Predicting Exoplanetary Features with a Residual Model for Uniform and Gaussian Distributions
\nAuthor: Andrew Sweet
\n\"In order to help bridge the gap between machine learning and astrophysics domain experts, the 2023 Ariel Data Challenge was hosted to predict posterior distributions of 7 exoplanetary features. The procedure outlined in this paper leveraged a combination of two deep learning models to address this challenge: a Multivariate Gaussian model that generates the mean and covariance matrix of a multivariate Gaussian distribution, and a Uniform Quantile model that predicts quantiles for use as the upper and lower bounds of a uniform distribution.\"
\n\"Training of the Multivariate Gaussian model was found to be unstable, while training of the Uniform Quantile model was stable.An ensemble of uniform distributions was found to have competitive results during testing (posterior score of 696.43), and when combined with a multivariate Gaussian distribution achieved a final rank of third in the 2023 Ariel Data Challenge (final score of 681.57).\"
\nSpectral and Auxiliary data
\n\"The Ariel Data Challenge at NeurIPS 2022 and ECML PPKD 2023 had three main differences: the size of the provided data sets, the number of exoplanetary features being predicted, and the scoring metrics. The methods here will focus on the data for ECML PPKD 2023. There were five data files distributed for the challenge, which can broadly be separated into input and output data for machine learning purposes, and contain simulated data for 41,423 (denoted as N below) planets. For input data there were two files:\"
\n– spectral data: which was composed of the wavelength grid, spectrum, uncertainty and bin width across 52 wavelength channels, of shape N × 4 × 52.\"
\n– auxiliary data: containing 8 auxiliary features for each planetary system, of shape N × 8. These features were the planet’s mass, orbital period, radius, semi-major axis, and surface gravity, as well as it’s host star’s mass, radius, temperature, and distance from Earth.\"
\n\"A few suggestions for future work include hyperparameter optimization, separate backbones, alternate feature engineering and feature selection, transfer learning with a pre-trained unsupervised learning scheme such as an encoder-decoder network, and combining the output distributions as part of the learning process such as with a Mixture-of-Experts (MoE).
\nhttps://arxiv.org/abs/2406.10771
\nhttps://astrobiology.com/tag/2023-ariel-data-challenge
\nSimulation-based Inference for Exoplanet Atmospheric Retrieval: Insights from winning the Ariel Data Challenge 2023 using Normalizing Flows
\nAuthors: Mayeul Aubin, Carolina Cuesta-Lazaro, Ethan Tregidga, Javier Viaña, Cecilia Garraffo, Iouli E. Gordon, Mercedes López-Morales, Robert J. Hargreaves, Vladimir Yu. Makhnev, Jeremy J. Drake, Douglas P. Finkbeiner, and Phillip Cargile
\nRetrieving Exoplanet Atmospheric Compositions - Transmission Spectroscopy
\n\"The most commonly used method to study the atmospheric composition of an exoplanet is called transmission spectroscopy, which consists of measuring how light from the host star gets absorbed by the planetary atmosphere during a transit, i.e. when the planet crosses in front of the disk of its star.\"
\n\"Retrieving the atmospheric properties of exoplanets from their transmission spectra is challenging. There have been many efforts in this direction. Due to the high dimensionality of the parameter space and the low-resolution spectra at hand, there are degeneracies, and a deterministic answer is not usually possible or informative. A number of codes, most based on Bayesian sampling algorithms but some on machine learning too, have been developed to do this.\"
\n\"In their study, the authors utilized Neural Spline Flows to model the posterior distribution of atmospheric parameters given the observed spectra. Neural Spline Flows, employ monotonic rational-quadratic splines to model the invertible mapping, and neural networks to predict the necessary parameters of these transformations. To implement Neural Spline Flows, they utilized the Zuko python package.\"
\nEnsembling the best models:
\n\"Once the hyperparameters optimization was complete, the authors ensembled
\nthe 10 best models to reduce model’s errors and increase robustness.\"
The Ariel Data Challenge 2023 Solution:
\nhttps://github.com/AstroAI-CfA/Ariel_Data_Challenge_2023_solution
\ntheanets.losses.GaussianLogLikelihood
\nWhy we consider log likelihood instead of Likelihood in Gaussian Distribution
\nCoding a Gaussian log-likelihood in R
\nAriel Gaussian Log Likelihood By Sohier Dane and Kaggle Competition Metrics (aka Kaggle Bot!)
", "messages": [ { "id": 2943916, "postDate": "2024-08-02T02:11:02.617Z", "content": "Predicting Exoplanetary Features with a Residual Model for Uniform and Gaussian Distributions
\nAuthor: Andrew Sweet
\n\"In order to help bridge the gap between machine learning and astrophysics domain experts, the 2023 Ariel Data Challenge was hosted to predict posterior distributions of 7 exoplanetary features. The procedure outlined in this paper leveraged a combination of two deep learning models to address this challenge: a Multivariate Gaussian model that generates the mean and covariance matrix of a multivariate Gaussian distribution, and a Uniform Quantile model that predicts quantiles for use as the upper and lower bounds of a uniform distribution.\"
\n\"Training of the Multivariate Gaussian model was found to be unstable, while training of the Uniform Quantile model was stable.An ensemble of uniform distributions was found to have competitive results during testing (posterior score of 696.43), and when combined with a multivariate Gaussian distribution achieved a final rank of third in the 2023 Ariel Data Challenge (final score of 681.57).\"
\nSpectral and Auxiliary data
\n\"The Ariel Data Challenge at NeurIPS 2022 and ECML PPKD 2023 had three main differences: the size of the provided data sets, the number of exoplanetary features being predicted, and the scoring metrics. The methods here will focus on the data for ECML PPKD 2023. There were five data files distributed for the challenge, which can broadly be separated into input and output data for machine learning purposes, and contain simulated data for 41,423 (denoted as N below) planets. For input data there were two files:\"
\n– spectral data: which was composed of the wavelength grid, spectrum, uncertainty and bin width across 52 wavelength channels, of shape N × 4 × 52.\"
\n– auxiliary data: containing 8 auxiliary features for each planetary system, of shape N × 8. These features were the planet’s mass, orbital period, radius, semi-major axis, and surface gravity, as well as it’s host star’s mass, radius, temperature, and distance from Earth.\"
\n\"A few suggestions for future work include hyperparameter optimization, separate backbones, alternate feature engineering and feature selection, transfer learning with a pre-trained unsupervised learning scheme such as an encoder-decoder network, and combining the output distributions as part of the learning process such as with a Mixture-of-Experts (MoE).
\nhttps://arxiv.org/abs/2406.10771
\nhttps://astrobiology.com/tag/2023-ariel-data-challenge
\nSimulation-based Inference for Exoplanet Atmospheric Retrieval: Insights from winning the Ariel Data Challenge 2023 using Normalizing Flows
\nAuthors: Mayeul Aubin, Carolina Cuesta-Lazaro, Ethan Tregidga, Javier Viaña, Cecilia Garraffo, Iouli E. Gordon, Mercedes López-Morales, Robert J. Hargreaves, Vladimir Yu. Makhnev, Jeremy J. Drake, Douglas P. Finkbeiner, and Phillip Cargile
\nRetrieving Exoplanet Atmospheric Compositions - Transmission Spectroscopy
\n\"The most commonly used method to study the atmospheric composition of an exoplanet is called transmission spectroscopy, which consists of measuring how light from the host star gets absorbed by the planetary atmosphere during a transit, i.e. when the planet crosses in front of the disk of its star.\"
\n\"Retrieving the atmospheric properties of exoplanets from their transmission spectra is challenging. There have been many efforts in this direction. Due to the high dimensionality of the parameter space and the low-resolution spectra at hand, there are degeneracies, and a deterministic answer is not usually possible or informative. A number of codes, most based on Bayesian sampling algorithms but some on machine learning too, have been developed to do this.\"
\n\"In their study, the authors utilized Neural Spline Flows to model the posterior distribution of atmospheric parameters given the observed spectra. Neural Spline Flows, employ monotonic rational-quadratic splines to model the invertible mapping, and neural networks to predict the necessary parameters of these transformations. To implement Neural Spline Flows, they utilized the Zuko python package.\"
\nEnsembling the best models:
\n\"Once the hyperparameters optimization was complete, the authors ensembled
\nthe 10 best models to reduce model’s errors and increase robustness.\"
The Ariel Data Challenge 2023 Solution:
\nhttps://github.com/AstroAI-CfA/Ariel_Data_Challenge_2023_solution
\ntheanets.losses.GaussianLogLikelihood
\nWhy we consider log likelihood instead of Likelihood in Gaussian Distribution
\nCoding a Gaussian log-likelihood in R
\nAriel Gaussian Log Likelihood By Sohier Dane and Kaggle Competition Metrics (aka Kaggle Bot!)
", "rawMarkdown": "#Predicting Exoplanetary Features\n\nPredicting Exoplanetary Features with a Residual Model for Uniform and Gaussian Distributions\n\nAuthor: Andrew Sweet\n\n\"In order to help bridge the gap between machine learning and astrophysics domain experts, the 2023 **Ariel Data Challenge** was hosted to predict posterior distributions of 7 exoplanetary features. The procedure outlined in this paper leveraged a combination of two deep learning models to address this challenge: a Multivariate Gaussian model that generates the mean and covariance matrix of a multivariate Gaussian distribution, and a Uniform Quantile model that predicts quantiles for use as the upper and lower bounds of a uniform distribution.\"\n\n\"Training of the Multivariate Gaussian model was found to be unstable, while training of the Uniform Quantile model was stable.An ensemble of uniform distributions was found to have competitive results during testing (posterior score of 696.43), and when combined with a multivariate Gaussian distribution achieved a final rank of third in the 2023 Ariel Data Challenge (final score of 681.57).\"\n\n**Spectral and Auxiliary data**\n\n\"The Ariel Data Challenge at NeurIPS 2022 and ECML PPKD 2023 had three main differences: the size of the provided data sets, the number of exoplanetary features being predicted, and the scoring metrics. The methods here will focus on the data for ECML PPKD 2023. There were five data files distributed for the challenge, which can broadly be separated into input and output data for machine learning purposes, and contain simulated data for 41,423 (denoted as N below) planets. For input data there were two files:\"\n\n– spectral data: which was composed of the wavelength grid, spectrum, uncertainty and bin width across 52 wavelength channels, of shape N × 4 × 52.\"\n\n– auxiliary data: containing 8 auxiliary features for each planetary system, of shape N × 8. These features were the planet’s mass, orbital period, radius, semi-major axis, and surface gravity, as well as it’s host star’s mass, radius, temperature, and distance from Earth.\"\n\n\"A few suggestions for future work include hyperparameter optimization, separate backbones, alternate feature engineering and feature selection, transfer learning with a pre-trained unsupervised learning scheme such as an encoder-decoder network, and combining the output distributions as part of the learning process such as with a **Mixture-of-Experts (MoE)**.\n\nhttps://arxiv.org/abs/2406.10771\n\nhttps://astrobiology.com/tag/2023-ariel-data-challenge\n\n#ADC 2023 - Exoplanet Atmospheric Retrieval\n\nSimulation-based Inference for Exoplanet Atmospheric Retrieval: Insights from winning the Ariel Data Challenge 2023 using Normalizing Flows\n\nAuthors: Mayeul Aubin, Carolina Cuesta-Lazaro, Ethan Tregidga, Javier Viaña, Cecilia Garraffo, Iouli E. Gordon, Mercedes López-Morales, Robert J. Hargreaves, Vladimir Yu. Makhnev, Jeremy J. Drake, Douglas P. Finkbeiner, and Phillip Cargile\n\nRetrieving Exoplanet Atmospheric Compositions - Transmission Spectroscopy\n\n\"The most commonly used method to study the atmospheric composition of an exoplanet is called **transmission spectroscopy**, which consists of measuring how light from the host star gets absorbed by the planetary atmosphere during a transit, i.e. when the planet crosses in front of the disk of its star.\"\n\n\"Retrieving the atmospheric properties of exoplanets from their transmission spectra is challenging. There have been many efforts in this direction. Due to the high dimensionality of the parameter space and the low-resolution spectra at hand, there are degeneracies, and a deterministic answer is not usually possible or informative. A number of codes, most based on Bayesian sampling algorithms but some on machine learning too, have been developed to do this.\"\n\n\"In their study, the authors utilized Neural Spline Flows to model the posterior distribution of atmospheric parameters given the observed spectra. Neural Spline Flows, employ monotonic rational-quadratic splines to model the invertible mapping, and neural networks to predict the necessary parameters of these transformations. To implement Neural Spline Flows, they utilized the **Zuko python package**.\"\n\nEnsembling the best models:\n\n\"Once the hyperparameters optimization was complete, the authors ensembled\nthe 10 best models to reduce model’s errors and increase robustness.\"\n\n**The Ariel Data Challenge 2023 Solution:**\n\nhttps://github.com/AstroAI-CfA/Ariel_Data_Challenge_2023_solution\n\n#Metric: Gaussian Log-Likelihood\n\n[theanets.losses.GaussianLogLikelihood](https://theanets.readthedocs.io/en/stable/api/generated/theanets.losses.GaussianLogLikelihood.html)\n\n[Why we consider log likelihood instead of Likelihood in Gaussian Distribution](https://math.stackexchange.com/questions/892832/why-we-consider-log-likelihood-instead-of-likelihood-in-gaussian-distribution)\n\n[Coding a Gaussian log-likelihood in R](https://stackoverflow.com/questions/44981549/coding-a-gaussian-log-likelihood-in-r)\n\n[Ariel Gaussian Log Likelihood](https://www.kaggle.com/code/metric/ariel-gaussian-log-likelihood/notebook) By Sohier Dane and Kaggle Competition Metrics (aka Kaggle Bot!)", "votes": 40 }, { "id": 2944691, "postDate": "2024-08-02T17:24:11.257Z", "content": "Thanks for sharing the detailed overview! @mpwolke
", "rawMarkdown": "Thanks for sharing the detailed overview! @mpwolke ", "votes": 1, "replies": [ { "id": 2944947, "postDate": "2024-08-02T22:00:11.753Z", "content": "I'm glad that you read the details. There were much more, however I'd rather make it shorter so that the Audience could read till the end.
", "rawMarkdown": "I'm glad that you read the details. There were much more, however I'd rather make it shorter so that the Audience could read till the end. ", "votes": 1 } ] }, { "id": 2943920, "postDate": "2024-08-02T02:31:47.103Z", "content": "Thanks for the detailed breakdown. The use of both Multivariate Gaussian and Uniform Quantile models is intriguing, and the performance metrics highlight the effectiveness of ensemble methods. The Gaussian Log-Likelihood metric is a solid choice for stability in evaluation. Looking forward to exploring future enhancements like hyperparameter tuning and advanced feature engineering
", "rawMarkdown": "Thanks for the detailed breakdown. The use of both Multivariate Gaussian and Uniform Quantile models is intriguing, and the performance metrics highlight the effectiveness of ensemble methods. The Gaussian Log-Likelihood metric is a solid choice for stability in evaluation. Looking forward to exploring future enhancements like hyperparameter tuning and advanced feature engineering", "votes": 2, "replies": [ { "id": 2944378, "postDate": "2024-08-02T11:37:41.167Z", "content": "Thank you for the support Abullrahman. It's always great trying to read/summarize/post something about the Competition metric. The learning path is pleasant and endless.
", "rawMarkdown": "Thank you for the support Abullrahman. It's always great trying to read/summarize/post something about the Competition metric. The learning path is pleasant and endless.", "votes": 2 } ] }, { "id": 2944034, "postDate": "2024-08-02T05:36:26.670Z", "content": "thanks for sharing @mpwolke
", "rawMarkdown": "thanks for sharing @mpwolke ", "votes": 2 }, { "id": 2949975, "postDate": "2024-08-07T05:41:22.807Z", "content": "thanks a lot for sharing!!!
", "rawMarkdown": "thanks a lot for sharing!!!" } ], "comments": [ { "id": 2944691, "author_name": "", "author_url": "", "post_date": "2024-08-02T17:24:11.257000", "content": "Thanks for sharing the detailed overview! @mpwolke
", "votes": 1, "replies": [ { "id": 2944947, "author_name": "Marília Prata", "author_url": "", "post_date": "2024-08-02T22:00:11.753000", "content": "I'm glad that you read the details. There were much more, however I'd rather make it shorter so that the Audience could read till the end.
", "votes": 1, "replies": [] } ] }, { "id": 2943920, "author_name": "Matin Maki Abdullrahman", "author_url": "", "post_date": "2024-08-02T02:31:47.103000", "content": "Thanks for the detailed breakdown. The use of both Multivariate Gaussian and Uniform Quantile models is intriguing, and the performance metrics highlight the effectiveness of ensemble methods. The Gaussian Log-Likelihood metric is a solid choice for stability in evaluation. Looking forward to exploring future enhancements like hyperparameter tuning and advanced feature engineering
", "votes": 2, "replies": [ { "id": 2944378, "author_name": "Marília Prata", "author_url": "", "post_date": "2024-08-02T11:37:41.167000", "content": "Thank you for the support Abullrahman. It's always great trying to read/summarize/post something about the Competition metric. The learning path is pleasant and endless.
", "votes": 2, "replies": [] } ] }, { "id": 2944034, "author_name": "Aadit Shukla", "author_url": "", "post_date": "2024-08-02T05:36:26.670000", "content": "thanks for sharing @mpwolke
", "votes": 2, "replies": [] }, { "id": 2949975, "author_name": "Parisa Karimi Darabi", "author_url": "", "post_date": "2024-08-07T05:41:22.807000", "content": "thanks a lot for sharing!!!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2943916": "#Predicting Exoplanetary Features\n\nPredicting Exoplanetary Features with a Residual Model for Uniform and Gaussian Distributions\n\nAuthor: Andrew Sweet\n\n\"In order to help bridge the gap between machine learning and astrophysics domain experts, the 2023 **Ariel Data Challenge** was hosted to predict posterior distributions of 7 exoplanetary features. The procedure outlined in this paper leveraged a combination of two deep learning models to address this challenge: a Multivariate Gaussian model that generates the mean and covariance matrix of a multivariate Gaussian distribution, and a Uniform Quantile model that predicts quantiles for use as the upper and lower bounds of a uniform distribution.\"\n\n\"Training of the Multivariate Gaussian model was found to be unstable, while training of the Uniform Quantile model was stable.An ensemble of uniform distributions was found to have competitive results during testing (posterior score of 696.43), and when combined with a multivariate Gaussian distribution achieved a final rank of third in the 2023 Ariel Data Challenge (final score of 681.57).\"\n\n**Spectral and Auxiliary data**\n\n\"The Ariel Data Challenge at NeurIPS 2022 and ECML PPKD 2023 had three main differences: the size of the provided data sets, the number of exoplanetary features being predicted, and the scoring metrics. The methods here will focus on the data for ECML PPKD 2023. There were five data files distributed for the challenge, which can broadly be separated into input and output data for machine learning purposes, and contain simulated data for 41,423 (denoted as N below) planets. For input data there were two files:\"\n\n– spectral data: which was composed of the wavelength grid, spectrum, uncertainty and bin width across 52 wavelength channels, of shape N × 4 × 52.\"\n\n– auxiliary data: containing 8 auxiliary features for each planetary system, of shape N × 8. These features were the planet’s mass, orbital period, radius, semi-major axis, and surface gravity, as well as it’s host star’s mass, radius, temperature, and distance from Earth.\"\n\n\"A few suggestions for future work include hyperparameter optimization, separate backbones, alternate feature engineering and feature selection, transfer learning with a pre-trained unsupervised learning scheme such as an encoder-decoder network, and combining the output distributions as part of the learning process such as with a **Mixture-of-Experts (MoE)**.\n\nhttps://arxiv.org/abs/2406.10771\n\nhttps://astrobiology.com/tag/2023-ariel-data-challenge\n\n#ADC 2023 - Exoplanet Atmospheric Retrieval\n\nSimulation-based Inference for Exoplanet Atmospheric Retrieval: Insights from winning the Ariel Data Challenge 2023 using Normalizing Flows\n\nAuthors: Mayeul Aubin, Carolina Cuesta-Lazaro, Ethan Tregidga, Javier Viaña, Cecilia Garraffo, Iouli E. Gordon, Mercedes López-Morales, Robert J. Hargreaves, Vladimir Yu. Makhnev, Jeremy J. Drake, Douglas P. Finkbeiner, and Phillip Cargile\n\nRetrieving Exoplanet Atmospheric Compositions - Transmission Spectroscopy\n\n\"The most commonly used method to study the atmospheric composition of an exoplanet is called **transmission spectroscopy**, which consists of measuring how light from the host star gets absorbed by the planetary atmosphere during a transit, i.e. when the planet crosses in front of the disk of its star.\"\n\n\"Retrieving the atmospheric properties of exoplanets from their transmission spectra is challenging. There have been many efforts in this direction. Due to the high dimensionality of the parameter space and the low-resolution spectra at hand, there are degeneracies, and a deterministic answer is not usually possible or informative. A number of codes, most based on Bayesian sampling algorithms but some on machine learning too, have been developed to do this.\"\n\n\"In their study, the authors utilized Neural Spline Flows to model the posterior distribution of atmospheric parameters given the observed spectra. Neural Spline Flows, employ monotonic rational-quadratic splines to model the invertible mapping, and neural networks to predict the necessary parameters of these transformations. To implement Neural Spline Flows, they utilized the **Zuko python package**.\"\n\nEnsembling the best models:\n\n\"Once the hyperparameters optimization was complete, the authors ensembled\nthe 10 best models to reduce model’s errors and increase robustness.\"\n\n**The Ariel Data Challenge 2023 Solution:**\n\nhttps://github.com/AstroAI-CfA/Ariel_Data_Challenge_2023_solution\n\n#Metric: Gaussian Log-Likelihood\n\n[theanets.losses.GaussianLogLikelihood](https://theanets.readthedocs.io/en/stable/api/generated/theanets.losses.GaussianLogLikelihood.html)\n\n[Why we consider log likelihood instead of Likelihood in Gaussian Distribution](https://math.stackexchange.com/questions/892832/why-we-consider-log-likelihood-instead-of-likelihood-in-gaussian-distribution)\n\n[Coding a Gaussian log-likelihood in R](https://stackoverflow.com/questions/44981549/coding-a-gaussian-log-likelihood-in-r)\n\n[Ariel Gaussian Log Likelihood](https://www.kaggle.com/code/metric/ariel-gaussian-log-likelihood/notebook) By Sohier Dane and Kaggle Competition Metrics (aka Kaggle Bot!)", "2944691": "Thanks for sharing the detailed overview! @mpwolke ", "2943920": "Thanks for the detailed breakdown. The use of both Multivariate Gaussian and Uniform Quantile models is intriguing, and the performance metrics highlight the effectiveness of ensemble methods. The Gaussian Log-Likelihood metric is a solid choice for stability in evaluation. Looking forward to exploring future enhancements like hyperparameter tuning and advanced feature engineering", "2944034": "thanks for sharing @mpwolke ", "2949975": "thanks a lot for sharing!!!" } }