Abstract: This project will develop a GPU implementation of gravitational-wave signal models to alleviate the computational bottleneck in LIGOs parameter estimation codes. Parameter estimation is an optimal tool for gravitational-wave signal detection but its current high cost prohibits current codes to be used as a search pipeline. The dominant cost of the codes is in evaluating signal models which are compared to data to obtain the likelihood of the data containing a gravitational-wave. The signal models are highly parallelizable. Hence there is strong motivation to implement the models on a GPU. The cost of signal modelling would need to be made between one and two orders of magnitude cheaper to develop a cheap gravitational-wave search pipeline.
Searching for gravitational waves using parameter estimation tools would allow LIGO to search a much larger volume of the observable universe. Additionally, we expect to be able to detect the “stochastic background” of gravitational waves using only a day of data, compared to around 40 months of data with current techniques. Detection of this background would be a big astronomical prize, allowing the study of the whole population of binary black holes and neutron stars, as well as having cosmological implications.
The project will produce a GPU parallelized implementation of a commonly used gravitational- wave signal model known as “IMRPhenomP”. This implementation would allow us to benchmark expected performance improvements to current parameter estimation tools. Looking further, the implementation would serve as a corner stone of a fully GPU-implemented Monte Carlo algorithm, the core algorithm used in parameter estimation.