Geometry optimization of a muon-electron scattering experiment (MUonE)
The past year started with the termination of my activities as coordinator of accelerator-based Physics research at my home institution, INFN-Padova. This freed up some commitments in my agenda, most notably a few board meetings in Rome and the organization of seminars at home. I spent the first few months finalizing a study which become a bulky single-author paper, where I demonstrated that by modifiying some design choices for the proposed MUonE experiment would result in the doubling of the sensitivity of the apparatus.
The collaboration ended up only using in part my suggestions, but my article is significant because it shows how it is possible to explore a vast space of possible design choices with fast simulations, in search of an optimized objective that summarizes the experimental goals, and to obtain very large gains in performance. Such observation laid the basis of a long-term research plan which focuses on the full end-to-end optimization of the design of experiments and other devices that employ the interaction of radiation with matter to operate.
MODE: Machine-learning Optimization of the Design of Experiments
The resulting collaboration I coalesced around this plan, called MODE, started its activities later last year, and we already submitted for publication an initial article where we describe our plan. It will be published in "Nuclear Physics News International" next March, if all goes well. I was also elected scientific coordinator of the collaboration and member of its Steering Board (with P. Vischia, J. Donini, A. Giammanco, and F. Ratnikov).
MODE includes 17 physicists and computer scientists from eight institutions, and plans to use differentiable programming and deep learning tools for the creation of a flexible, modular, scalable software architecture integrating event simulation, reconstruction, and inference extraction as a function of all interesting parameters describing in a continuous fashion the design choices and constraints for the construction of the detectors. In our mind such a tool should be applicable to the optimization of devices of different complexity, from detectors for muon tomography all the way to calorimeters for future colliders.
Such a plan is obviously exceedingly ambitious, but the potential gains of a full exploration of the space of design choices for any instrument (which is hundreds-dimensional even in simple cases) is so huge that it is a goal to pursue with the maximum energy.
Measuring TeV-Energy Muons
In June I started collaborating with a CERN physicist who is an expert in machine learning, Jan Kieseler, and with him and two INFN-Padova collaborators we studied a problem nobody had laid sight on. The issue is the precise measurement of very energetic muons. Muons and electrons are precious gemstones in particle collisions originated by protons, because their presence in the final states betray the intervention of electroweak interactions, which are rare and important to study; muons are also an excellent probe of possible new physics beyond the standard model, because new particles may decay by producing muon pairs.
However, our typical collider detectors are not suitable to measuring the energy of very energetic muons. The reason is that unlike electrons, that deposit all their energy inside the calorimeters by electromagnetic cascades (an electron emits a photon, the photon materializes an electron-positron pair, the latter emit more photons, and so on until the whole thing dies out, all the while the detector records the released energy), muons only leave a very small fraction of their energy in a dense medium. They can be measured by observing the curvature of their track in a magnetic field, but this method works well only up to several hundred GeV; for much higher muon momenta their track does not bend even in the strong magnetic fields of CMS and ATLAS; proof be that the resolution on the energy of a 1 TeV muon is in the 8-20% range.
What's worse, if you further double the muon energy, the relative resolution also worsens by a factor of two. Multi-TeV muons may not be an issue at the LHC, but they may be important in a future collider. Following an observation of Jan Kieseler, our group studied how it is indeed possible to infer the energy of muons beyond a TeV with multivariate algorithms focusing on the pattern of energy release along the muon path - which usually is neglected as next-to-useless information in present-day experiments.
After three intense months of work, we were able to show that resolutions of 30% can be achieved for 2 TeV muons in a preliminary article based on simple regressions based on the nearest-neighbor algorithm, but we are now working to improve on that by employing convolutional neural networks. Since this kind of thing has never been done before, and in view of its potential significance for future colliders, the work also well integrates with the activities of MODE.
Handling systematic uncertainties in physics measurements
Starting in March, last year I collaborated with Pablo de Castro Manzano in the preparation of the chapter of a book on Artificial Intelligence for Physics Research. The chapter deals with how to use machine-learning methods to reduce the impact of systematic uncertainties in physics measurement. The topic is one dear to me, and with Pablo we have published in 2019 a very innovative idea on how to obtain the full optimization of a classification task in presence of nuisance parameters (that's how Statisticians call systematic uncertainties). We produced a first draft and a preprint last July, and later we finalized the text for the book, which should be published early next year.
Talks at conferences and other dissemination and outreach activities
I was able to exploit my background studies for the above text in a few lectures, which I gave to a PhD school (the Yandex Machine Learning for HEP school), at a seminar in Lisbon, and at the International Data Analysis Olympiad in Moscow. The topic is also pretty much the subject of a talk I gave (the only one in person!) at the IX ICNFP conference in Crete. As far as presentations are concerned, I also gave a keynote lecture on artificial intelligence for physics analysis at the USERN congress in Tehran last November, and another seminar to the Statistics department of Carnegie-Mellon University, this time dealing with statistical problems in HEP.
Other talks of more outreach flavour were a video interview I gave online with Andrea Idini (MEETSCIENCE) and one with David Orban (Searching for The Question Live), and a talk in person on Art and Science here in Padova, for the Galileo Science Festival, with Prof. Giovanni Bianchi (an art historian).
Concerning ancillary activities, I have been elected as co-chair (with Marta Felcini) of the CMS Thesis Award Committee last September. I will thus organize the grading and prizing of the best Ph.D. theses produced by our young researchers. I was also recently accepted to be a member of the rising Ellis society, which deals with artificial intelligence research in Europe.
I was also appointed as member of the editorial board of a couple of journals (MDPI particles and MDPI symmetry). Ah, and I should not forget that I have continued to be part of the CMS Statistics Committee - really an entertaining way to spend a couple of hours every other Monday evening!
During 2020 I also refereed a PhD defense and half a dozen papers from IJMPA, PPNP, MDPI, SciPost, NeurIPS, and Nature Communications, plus a couple of grant applications, also from Horizon2020. Plus I helped publish a similar number of articles by being an editor of two Elsevier journals.
Finally, I published a proceedings paper for the talk I had given in 2019 at ICHEP VIII, "From SU(3) to Three Quark Families: Hard-Learned Lessons", an invited lecture at the Murray Gell-Mann memorial session of the conference. That article may be of interest to readers here, as it is mostly a historical discussion.
Last but not least, didactical activities. I gave a course in Statistics for Data Analysis for the PhD course in Physics last March, and a course on Particle Physics for the Master in Statistics in the fall. I also followed four students for their thesis; two of them graduated (Martina Fumanelli, Chiara Maccani), and two more are working at their research as we speak. Plus, the two super-graduate students I am supervising for the INSIGHTS ITN network, Lukas Layer and Hevjin Yarar, are doing cutting-edge research in applications of machine learning to HEP, and I expect nice results next year!
In Summary: 2020 was not that bad! And 2021?
So, if I look back at all what happened while I was (mostly) sitting in my home office, unable to go anywhere, I must say I cannot be dissatisfied.
And 2021? Well, first of all I am looking forward to the publication of the first article of MODE, as well as the book on AI for HEP. In the meantime, we will produce a journal publication on the muon regression study. I also am writing (with my friend Bruno Scarpa, a professor of Statistics in Padova, and a few students) an article on a new method for anomaly detection, which is surprisingly powerful. More on the longer time scale, I hope we will start real research activities within MODE, probably starting with the use case of muon tomography detector optimization. A MODE workshop on differentiable programming will be organized probably in September in Louvain-la-Neuve. I am also invited to lecture at a PhD school sometime in the spring, and at the Accademia Nazionale dei Lincei in Rome in March (on Artificial Intelligence for research).
With four students to supervise, two journals (Physics Open and Reviews in Physics) to be an editor of, two more to follow as member of the editorial board, two CMS committees to handle or follow (the thesis award committee and the statistics committee), a blog to attend to, my regular PhD and Master courses to give, and the scientific coordination of MODE just starting to become a thing, 2021 promises to be at least as productive as 2020 was. I only hope we can get rid of this virus as soon as possible, so that regular in-person interaction with my colleagues will restart!
And what about other things?
I promised I would also touch on non-work activities. First of all, the piano. I am taking weekly lessons of piano from Maestro Matteo Galzigna, and have been making steady progress. Last year I brought to a decent level the execution of two of Chopin's studies (Op. 25 n.1, "the eolic harp", and Op. 25 n.2) and started to work on Op. 12 n.10. I also worked on one piece from Schumann's Kreisleriana (n.5) and most of his Kinderszenen. Lately, I have been studying also Chopin's Ballade 1 and Fantasia-Improptu 4. All the above are continuing studies for 2021, of course (who can ever say they are done with Chopin's studies?). But I will try to add to the mix the famous Polacca in Ab Major by Chopin, maybe the piano piece I love the most.
As far as chess is concerned, I unfortunately did not attend any tournament. I only played a few blitz games on Chess24 (I should rather say a few thousand). My Elo rating there has been hovering between 2300 and 2500, but of course it is inflated by at least 300 points with respect to the real international Elo. I do not plan to improve my chess playing skills, though - I long realized that is really too time-consuming for my budget.
Finally, blogging. I did not blog nearly as much as I would have wanted, and I really have no excuse for that - with the lockdowns and all I had more time in my hands than in ordinary times. But perhaps this was also a less inspiring year than usual, given the lack of travel (except two conferences in New York and in Crete, I stayed at home). I will try to do better this year!