle competition and sharing your work with us.\nI briefly looked through the above mentioned papers. All the model performances were evaluated using R2, effectively MSE. However in this competition, we use MAE as the performance metric. Could you share with us some insights regarding why we change the metric.
\n\nSincerely
", "votes": 22, "replies": [ { "id": 455649, "author_name": "Abhishek Thakur", "author_url": "", "post_date": "2019-01-14T09:44:11.623000", "content": "I would like to know this too :)
", "votes": 0, "replies": [] }, { "id": 456509, "author_name": "Bertrand RL", "author_url": "", "post_date": "2019-01-16T00:02:39.963000", "content": "Hi Eliott, thanks for participating in our competition. Machine learning is in its infancy in earth science and seismology and there is no benchmark or commonly used metric. We chose to go with MAE for the competition so that the metric is a physical quantity, time.
", "votes": 19, "replies": [] }, { "id": 522930, "author_name": "CPMP", "author_url": "", "post_date": "2019-04-25T08:46:13.167000", "content": "\n\n\nso that the metric is a physical quantity, time.
\n
Root mean squared error would also be time.
", "votes": 2, "replies": [] } ] }, { "id": 460846, "author_name": "ultragamza", "author_url": "", "post_date": "2019-01-24T14:33:19.437000", "content": "I have received the dataset mentioned in the paper(p4581 and p2394 dataset). I could see the acoustic data data similar to this competition, but I could not find the time to failure data.\nIs the time to failure data mentioned in the article publicly available?\nThanks.
", "votes": 3, "replies": [ { "id": 461814, "author_name": "Elliot", "author_url": "", "post_date": "2019-01-27T03:52:11.263000", "content": "I was looking at those two dataset as well. But up to my current understanding, they are of little use to our competition.
\n\nP4581, yes, was recorded using the same device that recorded our data. But the dataset was used for different purposes.
\n\nI could find two papers using P4581, one is to estimate friction, one is to identify abnormal event.
\n\nIn our competition, of course, you can imagine it as estimating both friction and identifying abnormal events, then using the results to estimate fault slip time. If you are building some complicated pipeline like this, the dataset might be useful to you..
\n\nIf the host can help check my understanding, will be grateful.
\n\nI do have features to pick up abnormal events anyway, but they show little/no contributions to my current model. They can't help improve the systematic estimation error occurs for the 3 lengthiest fault slips (time to failure starting with >13) in the training set. I'm afraid because of this, those who are implementing CNN/RNN would hardly find any surprises as well.
", "votes": 3, "replies": [] }, { "id": 466166, "author_name": "", "author_url": "", "post_date": "2019-02-04T19:48:16.647000", "content": "", "votes": 0, "replies": [] }, { "id": 466167, "author_name": "Bertrand RL", "author_url": "", "post_date": "2019-02-04T19:48:56.533000", "content": "Hi Elliott, p4581 is an experiment that has the same setup in terms of machine, material and recording apparatus. The stress conditions applied on the artificial faults were different however, resulting in very different earthquake cycles. Time to failure is identified from the friction and therefore using p4581 to train a model to predict friction could indeed help for the competition, since the properties of the seismic data scale with the stress conditions, as is shown is the second paper referenced above.
", "votes": 2, "replies": [] }, { "id": 467547, "author_name": "Elliot", "author_url": "", "post_date": "2019-02-07T09:58:22.903000", "content": "Thank you Bertrand, this explains it, crystal clear!
", "votes": 1, "replies": [] }, { "id": 500688, "author_name": "MrBurnst", "author_url": "", "post_date": "2019-03-26T12:23:55.513000", "content": "interesting. about experiment p4581 - I could not find the stress conditions or friction values in the files, only the acoustic signal. Are they at all available?
", "votes": 0, "replies": [] } ] }, { "id": 464841, "author_name": "ChrisWestland", "author_url": "", "post_date": "2019-02-01T16:03:12.020000", "content": "Bertrand, thank you for hosting this competition and congratulations on your Nature Geoscience publications. I had actually hoped, when I saw the announcement for this competition, that you would have spatio-temporal data (kind of like weather forecasting). Time-to-failure 'only' data makes this more like an exercise in technical stock-price forecasting (i.e., based on price time-series only, rather than news and financials). I'm guessing you will find a lot of artifacts from overfitting; my financial analogue would be Elliott Wave Theory. Anyway, I think I'll sit this one out, but if you ever post spatio-temporal tremors data, I would be keen to try my hand.
", "votes": 1, "replies": [] }, { "id": 486538, "author_name": "Neelima Mehta", "author_url": "", "post_date": "2019-03-08T23:46:36.207000", "content": "Hi Bertrand,
\n\nThis is an exciting challenge. I've been reviewing the various papers you've shared, and a few others from other sources that came up on USGS. I have a few basic questions (I apologize if I'm way off in my understanding here):
\n\nSome of the above papers refer to \"repeating earthquakes\" with \"continuous seismic signals\", while the data shared is for \"aperiodic earthquakes\". Is there any literature that has used the classic laboratory model to create these \"aperiodic earthquakes\"? It would be good to get some literature on these specific type of earthquakes to find a good way to frame the ML question, before applying a model.
In one of your responses below, you've mentioned using another data set to help determine friction that can be helpful in applying to this competition - can you please elaborate what you mean by that? Did you mean estimating the friction as a function of the acoustic signal?
Is the failure observed for this competition's dataset, a.k.a. \"aperiodic earthquakes\" due to stick-slip failure? If not, what is the technical term for the type of failure that triggers these aperiodic earthquakes?
Thanks!
", "votes": 2, "replies": [] }, { "id": 469542, "author_name": "林湧森 (Dyson Lin)", "author_url": "", "post_date": "2019-02-11T11:55:04.487000", "content": "You may be interested in my air voltage signals for quake forecast :-)\nhttps://www.facebook.com/dysondyson
", "votes": -2, "replies": [] }, { "id": 524782, "author_name": "Shashank Pathak", "author_url": "", "post_date": "2019-04-29T14:12:29.507000", "content": "What is the expected degree of precision?
", "votes": 0, "replies": [] }, { "id": 515364, "author_name": "Licheng Zhang", "author_url": "", "post_date": "2019-04-12T13:53:57.047000", "content": "Hi Bertrand,
\n\nThanks for the info. I wonder whether I can get information about the physical model of the experiments.
", "votes": 0, "replies": [] }, { "id": 497347, "author_name": "arpitr07", "author_url": "", "post_date": "2019-03-23T12:09:25.940000", "content": "Does the signal data provided contain displacement amplitudes or acceleration?
", "votes": 0, "replies": [] }, { "id": 478766, "author_name": "Olivier Moulin", "author_url": "", "post_date": "2019-02-26T15:51:35.620000", "content": "Great Challenge ... lot of possiblities and it will change a little bit of house price estimation, image recognition or malware classification ... Thanks a lot :)
", "votes": 0, "replies": [] }, { "id": 477084, "author_name": "cascio", "author_url": "", "post_date": "2019-02-23T20:42:16.970000", "content": "Bertrand, I enjoyed reading the articles you posted and am looking forward to participating in this competition.
", "votes": 0, "replies": [] }, { "id": 464981, "author_name": "JohnDillinger", "author_url": "", "post_date": "2019-02-01T23:52:27.643000", "content": "Hi Bertrand,\nwhat is the difference between the dataset provided in \"LANL Earthquake Prediction\"-Kaggle Challenge\nand the dataset which was used in the following paper:\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2017GL074677 ???
\n\nSame testbench? What is different about this experiment?
\n\n\"For this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly.\"\nCan you provide more details, please? Where is the difference between this experiment and the experiment in the paper above?
\n\nThanks in Advance
", "votes": 0, "replies": [ { "id": 466157, "author_name": "Bertrand RL", "author_url": "", "post_date": "2019-02-04T19:35:57.383000", "content": "Hi JohnDillinger, the experimental data comes from the same machine with the same fault material. The main difference is that the stress applied to the fault is lower, leading to more unstable and random lab earthquakes. An other difference is that the device used to record seismic data is much more sensitive, it is a piezoceramic sensor instead of an accelerometer.
", "votes": 1, "replies": [] } ] }, { "id": 464577, "author_name": "D.Baranzini", "author_url": "", "post_date": "2019-02-01T05:24:26.550000", "content": "a curiosity Bertrand, we know this prediction on a test set (after training whatever ensemble model will appear) would be the \"prediction of the century\"....nobody has ever predicted real (non laboratory) earthquakes ...so what brings you to believe that this important exercise will work here : ) .... and let me say that I do believe you and kaggle are doing the right thing : )
", "votes": 0, "replies": [] }, { "id": 460918, "author_name": "Matheus ", "author_url": "", "post_date": "2019-01-24T18:02:07.037000", "content": "Thank you and yout team for setting up such challenge. I will dive into those articles to get a better vision of the problem.\nRegards.
", "votes": 0, "replies": [] }, { "id": 500808, "author_name": "", "author_url": "", "post_date": "2019-03-26T15:23:32.070000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "455430": "Hello everyone, \n\nI am one of the Los Alamos researchers on the organizing team. We are a small group of machine learning scientists and geophysicists, and we have been using and developing ML methods on seismic data for the last 3 years.\n\nOur published work on the topic is in the following papers:\nhttps://doi.org/10.1002/2017GL074677\nhttps://doi.org/10.1002/2017GL076708\nhttps://rdcu.be/bdG8Y\nhttps://rdcu.be/bdG9r\n\nThe first 3 papers will explain our initial work on data similar to the one you have for the challenge. The 4th paper will show initial results on applying the methods developed on lab data on field data with success, on a particular type of earthquakes known as slow earthquakes.\n\nThe data for this challenge comes from a classic laboratory earthquake experiment, that has been studied in depth as a tabletop analog of seismogenic faults for decades. A number of physical laws widely used by the geoscience community have been derived from this earthquake machine.\n\nFor this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly. \n\nThe data from this classic earthquake machine had never been studied using ML until our early efforts 3 years ago, and much remains to be explored and discovered.\n\nMany thanks to Kaggle for hosting the competition *pro bono*, we are looking forward to seeing what the community can discover in our data!\n\nBertrand Rouet-Leduc\nhttps://twitter.com/bertrandrl\n\n\n\n\n\n", "455466": "Hi Bertrand, thank you and your team for hosting this meaningful kaggle competition and sharing your work with us.\nI briefly looked through the above mentioned papers. All the model performances were evaluated using R2, effectively MSE. However in this competition, we use MAE as the performance metric. Could you share with us some insights regarding why we change the metric.\n\nSincerely", "460846": "I have received the dataset mentioned in the paper(p4581 and p2394 dataset). I could see the acoustic data data similar to this competition, but I could not find the time to failure data.\nIs the time to failure data mentioned in the article publicly available?\nThanks.", "464841": "Bertrand, thank you for hosting this competition and congratulations on your Nature Geoscience publications. I had actually hoped, when I saw the announcement for this competition, that you would have spatio-temporal data (kind of like weather forecasting). Time-to-failure 'only' data makes this more like an exercise in technical stock-price forecasting (i.e., based on price time-series only, rather than news and financials). I'm guessing you will find a lot of artifacts from overfitting; my financial analogue would be Elliott Wave Theory. Anyway, I think I'll sit this one out, but if you ever post spatio-temporal tremors data, I would be keen to try my hand.", "486538": "Hi Bertrand,\n\nThis is an exciting challenge. I've been reviewing the various papers you've shared, and a few others from other sources that came up on USGS. I have a few basic questions (I apologize if I'm way off in my understanding here):\n\n1. Some of the above papers refer to \"repeating earthquakes\" with \"continuous seismic signals\", while the data shared is for \"aperiodic earthquakes\". Is there any literature that has used the classic laboratory model to create these \"aperiodic earthquakes\"? It would be good to get some literature on these specific type of earthquakes to find a good way to frame the ML question, before applying a model.\n\n2. In one of your responses below, you've mentioned using another data set to help determine friction that can be helpful in applying to this competition - can you please elaborate what you mean by that? Did you mean estimating the friction as a function of the acoustic signal?\n\n3. Is the failure observed for this competition's dataset, a.k.a. \"aperiodic earthquakes\" due to stick-slip failure? If not, what is the technical term for the type of failure that triggers these aperiodic earthquakes?\n\nThanks!", "469542": "You may be interested in my air voltage signals for quake forecast :-)\nhttps://www.facebook.com/dysondyson", "524782": "What is the expected degree of precision? ", "515364": "Hi Bertrand,\n\nThanks for the info. I wonder whether I can get information about the physical model of the experiments. ", "497347": "Does the signal data provided contain displacement amplitudes or acceleration?", "478766": "Great Challenge ... lot of possiblities and it will change a little bit of house price estimation, image recognition or malware classification ... Thanks a lot :)", "477084": "Bertrand, I enjoyed reading the articles you posted and am looking forward to participating in this competition.", "464981": "Hi Bertrand,\nwhat is the difference between the dataset provided in \"LANL Earthquake Prediction\"-Kaggle Challenge\nand the dataset which was used in the following paper:\nhttps://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2017GL074677 ???\n\nSame testbench? What is different about this experiment?\n\n\"For this challenge we selected an experiment that exhibits a very aperiodic and more realistic behavior compared to the data we studied in our early work, with earthquakes occurring very irregularly.\"\nCan you provide more details, please? Where is the difference between this experiment and the experiment in the paper above?\n\nThanks in Advance", "464577": "a curiosity Bertrand, we know this prediction on a test set (after training whatever ensemble model will appear) would be the \"prediction of the century\"....nobody has ever predicted real (non laboratory) earthquakes ...so what brings you to believe that this important exercise will work here : ) .... and let me say that I do believe you and kaggle are doing the right thing : ) ", "460918": "Thank you and yout team for setting up such challenge. I will dive into those articles to get a better vision of the problem.\nRegards. ", "500808": "" } }