Machine and Deep Learning in Science: Particle Physics


 In most of the research I performed as an experimental physicist, a signal was searched buried in the noise. In the case of Belle II searches, such as the search for invisible decays of a Z' boson, not even the mass of such particle is known, making the search even more complex. Over the years, it required smarter and smarter solutions. One such solution was the implementation of the so-called Punzi-net, a neural network architecture that minimises a loss function typical of particle physics experiments and maximises the sensitivity of the search for the Z'. 

When I developed, together with my team, the first version of the Punzi-net, we expected some improvements over standard analysis methodologies, but allowing the network to learn features typical of the search for the invisible Z' and at the same time minimising, first a standard loss-function, and then in sequence a loss-function that we derived from statistics applied to high-energy physics, we discovered that the neural network was capable not only of identifying potential signals more efficiently than standard tools, but it had the capabilities to simultaneously generalise to all mass hypotheses of the search, as shown below.  Not surprisingly, this has attracted the attention of local media in Austria!

Screenshot 2024-10-16 at 13.15.36
Screenshot 2024-10-16 at 13.15.36
Screenshot 2024-10-16 at 13.16.10
Screenshot 2024-10-16 at 13.16.10
Screenshot 2024-10-16 at 13.17.36
Screenshot 2024-10-16 at 13.17.36

 

Machine and Deep Learning in Science: Gravitational Waves

Gravitational wave signals are generally buried in the noise of the detector and are usually extracted from the noise through a method called matched filtering, which cross-correlates data with a bank of waveform templates to see if any of the templates match the observed data. This method is very powerful, but it requires a-priori knowledge of the exact morphology of the gravitational waves. Events whose waveforms deviate significantly from the generated template bank, due to unusual morphology (exotic objects, highly eccentric mergers, etc.) might be overlooked. 


I am involved in developing deep-learning algorithms that use a completely different approach: anomaly detection. Gravitational wave data consists mainly of the detector strain, which is essentially the detector noise, except when a gravitational wave passes through the detector. Algorithms, such as autoencoders, can be used for such a task.

Autoencoders are deep-learning tools that encode the information on which they are trained in a lower dimensional space and then decode it again. The aim of the autoencoder is to reconstruct an output as close as possible to the input data, with a reconstruction error (related to the difference between output and input) as small as possible.



In my approach, I train a deep convolutional autoencoder (CAE) on a large sample of simulated detector noise and test if when a signal is included, it is flagged as an anomaly. The work is ongoing, and the model performs well, suggesting that it might be employed in real searches with further studies.