FastParam

About the FastParam Project

FastParam is a research project funded by the ICSC – Centro Nazionale di Ricerca in HPC, Big Data e Quantum Computing, as part of the PNRR, and supported by the European Union – NextGenerationEU. It aims at reducing inference times in the context of Pulsar Timing Arrays, galactic-scale observational instruments consisting of sets of pulsars, whose beams' time of arrival are registered over a time window of 15–20 years. Many of such datasets exist ranging from EPTA to PPTA, accounting for a growing community of researchers. The state-of-the-art software enables complex modelling and inference of astrophysical and cosmological sources of delay in the time of arrival of the pulsar signals, but faces enormous challenges in terms of computational time.

We have explored and implemented dimensionality reduction and machine learning-driven sampling solutions to reduce such time, allowing rapid testing of physical hypotheses pushing innovation and theoretical development in the field. We used Hierarchical Bayesian Models with normalizing flow boosted sampling for fast execution and hyperparameter reparametrisation for better accuracy. We also developed a novel Empirical Bayes approach for efficient application of nested sampling to the Pulsar Timing Array problem. The method has shown promising results in terms of both efficiency and speed as well as accuracy with respect to the injected target values from synthetic data. The project delivers openly accessible tools: frameworks, PTA analysis algorithms, and source code repositories, all provided under transparent licensing terms. In this way, FastParam contributes to reproducible science and accelerates discovery for both the astrophysics and AI communities.

Project outputs

• Project outputs: A software package for rapid inference on Pulsar Timing Arrays. On input an enterprise PTA object, our pipeline can follow two possible roads stemming from the hierarchical Bayesian modeling of the problem. An accurate inference reparametrisation of prior hyperparameter space is available for posterior independence from possibily misspecified priors. An empirical Bayes approach has been developed for non-hierarchical modeling with minimal loss of accuracy but enormous gain in speed.
• Synthetic datasets: Have been produced for software testing, which are modeled over real pulsars according to their physical description. Synthetic data is obtained using widely acknowledge instruments like the tempo2 package. Injecting noise and Gravitational Wave parameters, we have been able to recover such values in record time.
• Inference: A wide demo set of posterior and plots which showcases the output of our new inference pipeline. One can browse by prior choice, number of pulsars and physical model assumptions: we have pushed the boundaries of our research by stress testing on the widest possible scenarios.

Contact & Acknowledgements