This paper proposes TransientGAN, a novel conditional Wasserstein Generative Adversarial Network (cWGAN) framework for probabilistic forecasting of multivariate astronomical light curves. It combines a temporal encoder-decoder architecture with adversarial training to generate realistic future observations conditioned on historical data, enabling early-time prediction of supernova peak magnitudes, fade rates, and temporal classification.
Key findings
TransientGAN uses a cWGAN to predict future astronomical events from sparse, early-time observations.
The model includes a temporal encoder-decoder architecture with attention mechanisms for handling irregular, multivariate time series.
Extensive studies demonstrate the potential for GAN-based approaches to improve transient prioritization for follow-up observations.
Limitations & open questions
The paper does not discuss the scalability of TransientGAN to larger datasets.
Further research is needed to refine the model's predictive accuracy for rare and novel astronomical events.