Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

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

NeurIPS 2026 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: PAC-Bayes and decoupled GP learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NightLANP, an Attentive Neural Process meta-trained on 15 simulated transient classes to reconstruct sparse multi-band LSST-like light curves amortized in microseconds, outperforming GP and neural baselines on regression, peak recovery and calibration metrics.

Problem

Light curves from LSST-like surveys are sparse, irregular and multi-band, and standard interpolators (Gaussian Processes) need per-object kernel fitting and scale poorly, while neural interpolators are often miscalibrated.

Approach

Applies Attentive Neural Processes meta-learned across a class-agnostic population of simulated transients, using band-as-coordinate inputs, dynamic slicing and temporal re-zeroing, so one amortized model reconstructs all bands at inference without per-curve fitting.

Key results

• ANP achieves median MSSE 3.01 versus 6.56 for the best GP variant and 7.95 for BNN; NLPD 0.06 versus 1.36 for the next-best GP.
• ANP PICP at 95% is 92.7%, closest to nominal, while GPs reach 86.9-88.3% and BNN only 14.2%.
• ANP inference takes about 4.92e-6 s per light curve on two T4 GPUs, roughly four orders of magnitude faster than MLP and five or more faster than GPs (~1.2 s).

Surprising findings

• Out-of-the-box ANP forecasting of peak time from rise-only contexts failed without a separately trained forecasting variant.
• SN Ia reconstruction errors are higher than for core-collapse and stripped-envelope classes, attributed to long zero-flux padding in simulations.

Limitations

• Trained and tested only on PLAsTiCC simulations, not real survey data.
• No architecture or hyperparameter search; observational flux uncertainties not used as inputs.
• Only explosive transient classes tested; generalization to other variable phenomena untested.

Experimental setup

Datasets: PLAsTiCC simulations (SNANA, RESSPECT high-cadence subset) LSST OpSims v5.1.1 cadence

Benchmarks: PLAsTiCC-derived held-out test set with 3,750 light curves under LSST OpSims cadence

Metrics: MSE MSSE MAPE MASE chi2 NRMSE_p NRMSE_o NLPD CRPS PICP

Baselines: Gaussian Process (GP) GP2 GP3 Bayesian Neural Network Normalizing Flow MLP (pytorch) MLP (sklearn)

Builds on: Attentive Neural Process Gaussian Process Conditional Neural Process Neural Process fulu

Closest papers by meaning

• Deep Random Features for Scalable Interpolation of Spatiotemporal Data (ICLR 2025 · similarity 0.84)
• AstroCompress: A benchmark dataset for multi-purpose compression of astronomical data (ICLR 2025 · similarity 0.83)
• Unified Framework for Diffusion Generative Models in SO(3): Applications in Computer Vision and Astrophysics (AAAI 2024 · similarity 0.82)
• A systematic evaluation of vision-language models for observational astronomical reasoning tasks (NeurIPS 2026 · similarity 0.81)
• Generating Full-field Evolution of Physical Dynamics from Irregular Sparse Observations (NeurIPS 2025 · similarity 0.81)
• Learning What's Real: Disentangling Signals and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics (NeurIPS 2026 · similarity 0.81)