The European Commission pays close attention to document the work of the projects that benefited of its funding. With that intent, the AMVA4NewPhysics network has been described, along with its goals, in a 2016 article on the Horizon magazine.

And what is this AMwhatchamacallit network anyway? It is a "European Training Network", a joint venture of a dozen research institutes and universities in Europe who, together with a few industrial partners, provide advanced training to Ph.D. students in Physics and Statistics, and does research at the Large Hadron Collider by participating in the CMS and ATLAS experiments. The network has a blog, where the students and the other members write about machine learning and particle physics, among other things.

Since the network is nearing the end of its lifetime, I was asked by the EU to provide some information for an update of the above article. I think it is useful to share the questions and my own answers below. Of course, the mentions I made below of some of the network output are only a partial representation, and are not meant to pay justice to the many more product, algorithms, articles, and scientific results that have been produced by the 11 early-stage researchers and their research groups in the past 3.5 years - so I apologize in advance to my colleagues if they would have liked to see more emphasis on the results of their liking. This should therefore become a motivation to write dedicated posts on the matter in the near future!

  1. By the time the project comes to an end this August, how much progress will have been made toward deploying advanced statistical learning techniques for data analyses at the LHC? What are the advantages and potential benefits of this approach?

The AMVA4NewPhysics network started its operations in September 2015. In these four years there has been an enormous effort by the LHC collaborations, especially the ATLAS and CMS experiments of which network members are collaborators, to improve their measurement and search capabilities using new advanced methods. This has been spurred by two separate facts: first, the booming of new machine learning and deep learning technologies in the society at large; and second, the increased confidence of the particle physics community toward using those new tools in fundamental physics research. AMVA4NewPhysics has provided a visible and important contribution to this effort, with the development and application of several entirely new algorithms to the physics cases of interest of the ATLAS and CMS collaborations. One example taken from CMS data analyses will suffice. One of the most important analysis tasks of relevance for both searches of new physics and detailed measurements of the Higgs boson properties is the identification of b-quark-originated jets, the so-called “b-tagging”. Members of the network have developed a neural-network-based b-tagging algorithm, DeepFlavour, which significantly outperforms previous methods. Its application to physics analyses and searches has considerably improved the measurement capabilities and discovery reach of the experiment.

  1. What steps have been taken to date? And what do you foresee being the key achievements by the end of the project?

While the network will continue to operate for five more months, and its research activities will be continued by its members after that, all but three of the early-stage researchers (ESRs) enrolled in PhD programs by the participating beneficiary nodes have already concluded their contract. Most of the scientific objectives of the network have thus already been achieved, with only one scientific deliverable still to be produced (it is due April 30th). Among the key achievements, besides the already mentioned DeepFlavour algorithm, three in particular should be cited. One is the development of a new software package for matrix element calculations, MoMEMta (, which significantly improves our capability to produce theory predictions for our physics searches. A second is a data-driven density estimation technique for proton-proton collisions producing multiple hadronic jets, which has allowed an accurate modelling of the complex background from quantum chromodynamics processes in the search for the elusive process of pair-production of Higgs bosons, as well as the production of a CMS scientific article that has just been accepted for publication ( A third is an innovative neural-network-based method that allows optimal inference in situations where systematic uncertainties affect significantly the measurement process; the algorithm, called “INFERNO” (from “inference-aware neural optimization”) produces a summary statistic which is robust to the degrading effect of unknown values for nuisance parameters, strongly boosting the precision of the inference. A publication has been submitted to Computer Physics Communications (; we are in the process of testing this algorithm on several use cases with CMS and ATLAS data, with excellent prospects of significantly improving the resulting scientific output.

  1. Can you briefly outline the highlights and practical examples of your work so far? Have any insights or discoveries of note emerged from the data analyses?

I would refer to the previous answers for the first question above. As for the second, the focus of the AMVA4NewPhysics network has been on the development of new techniques. Particle physics data are complex and a cycle from data collection to publication of results typically takes two to three years; hence it is too early to say if those new techniques, which have been made available only during the past year, will produce new discoveries. However, I can say I have no doubt that they will significantly increase our insight in many of the studied physics processes.

  1. What would you describe as the most significant social or economic benefits arising from your work (including related to the training aspects of the network)?

The network is indeed based on the pillar of innovative training, providing a generation of young researchers with better tools to become leaders in research or industry in the future. It is here that I think the AMVA4NewPhysics network has had its largest socio-economic impact, as not only we could train 11 early-stage researchers in fundamental physics and machine learning in a much more thorough and tailored way than conventional PhD programs can do, but we also created and established important and lasting ties with four non-academic industries: B12 consulting (, SDG group (, MathWorks (, and YANDEX (

[In passing, I cannot avoid thanking here the PIs of these partners of our network, respectively Michel Herquet, Maurizio Sanarico, Ilya Narsky, and Andrey Ustyuzhanin – they have offered invaluable training and supervision to our ESRs.]

By providing a continued sponsoring, e.g., of the YANDEX summer school on Machine Learning for High-Energy Physics (see e.g. we have contributed to making that event a truly unique provider of knowledge at the cutting edge of research in machine learning. It is not by chance that the AMVA4NewPhysics network will hold its final workshop and offer a public lecture on artificial intelligence in cooperation and in coincidence with this year’s edition of that school, which will take place at the DESY laboratories of Hamburg, Germany at the beginning of July.

  1. How can the project results be used (e.g. will they be implemented for future work at the LHC or for other particle physics applications, taken up by partners, or lead to further research in the field)?

Most of the products of the AMVA4NewPhysics network are publically available on dedicated web pages and software repositories and are well documented, so they can be used by the community without limitations. As mentioned above, some of them have already been employed in physics analyses, and they will continue to be; others are in the process of being adapted to future analyses not yet carried out in full. Some of the algorithms may also be of interest outside high-energy physics, and in all cases they contribute to the general progress in computer science, such as e.g. is the case of the INFERNO algorithm, which addresses the general issue of likelihood-free inference that has seen a considerable interest in the recent past.

6. Any other highlights worth mentioning in the article?

One of the goals of the AMVA4NewPhysics network was to increase the outreach action and the culture of outreach within the particle physics community. We designed a very comprehensive plan which included social media action (through a blog, , and twitter account), public lectures and engagement events, interaction with high-school students, and crucially, a workshop to perform a self-examination of the impact of those outreach practices, with a view to improving their effectiveness. All of the above activities have been carried out successfully, although the blog – which was supposed to receive at least two monthly contributions by each one of the ESRs – suffered from the very busy agenda of our fellows, torn between research, training, conferences, secondments, and the goal of achieving their PhD. The blog ended up being a lower throughput web site than originally foreseen (still, the 295 articles published there have so far received overall over 110k visits). I would like to mention that our self-assessment of the outreach action, which was performed as planned during a special workshop held last September at CERN, was deemed by the participants an extremely fruitful process, and all the invited outreach experts who participated to the workshop agreed that this initiative must and will continue in the future.


Tommaso Dorigo is an experimental particle physicist who works for the INFN at the University of Padova, and collaborates with the CMS experiment at the CERN LHC. He coordinates the European network AMVA4NewPhysics as well as research in accelerator-based physics for INFN-Padova, and is an editor of the journal Reviews in Physics. In 2016 Dorigo published the book “Anomaly! Collider physics and the quest for new phenomena at Fermilab”. You can get a copy of the book on Amazon.