{
  "id": 586581,
  "title": "Top 3 Solutions from 2024 NeurIPS Ariel Data Challenge summarized (in tabular format)",
  "url": "/competitions/ariel-data-challenge-2025/discussion/586581",
  "author_name": "Athar Sayed",
  "post_date": "2025-06-27T09:22:30.849000",
  "votes": 19,
  "comment_count": 2,
  "views": 0,
  "content": "<table>\n<thead>\n<tr>\n<th><strong>Rank</strong></th>\n<th><strong>Team / Key Approach</strong></th>\n<th><strong>Core Strategy &amp; Modeling</strong></th>\n<th><strong>Key Innovations / Insights</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>🥇 1st</td>\n<td><strong>\"Hacking\" the Simulator</strong><br>Gain Drift Fitting &amp; Foreground Processing + Ensemble (GPR, AutoEncoder, NMF)</td>\n<td>- Deep analysis of ExoSim2 and TauREx3 simulation code to uncover exact gain drift (1 + f(t)·g(λ)) and foreground signal.<br>- Two-stage analytical fitting for transit dip and star spectrum, incorporating the identified gain drift.<br>- Ensemble of Gaussian Process Regression, AutoEncoder, and Non-negative Matrix Factorization (NMF) for robust dip estimation.</td>\n<td>- Disabling hot pixel processing and foreground subtraction based on simulator code.<br>- Direct modeling of simulator’s noise/drift mechanisms.<br>- Bootstrapping for wavelength-dependent error estimation.</td>\n</tr>\n<tr>\n<td>🥈 2nd</td>\n<td><strong>Pure Bayesian Inference (BI) + Gaussian Processes (GPs)</strong></td>\n<td>- Full Bayesian Inference framework modeling detector noise, drift, transit behavior &amp; depth as formal distributions.<br>- GPs model smooth functions like drift and transit depth.<br>- Iterative linearization for solving the nonlinear Bayesian model.</td>\n<td>- Posterior-based uncertainty estimation.<br>- Clean separation of physics (prior) from math (solver).<br>- PCA on rough fits to extract shared depth shapes across planets.</td>\n</tr>\n<tr>\n<td>🥉 3rd</td>\n<td><strong>Polynomial Fitting + PCA</strong></td>\n<td>- Polynomial fitting of time-series (excluding ingress/egress) for detrending.<br>- Averaging across neighboring wavelengths to reduce noise.<br>- Adaptive polynomial degree to avoid overfitting.<br>- PCA applied per star to refine final spectra.</td>\n<td>- Empirical noise reduction: hot pixel retention, top 50% intensity selection, wavelength weighting, and dead pixel handling.<br>- Adaptive sigma estimation based on spectrum variability.<br>- Slight spectrum inflation helped, possibly due to foreground signal.</td>\n</tr>\n</tbody>\n</table>",
  "messages": [
    {
      "id": 3233879,
      "postDate": "2025-06-27T09:22:30.850Z",
      "content": "<table>\n<thead>\n<tr>\n<th><strong>Rank</strong></th>\n<th><strong>Team / Key Approach</strong></th>\n<th><strong>Core Strategy &amp; Modeling</strong></th>\n<th><strong>Key Innovations / Insights</strong></th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>🥇 1st</td>\n<td><strong>\"Hacking\" the Simulator</strong><br>Gain Drift Fitting &amp; Foreground Processing + Ensemble (GPR, AutoEncoder, NMF)</td>\n<td>- Deep analysis of ExoSim2 and TauREx3 simulation code to uncover exact gain drift (1 + f(t)·g(λ)) and foreground signal.<br>- Two-stage analytical fitting for transit dip and star spectrum, incorporating the identified gain drift.<br>- Ensemble of Gaussian Process Regression, AutoEncoder, and Non-negative Matrix Factorization (NMF) for robust dip estimation.</td>\n<td>- Disabling hot pixel processing and foreground subtraction based on simulator code.<br>- Direct modeling of simulator’s noise/drift mechanisms.<br>- Bootstrapping for wavelength-dependent error estimation.</td>\n</tr>\n<tr>\n<td>🥈 2nd</td>\n<td><strong>Pure Bayesian Inference (BI) + Gaussian Processes (GPs)</strong></td>\n<td>- Full Bayesian Inference framework modeling detector noise, drift, transit behavior &amp; depth as formal distributions.<br>- GPs model smooth functions like drift and transit depth.<br>- Iterative linearization for solving the nonlinear Bayesian model.</td>\n<td>- Posterior-based uncertainty estimation.<br>- Clean separation of physics (prior) from math (solver).<br>- PCA on rough fits to extract shared depth shapes across planets.</td>\n</tr>\n<tr>\n<td>🥉 3rd</td>\n<td><strong>Polynomial Fitting + PCA</strong></td>\n<td>- Polynomial fitting of time-series (excluding ingress/egress) for detrending.<br>- Averaging across neighboring wavelengths to reduce noise.<br>- Adaptive polynomial degree to avoid overfitting.<br>- PCA applied per star to refine final spectra.</td>\n<td>- Empirical noise reduction: hot pixel retention, top 50% intensity selection, wavelength weighting, and dead pixel handling.<br>- Adaptive sigma estimation based on spectrum variability.<br>- Slight spectrum inflation helped, possibly due to foreground signal.</td>\n</tr>\n</tbody>\n</table>",
      "rawMarkdown": "| **Rank** | **Team / Key Approach** | **Core Strategy & Modeling** | **Key Innovations / Insights** |\n|:--------:|--------------------------|-------------------------------|--------------------------------|\n| 🥇 1st | **\"Hacking\" the Simulator**<br>Gain Drift Fitting & Foreground Processing + Ensemble (GPR, AutoEncoder, NMF) | - Deep analysis of ExoSim2 and TauREx3 simulation code to uncover exact gain drift (1 + f(t)·g(λ)) and foreground signal.<br>- Two-stage analytical fitting for transit dip and star spectrum, incorporating the identified gain drift.<br>- Ensemble of Gaussian Process Regression, AutoEncoder, and Non-negative Matrix Factorization (NMF) for robust dip estimation. | - Disabling hot pixel processing and foreground subtraction based on simulator code.<br>- Direct modeling of simulator’s noise/drift mechanisms.<br>- Bootstrapping for wavelength-dependent error estimation. |\n| 🥈 2nd | **Pure Bayesian Inference (BI) + Gaussian Processes (GPs)** | - Full Bayesian Inference framework modeling detector noise, drift, transit behavior & depth as formal distributions.<br>- GPs model smooth functions like drift and transit depth.<br>- Iterative linearization for solving the nonlinear Bayesian model. | - Posterior-based uncertainty estimation.<br>- Clean separation of physics (prior) from math (solver).<br>- PCA on rough fits to extract shared depth shapes across planets. |\n| 🥉 3rd | **Polynomial Fitting + PCA** | - Polynomial fitting of time-series (excluding ingress/egress) for detrending.<br>- Averaging across neighboring wavelengths to reduce noise.<br>- Adaptive polynomial degree to avoid overfitting.<br>- PCA applied per star to refine final spectra. | - Empirical noise reduction: hot pixel retention, top 50% intensity selection, wavelength weighting, and dead pixel handling.<br>- Adaptive sigma estimation based on spectrum variability.<br>- Slight spectrum inflation helped, possibly due to foreground signal. |\n",
      "votes": 18
    },
    {
      "id": 3233884,
      "postDate": "2025-06-27T09:25:38.163Z",
      "content": "<p>nice summary :)! </p>",
      "rawMarkdown": "nice summary :)! ",
      "votes": 1
    },
    {
      "id": 3233883,
      "postDate": "2025-06-27T09:24:17.533Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 3233884,
      "author_name": "Gordon Yip",
      "author_url": "",
      "post_date": "2025-06-27T09:25:38.163000",
      "content": "<p>nice summary :)! </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 3233883,
      "author_name": "",
      "author_url": "",
      "post_date": "2025-06-27T09:24:17.533000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "3233879": "| **Rank** | **Team / Key Approach** | **Core Strategy & Modeling** | **Key Innovations / Insights** |\n|:--------:|--------------------------|-------------------------------|--------------------------------|\n| 🥇 1st | **\"Hacking\" the Simulator**<br>Gain Drift Fitting & Foreground Processing + Ensemble (GPR, AutoEncoder, NMF) | - Deep analysis of ExoSim2 and TauREx3 simulation code to uncover exact gain drift (1 + f(t)·g(λ)) and foreground signal.<br>- Two-stage analytical fitting for transit dip and star spectrum, incorporating the identified gain drift.<br>- Ensemble of Gaussian Process Regression, AutoEncoder, and Non-negative Matrix Factorization (NMF) for robust dip estimation. | - Disabling hot pixel processing and foreground subtraction based on simulator code.<br>- Direct modeling of simulator’s noise/drift mechanisms.<br>- Bootstrapping for wavelength-dependent error estimation. |\n| 🥈 2nd | **Pure Bayesian Inference (BI) + Gaussian Processes (GPs)** | - Full Bayesian Inference framework modeling detector noise, drift, transit behavior & depth as formal distributions.<br>- GPs model smooth functions like drift and transit depth.<br>- Iterative linearization for solving the nonlinear Bayesian model. | - Posterior-based uncertainty estimation.<br>- Clean separation of physics (prior) from math (solver).<br>- PCA on rough fits to extract shared depth shapes across planets. |\n| 🥉 3rd | **Polynomial Fitting + PCA** | - Polynomial fitting of time-series (excluding ingress/egress) for detrending.<br>- Averaging across neighboring wavelengths to reduce noise.<br>- Adaptive polynomial degree to avoid overfitting.<br>- PCA applied per star to refine final spectra. | - Empirical noise reduction: hot pixel retention, top 50% intensity selection, wavelength weighting, and dead pixel handling.<br>- Adaptive sigma estimation based on spectrum variability.<br>- Slight spectrum inflation helped, possibly due to foreground signal. |\n",
    "3233884": "nice summary :)! ",
    "3233883": ""
  }
}