1. Introduction
The transit method has become one of the most widely used techniques for detecting and characterizing exoplanets, not only for its sensitivity but also because transits can be observed from ground-based facilities Perryman (Reference Perryman2018). However, it remains highly affected by instrumental noise, atmospheric distortions, and incomplete ground-based coverage, which introduce gaps and systematics that limit parameter accuracy.
To address these challenges, Generative AI offers promising tools for denoising and characterizing transit signals. Our previous framework, GANexo v1.0, was limited to synthetic transits, serving as a proof of concept for exoplanet analysis. In this work, we present preliminary results from GANexo v2.0, an improved framework that integrates real observational data and represents a critical step toward more robust and reliable exoplanet characterization. For a detailed description of GANEXO v1.0 architecture, we refer the reader to Meza-Obando & Crespo-Mariño (Reference Meza-Obando and Crespo-Mariño2024).
2. GANexo v2.0: Innovations and Architecture
Figure 1 compares GANexo v1.0 and v2.0. While v1.0 was limited to synthetic signal generation, v2.0 integrates real observational data, enabling operation under realistic conditions where noise and data gaps are unavoidable.The updated version includes fine-tuning of both the autoencoder and GAN components for improved noise reduction and generative stability, and introduces direct extraction of transit parameters through GAN-based regression.
Comparison of GANexo v1.0 on the left with synthetic transits from Meza-Obando & Crespo-Mariño (Reference Meza-Obando and Crespo-Mariño2024), and GANexo v2.0 on the right with real data.
Figure 1 Long description
Panel A: A flowchart showing the process of GANexo v1.0. It starts with synthetic transit data, which passes through a noise model. The noisy data is then processed by an autoencoder for noise removal. The cleaned data undergoes a conditional Generative Adversarial Network (cGAN) regression fit to characterize exoplanets. Panel B: A flowchart showing the process of GANexo v2.0. It starts with real data from the Hubble Space Telescope, Exoplanet Archive, and other sources. The data is processed by an autoencoder for noise removal. The cleaned data undergoes a cGAN regression fit to characterize exoplanets.
The architecture is structured in two main stages. First, autoencoder-based filtering is applied to reduce both white and red noise, enhancing the clarity of transit signals. Second, a one-dimensional conditional GAN performs generative regression to estimate critical planetary parameters such as transit depth and duration. Together, these stages improve robustness against observational noise and enhance the reliability of transit-based characterization. Data for this study were retrieved from the NASA Exoplanet Archive Akeson et al. (Reference Akeson and Chen2013).
2.1. Results
GANexo v2.0 has so far been tested on three randomly selected exoplanet cases: HAT-P-3 b, XO-7 b, and HAT-P-37 b, focusing on the estimation of planetary radius and semi-major axis. Results are summarized in Table 1. Figure 2 shows HAT-P-3 b, where the original signal (left) is processed by the autoencoder to produce a denoised and more distinct transit profile (middle), and then refined through cGAN regression (right) for accurate characterization. Additional cases are currently being tested.
Comparison of some exoplanet parameters from GANexo and the NASA Exoplanet Archive.
Table 1 Long description
A table comparing exoplanet parameters from GANexo and NASA Exoplanet Database. The table has three rows and four columns. Column headers are Exoplanet, GANEXO Exoplanet Radius (R⊕), GANEXO Semi-major Axis (AU), NASA EXOPLANET DATABASE Exoplanet Radius (R⊕), NASA EXOPLANET DATABASE Semi-major Axis (AU). Row 1: HAT-P-3 b, 10.96, 0.040, 10.53, 0.038. Row 2: XO-7 b, 14.92, 0.044, 15.38, 0.044. Row 3: HAT-P-37 b, 12.95, 0.037, 12.74, 0.037.
HAT-P-3 b transit: raw light curve (left), denoised signal with autoencoder (center), and cGAN regression for dip measurement (right).
2.2. Discussion
GANexo v2.0 extends our earlier synthetic-based framework to real observational data, with fine-tuning of both the autoencoder and cGAN components yielding improved adaptation to noisy and incomplete light curves. These refinements enhance noise suppression and parameter regression, resulting in more accurate and robust exoplanet characterization under realistic observational conditions.
Looking ahead, future developments will focus on extending GANexo v2.0 to multi-telescope datasets, adapting it for space-based missions, and conducting systematic tests with ongoing ground-based measurements. In parallel, the possibility of reintroducing synthetic generation is being explored as a way to reconstruct missing parts of ground-based transits when interruptions occur. Together, these directions aim to consolidate GANexo as a versatile and reliable tool for exoplanet transit analysis.
