I have been preparing a poster for the MODE Workshop, which will take place in the splendid location of Kolymbari, in the Greek island of Crete, from September 1 to 7. The poster is not about one specific research idea or output, as commonly happens. Rather, I conceived it as a summary of the 10 past years of my research, which centered on how to further fundamental science with the new technology that AI has been providing us.
I decided to put ChatGPT to work on a layout that could summarize the connections, the flow, and the main ideas and results. The outcome is the graph shown above, which is unreadable inside this column. I therefore broke it in pieces for this post, to focus on the main elements below.
First, on the left of the graph we find a few motivational inputs: observations I made. Anybody realized in 2012 that a paradigm shift was in place: by boosting our sensitivity to the Higgs boson in LHC searches, Machine Learning had entered the fray of data analysis in particle physics, and was there to stay.
In the face of this, a further observation I made whas that the detector builders community - experimental physicists who design the instruments that collect the data for our physics analyses and searches - appeared to not show the same level of enthusiasm to that new technology: the attitude of expert detector designers was that their job could not be helped much by a machine learning algorithm, let alone taken on entirely. This in my mind was shortsighted, especially since those ML tools were already demonstrating how the information extraction from those detectors outputs was being revolutioned: it was impossible to imagine that this had no impact on the optimal design of the instruments that produced those data, selecting in so doing the kind of information to retrive from a breadth of possible information-generation mechanisms.
And then there was this other small revolution in the making: hadron calorimeters, for decades thought to be the least important piece of hardware of a collider detector, only providing a rough estimate of the collective energy of a stream of particles, were now recognized as crucial enablers of new physics searches at high energy, provided that they returned an image of the energy deposition fine enough to allow the retrieval of information of the pattern of the energy flow. This is because heavy particles like top quarks, W and Z bosons, and Higgs bosons often produce several distinct streams of hadrons when they decay, and the chance of image these streams as distinct objects, in a kinematic regime where they are not widely separated in angle, enables a powerful disentanglement of the signal from otherwise impossible to tame QCD backgrounds.
The listed observations fueled ideas for research work that are shown in the second column: in general, push for a more intensive and extensive use of AI in fundamental science, going beyond the improvement of data analyses and inference extraction. And then, of course these new technologies could improve our capability to design effective detection instruments. Finally, a rethink of the purpose of detection elements was called for: if in the past we had trackers and then calorimeters, assigning a separate task to each, in the future we were going to have to look into hybridizing the two systems, to jointly optimize them for best performance overall.
In 2015, a proposal for exploiting machine learning in advanced data analysis was funded by the European Community. Two years later, another similar project was also funded. These two projects produced a wealth of interesting results, through the hard work of the hired PhD students. One enabling step was the effort of my PhD student Pablo de Castro, who designed the algorithm called INFERNO. INFERNO leveraged differentiable programming to produce an end-to-end optimization of physics analyses. It was a truly novel concept, which focused on how to account for the effect of systematic uncertainties in a physics measurement, and it was an enabling step for the work we would produce later on in the MODE Collaboration.
In 2019 I founded the MODE Collaboration to exploit the idea of INFERNO and the availability of tools for differentiable programming, to put to work the ideas I mentioned above and work toward the end-to-end optimization of scientific experiments - this time looking not only at the information extraction, but also at the design of the instruments as an optimizable procedure. MODE is now a mature collaboration with its own yearly workshop, collaborators from four continents and over 40 institutions, and a wealth of produced demonstrations of end-to-end optimization.
In parallel to MODE, a few years later EUCAIF was founded, and I took charge of leading (with Pietro Vischia) the working group 2 called "Codesign". The idea of studying codesign is in demonstrating that one cannot optimize a piece of hardware in isolation, without accounting for the information extraction procedures that the hardware enables. We produced an important publication (now submitted to Report of Progress in Physics), and are now exploring many demonstrations of co-design of scientific experiments.
One important area of activity that has been fueled by the MODE Collaboration concerns hadron calorimeters, because the high granularity that has been identified as a crucial ingredient to enable heavy objects tagging a decade ago can now be studied as an enabler of an additional channel of information out of hadron showers: the identification of the particles that originate the showers. This is a very hard task, but CNNs and graph neural networks may allow it. We proved it in a publication and in another we examined whether the very high granularity that is needed to harvest the necessary information to do particle identification could be by-passed by novel methods to read out the energy deposition patterns. This led to investigate neuromorphic computing as a solution, and to the writing the winning proposal of a EIC-Pathfinder project, PHINDER, which has just started last April.
PHINDER exploits nanophotonics - in particular, indiun arseniate nanowires - to perform two important tasks: act as light-sensitive elements in a particle detector where the active material is a scintillating plastic, and act as neurons in a neuromorphic sensing and computing network. There are several distinct elements of novelty in its concept: photons are used for computation, and their time series becomes the central vector of information extraction. This allows us to extract topological information from particle showers without the need to chop our calorimeter into millions of independently read-out electronic channels. I realize it sounds rather nebulous stated as above, but I think this post is too long to get into more detail on that here. Perhaps I will write more about PHINDER research in the near future.
So, going back to the big picture - it makes me feel proud to look back at these 10 years and to realize how much we were able to accomplish. Besides three funded projects, with over 9 million euros granted by the European Community, we published two dozen articles that leverage new ideas, often groundbreaking ones. I am really grateful to all the undergraduate and graduate students who made this possible!
