Accelerating Posterior Inference from Pulsar Light Curves via Learned Latent Representations and Local Simulator-Guided Optimization

Posterior inference from pulsar light curves is commonly performed using Markov chain Monte Carlo (MCMC), which is accurate but computationally expensive. We introduce a framework that accelerates inference while maintaining accuracy by combining learned latent representations with simulator-guided optimization. A masked U-Net is pretrained to reconstruct complete light curves from partial observations and to produce informative latent embeddings. For a query light curve, we retrieve similar simulations from a precomputed bank using distances in this embedding space, yielding an initial empirical approximation to the posterior over parameters. This estimate is then refined via local hill-climbing updates guided by a forward simulator, progressively concentrating on higher-likelihood regions. Experiments on the observed light curve of PSR J0030+0451, captured by NASA’s Neutron Star Interior Composition Explorer (NICER), show that our method recovers key posterior support and multimodal structure while remaining broadly consistent with MCMC, achieving a 120x speedup and reducing inference time from 24 hours to 12 minutes.