Probabilistic Data-Driven Modelling of Astrophysical Transients: The Neural Process Family for Ultrafast and Class-Agnostic Light Curve Reconstruction

Astronomical light curves --- the sparse, irregular, multi-wavelength time series produced by photometric surveys --- are notoriously hard to model, with Gaussian Processes (GPs) the current standard despite their poor scaling, *a priori* kernel specification, and inability to share information across objects. We introduce the use of Attentive Neural Processes (ANPs) as a fast, data-driven, morphologically agnostic interpolator: a single meta-learned model handles 14 explosive transient classes without specifying the class at inference. Evaluated on realistic LSST cadences, the ANP outperforms seven GP- and neural-network-based benchmarks on all regression, feature-recovery, and probabilistic-calibration metrics, while running in $\sim 10^{-6}$ s per light curve --- four to five orders of magnitude faster than benchmark models --- fast enough for on-line use on the LSST alert stream of $\sim 10^{7}$ events/night. We demonstrate how the ANP avoids neural networks' overconfidence and GPs' underconfidence, delivering sharp, calibrated uncertainties suitable for downstream science. This work establishes the neural process family as a scalable, probabilistic foundation for real-time transient science in the LSST era.