Title: Leveraging the Neural Process Family for Ultrafast, Class-Agnostic Light Curve Reconstruction in Astrophysical Transients: The NightLANP Approach to Probabilistic Data-Driven Modeling

Abstract

Ground-based astrophysical observations frequently yield sparse and irregular light curves due to limitations imposed by weather, environmental factors, and scientific protocols. As the Vera C. Rubin Observatory prepares for its Legacy Survey of Space and Time, the resulting data volume presents unprecedented opportunities for transient research. However, the survey’s cadence—characterized by sparsity and irregularity across six photometric bands—remains a significant hurdle for accurate inference. While interpolation techniques are commonly employed to address this gap, Gaussian Processes (GPs), the traditional standard, face notable limitations. They struggle with cross-band correlations, demand prior kernel specification, and require individual fitting for each light curve, resulting in poor scalability.

To overcome these challenges, we introduce the application of the neural process family for light curve reconstruction. This approach merges the probabilistic rigor of Gaussian Processes with the computational efficiency of deep learning. By utilizing meta-learning on a diverse array of simulated transients, our method, Attentive Neural Processes, moves the heavy computational burden to the training phase. This enables rapid, amortized inference using a class-agnostic model.

We evaluated our approach against realistic Rubin cadences covering 15 distinct transient classes. The results demonstrate that an unoptimized, off-the-shelf Attentive Neural Process consistently surpasses all benchmarks—including various Gaussian Processes and neural networks—ac every tested metric. These metrics encompass regression quality, the recovery of astrophysical features, and probabilistic calibration. Furthermore, our model performs simultaneous interpolation across all bands in microseconds. This speed is over four orders of magnitude faster than the next-best neural benchmark and five times faster than Gaussian Processes, highlighting the potential of neural processes for handling the nightly Rubin alert stream. Additionally, Attentive Neural Processes mitigate the overconfidence typical of standard neural networks and the underconfidence associated with Gaussian Processes, providing sharp and well-calibrated uncertainty estimates. This study establishes the neural process family as a scalable, probabilistic foundation for real-time transient science in the Rubin era.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC