Support for young scientists from all over the world
Doctoral projects
© Stephen R. Green

Machine learn­ing methods for gravi­ta­tional-wave data analysis

Maxim­il­ian Dax – Hector Fellow Bernhard Schölkopf

The detec­tion of gravi­ta­tional waves (GWs) has opened a new window on the universe, through which we can study the physics of black-hole and neutron-star mergers. By analyz­ing GWs we can infer proper­ties of the corre­spond­ing astro­phys­i­cal systems. Current analy­sis methods are however too compu­ta­tion­ally expen­sive to deal with the growing amount of data. My research is thus concerned with the devel­op­ment of more efficient methods for the GW analy­sis using power­ful machine learn­ing methods.

The coales­cences of binary systems of black holes or neutron stars emit gravi­ta­tional waves (GWs) that encode infor­ma­tion about the dynam­ics of the system. The detec­tion of GWs there­fore offers the excit­ing and unique oppor­tu­nity to gain insight into these events. Current analy­sis methods are however slow and compu­ta­tion­ally expen­sive. This becomes increas­ingly problem­atic with the growing density of detected signals due to the contin­u­ous improve­ment of the detec­tor sensi­tiv­ity. More efficient analy­sis methods are there­fore essen­tial for the progress in gravi­ta­tional physics. My research is concerned with the devel­op­ment of machine learn­ing methods to accel­er­ate the analy­sis of GWs.

My current focus is on infer­ring the parame­ters of astro­phys­i­cal systems (e.g., the masses) from the GW data observed in detec­tors (‘strain’). This task can be inter­preted as an inverse problem. In the forward direc­tion the waveform for a given set of parame­ters can be simulated within the theory of general relativ­ity. We are inter­ested in the backward direc­tion; given the measured strain data that contains the GW waveform and detec­tor noise we estimate the parame­ters using power­ful methods for simula­tion-based infer­ence. Specif­i­cally, we train a deep neural network to model the so-called poste­rior distri­b­u­tion. Due to the amorti­za­tion of compu­ta­tion for differ­ent events and the capabil­ity to gener­ate poste­rior densi­ties using only a forward pass through a neural network at infer­ence time, our approach allows for rapid low-latency analy­sis of GWs.

Machine Learning Methoden für die Analyse von Gravitationswellen

Figure 1: Poste­rior densi­ties of the masses of the binary-black hole system associ­ated with the first ever detected gravi­ta­tional wave GW150914. The result of the expen­sive standard method is shown in blue, the output of our neural density estima­tor in orange. The contours repre­sent the 50% and 90% credi­ble regions. Reprinted by permis­sion of Stephen R. Green.

Maxim­il­ian Dax

Max Planck In­sti­tute for Intel­li­gent Systems

Super­vised by

Prof. Dr.

Bernhard Schölkopf

Infor­mat­ics, Physics & Mathematics

Hector Fellow since 2018Disziplinen Bernhard Schölkopf