Parallelized Inference for Gravitational-Wave Astronomy

The likelihood functions we typically have to evaluate in gravitational-wave astronomy typically involve combining many independent measurements, whether that’s combining information from many frequency bins and detectors when analysing single transient events, or many different events when performing a population analysis. This means that our analyses can quite easily become limited by our ability to add or multiply large arrays of numbers.

In this work we showed how using GPUs to parallelise the computation of models for either the gravitational waveforms or population models likelihood evaluation times can be dramatically reduced and our astrophysical inference is dramatically accelerated.

With GPU-accelerated code we will be able to easily analyse much longer signals allowing us to look at more low frequency content. Using our GPU accelerated code we perform the first inference on a simulated binary neutron star inspiral lasting longer than 500s. The angular momenta “spin” of the neutron stars are constrained approximately twice as well using this longer signal compared to an analysis of the last two minutes.

Additionally, using the GPU-accelerated population inference code we will be able to analyse hundreds of events together an order of magnitude faster than the single-threaded CPU code used to analyse the first gravitational-wave transient catalogue.

Code developed for this paper can be found at:

• GPUCBC - GPU compatible code for analysing single events.
• GWPopulation - GPU compatible code for analysing populations.
• Gregory Poole’s CUDA enabled version of IMRPhenomPv2

See interactive examples of using the code for analysing a simulated binary neutron star inspiral (colab github) and GWTC-1 (colab github).