I Using AI for Gravitational Wave Generation

[kelly0303]

Hello! Can someone point me towards some papers that use AI for gravitational wave generation? I found many papers using AI to analyze data, but not really something where AI is used to actually generate waveforms. For example this was done for particle physics simulations to increase the speed by a few orders of magnitude (at some loss in accuracy), so given that a GW simulation takes a few weeks (based on SXS data), I assume this would be a good place to try AI simulations (I know there are some non-AI based codes out there providing simulations, but I am interested in codes using AI). By AI I mean deep neural networks.

 

[romsofia]

I don't see how AI would be helpful in creating NEW waveforms as I believe it would most likely overfit to whatever data you train it on? So, maybe you could use AI to help decide which waveform would be best given a bunch of them (outside of what is being used right now, since ML is being used in GW stuff).

The closest paper that I know of that would do this would be: https://arxiv.org/pdf/1909.10986.pdf
It might be a good place to start, but I don't think AI/Machine Learning can really do what you ask, but I've only dabbled in it.

 

[kelly0303]

I don't see how AI would be helpful in creating NEW waveforms as I believe it would most likely overfit to whatever data you train it on? So, maybe you could use AI to help decide which waveform would be best given a bunch of them (outside of what is being used right now, since ML is being used in GW stuff).

The closest paper that I know of that would do this would be: https://arxiv.org/pdf/1909.10986.pdf
It might be a good place to start, but I don't think AI/Machine Learning can really do what you ask, but I've only dabbled in it.

I will look over that paper, but AI can certainly do that. As I said there are lots of papers used for particle physics simulations (jets for example at LHC using GANs) and obviously they are not just reproducing already existing simulations, but interpolating among them and hence creating new event topologies much faster. So I see no reason why this can't be done for GW.

 

[romsofia]

I will look over that paper, but AI can certainly do that. As I said there are lots of papers used for particle physics simulations (jets for example at LHC using GANs) and obviously they are not just reproducing already existing simulations, but interpolating among them and hence creating new event topologies much faster. So I see no reason why this can't be done for GW.

I'm not sure about particle physics simulations because that isn't my field, so I'd have to see if there are any similarities.

Waveforms are already generated via algorithms for the most (as stated in the paper), and yes, i was being dramatic that ML can't be used, but at the current point, I see no reason why it'd be mainstream. They do talk about using machine learning for it in the paper as well, but I haven't seen any development on the techniques, but hopefully that paper sets you off in the right direction!

Good luck.