In this presentation, We will detail our solution to the Ariel Data Challenge 2024, focusing on exoplanet spectrum calibration and prediction. The challenge included handling noisy pixel data and accurately predicting transit boundaries, where we applied Gaussian smoothing along the frequency axis to mitigate outliers. To improve transit zone predictions, we used Savitzky-Golay smoothing in the time direction, ensuring flux behavior aligned with expected transit patterns. For feature construction, we built a polynomial model to align the transit and non-transit parts of the spectrum. By using a combination of genetic algorithms and manual selection, we identified key frequency intervals, improving model performance by targeting relevant absorption zones for gases such as CH4, H2O, and CO2.

The solution employed both heuristic models and convolutional neural networks (CNN). The heuristic model used manually tuned weights for prediction, while the CNN model predicted spectrum points and average sigma, with careful tuning to avoid overfitting, including a window size of 21 frequencies. The final model achieved 0.692 on the public leaderboard, blending heuristic and CNN approaches to enhance prediction accuracy. This talk will provide an in-depth look at the methods used to achieve these results.