Tommaso Dorigo – a Visiting Researcher

Luleå University of Technology is conducting a large amount societal beneficial, cutting-edge and interesting research. To accomplish this, we collaborate with other incredible people who are at the forefront in their field – one of these people is Tommaso Dorigo, particle physicist and computer scientist.

Tommaso is currently collaborating with the Machine Learning group at Luleå University of Technology due to our expertise in neuromorphic computing.

"I was invited to a key lecture here in October 2022. I was talking to Fredrik Sandin, Professor in Machine Learning at Luleå University of Technology, and we found out that our research was closely related. So, we hired a joint PhD student and are currently working on a design for a calorimeter as well as detector optimisation", Tommaso Dorigo says.

In particle physics a calorimeter is a detector which is used to measure the kinetic energy of particles. Particles enter the calorimeter and initiate a particle shower in which their energy is deposited in the calorimeter and are then collected and measured.

"What we want to do is essentially to exploit Einstein's famous formula E=mc^2 by allowing the energy to convert into mass of other new particles that are created in collisions with the nuclei of atoms in the calorimeter. This produces a multiplication of particles – a shower. We can measure E if we collect the signal of all produced particles, so we shoot a particle into a dense enough block so that the entire shower stays within the block without losing secondary particles, which would degrade the energy measurement. This way we can measure the kinetic energy more precisely", Tommaso explains.

Building more performant calorimeters is very important, because getting more precise measurement of particle energies enables us to take clearer pictures of what happens in the high-energy collisions, we produce in particle colliders and therefore to understand more of the physics that governs the subnuclear realm. But calorimeters are also used in hadron therapy, when we want to precisely image the interior of the human body before we irradiate a tumour, for example. So, the research is interdisciplinary and has great potential for societal benefits.

For further optimisation of the detector the collaboration also includes NanoLund, a research organisation where material scientists study nano-photonics which can be used in the optimisation of the calorimeter. Specifically, it is the nanowires which detect photons released by the energy release processes in the calorimeter, the utilisation of the nanowires in a calorimeter would thus increase the overall quality of the analysis.

Challenges in designing particle detectors

Tommaso Dorigo has an extensive and impressive academic record. During his PhD he worked on the CDF-experiment (Collider Detector at Fermilab) and was part of the group which found the top quark, which was missing at the time. In 2012 he was the Chair of the statistics committee in the research project which discovered the Higgs boson, one of the most important discoveries in particle physics.

"It was huge discovery for modern particle physic science, not only for the discovery itself but also because it started changing the culture of how particle research was conducted. Prior to the discovery the use of machine learning was frowned upon by particle physicists. However, as the Higgs boson was discovered by the help of machine learning the opposition was overcome", Tommaso says.

Machine learning has been embraced by the researchers as it gives more precise results. Tommaso Dorigo believes that the next paradigm shift might be the use of artificial intelligence, AI, in particle physics. Shortly, it is because AI can assist in designing experiments to enable these devices to produce better data from particle collisions, which ultimately enables us to do better science with them.

Designing and building a full-scale detector for a large particle physics facility, such as the Large Hadron Collider at CERN, takes approximately 20 years. A detector builder may use a high-fidelity simulation of the physical processes that take place when particles created in the collisions interact with the detector, to assess the merits of a specific design. This is CPU-costly, so one may only afford to test a few different configurations - say 10. Of those ten simulations the detector builder will conclude that one of those ways is more effective than the other nine.

"The problem with this process is that we probe, let us say, ten points in a thousand-dimensional space. To solve this, I want to construct an algorithm that works like a neural network, which is powered by differentiable programming and exploits the chain rule of differential calculus to propagate derivatives of the loss function back to the parameters to be optimized by gradient descent. Essentially, I want to create the detector in this machinery and optimise it", Tommaso Dorigo says and continues:

"It is rather simple when you think about it, if you are looking for a bakery you would not just walk in different directions and hope you end up at a bakery. This is what is currently happening with the probing of ten random points in this dimensional space. I want to find the bakery by sniffing for the bread. In short, this is what the AI will do in the thousand-dimensional space. It is progressively walking in the direction where the smell of bread gets more intense, that is what gradient descent does in a nutshell."