Reference : UMR7164-KEVVEL-004
Workplace : PARIS 13
Date of publication : Tuesday, September 20, 2022
Scientific Responsible name : Edward K. Porter
Type of Contract : PhD Student contract / Thesis offer
Contract Period : 36 months
Start date of the thesis : 1 November 2022
Proportion of work : Full time
Remuneration : 2 135,00 € gross monthly
Description of the thesis topic
To extract the scientific information from binary mergers, the LVK uses Bayesian inference to estimated the physical parameters of the binary. At present, the LVK uses both Markov chain Monte Carlo and nested sampling algorithms. As these algorithms both come from a family of random walk samplers, their convergence is slow, requiring runtimes of weeks to months. The group at APC has been working on the development of a state of the art Bayesian inference algorithm called DeepHMC. This algorithm is a Hamiltonian Markov chain sampler that is a non-random walk sampler, and as it has an exponential convergence, is an order of magnitude faster that the current algorithms in use by the GW community. This acceleration comes in a large part from a machine learning engine at the heart of the algorithm. The goal of this PhD thesis is to further optimise and develop the machine learning basis of the algorithm, and to parallelise the algorithm where possible for use with GPUs/CPUs, further reducing the runtime of the algorithm and potentially allowing it to be used for near real-time, low latency parameter estimation. During the duration of the thesis, it is expected that the student will split their time 50/50 between the LVK and ET collaborations. Within the LVK, the student will be part of the Compact Binary Coalescence group, and in particular will be expected to play a role in both the parameter estimation and nuclear astrophysics groups. Within the ET collaboration, the student will be part of the Observational Science Board, and will work on problems regarding the parameter estimation and science extraction of compact binary coalescences.
The plan for the three years is as follows:
Year 1: investigate and optimise the architectural structure of the deep neural network by experimenting with the number of layers, and the number of neurons in each layer, as well as investigating different activation functions and their combinations. Where possible, parallelising the algorithm for application on GPUs and/or CPU clusters. Increasing the dimensionality of the algorithm to include missing physical effects such as binary precession and eccentricity.
Years 2/3: apply the algorithm to real data from LIGO-Virgo-Kagra's fourth science run. Take part in the ET Science Challenges which will use simulated data and work on the development of a viable inference algorithm for long duration, overlapping GW signals for 3G GW detectors.
Until very recently, gravitational waves (GWs) were the last unverified prediction of Einstein's theory of general relativity (GR). While indirect evidence existed from the orbital decay of the Hulse-Taylor binary pulsar, it was only with the observation of the merger of the binary black hole (BBH) GW150914 that we finally had direct evidence for the existence of GWs. This event opened a new window on the universe and was the birth of multimessenger astronomy. Within two years, the first observation of GWs from the binary neutron star (BNS) merger GW170817 was made. This seminal event coincided with the observation of a gamma ray burst from the Fermi satellite. Within 10 hours, an optical counterpart was identified by a number of telescopes world wide. And in the next few days to weeks, were followed by observations across the electromagnetic spectrum. This one event had huge consequences across a number of different fields. The arrival of the gamma rays within two seconds of the GW signal confirmed the prediction that GWs travel at the speed of light, an observation that severely constrained alternative theories of gravity. Combining the GW and EM signals not only put a tight constraint on the nuclear equation of state for neutron stars, but also provided evidence that the heaviest elements in the universe might actually come from BNS mergers, and not supernovae as was previously thought. Finally, as GWs allow us to measure the luminosity distance to the source, combining this information with the measure redshift from EM observations allowed us to put the first constraint on the value of Hubble's constant that did not require a cosmic distance ladder or observations of the CMB. Since then, a total of 90 GW events have been observed including the first direct evidence of mixed neutron star-black hole binaries and the formation of an intermediate mass black hole.
While originating from some of the most violent events in the universe, GWs are incredibly weak, and as a consequence, very hard to detect. A global network of GW detectors are in the process of being upgraded in time for the 4th LIGO-Virgo-Kagra (LVK) at the end of 2022. This network consists of the two 4km LIGO detectors in the USA, the 3km Virgo detector in Italy and the 3km KAGRA detector in Japan. In the next few years it is also expected that another 4km LIGO detector, planned for construction in India, will also join the network. As we approach the constraints of the current detectors, we are limited as to how low we can push the low frequency performance of the detectors, and will have to construct new detectors for the future. In Europe, the Einstein Telescope (ET) project has recently become part of the ESFRI roadmap for large scale projects. This cryogenic 10km underground detector will allow us to descend to a frequency of around 2 Hz, opening up the possibility of detecting GW sources out to a redshift of z = 100, with an expected rate of ~10^5 sources per year. In the USA, plans are being studied for the construction of a second 3rd generation detector called Cosmic Explorer, planned to work in collaboration with ET.
Constraints and risks
Travel in France or abroad is to be expected.
No particular risk.
We talk about it on Twitter!