Periodic Reporting for period 4 - mPP (machine learning for Particle Physics)
Reporting period: 2022-10-01 to 2023-06-30
This ERC project aimed at developing Machine Learning applications for High Energy Physics, and in particular for the experiments operating at the CERN Large Hadron Collider (LHC). The experimental environment at the LHC is peculiar and challenging for many aspects:
- Generated data are sparse sets of electronic signals recorded by irregularly arranged arrays of sensors.
- The LHC delivers an unprecedented amount of data, that has to be filtered in real time and with short latency.
- The data processing is operated on custom computing resources, such as FPGAs mounted on electronic boards,
dedicated computer clusters, etc. This implies specific constraints in terms of latency, memory footprint, etc.
- Typical data analysis workflows also rely on large sets of simulated data, which require sizeable resources to be
produced.
This ERC project demonstrated how Deep Learning can facilitate this data processing, freeing computing resources to enlarge
the scientific program of the experiments and developing new techniques to extend their reach beyond traditional techniques.
Thanks to the work performed in the last five years, these techniques are now accepted by the community as mainstream part
of the scientific exploitation of LHC data
- Generated data are sparse sets of electronic signals recorded by irregularly arranged arrays of sensors.
- The LHC delivers an unprecedented amount of data, that has to be filtered in real time and with short latency.
- The data processing is operated on custom computing resources, such as FPGAs mounted on electronic boards,
dedicated computer clusters, etc. This implies specific constraints in terms of latency, memory footprint, etc.
- Typical data analysis workflows also rely on large sets of simulated data, which require sizeable resources to be
produced.
This ERC project demonstrated how Deep Learning can facilitate this data processing, freeing computing resources to enlarge
the scientific program of the experiments and developing new techniques to extend their reach beyond traditional techniques.
Thanks to the work performed in the last five years, these techniques are now accepted by the community as mainstream part
of the scientific exploitation of LHC data
At the beginning, the project focused on establishing Deep Learning as a useful tool, using non controversial tasks (data quality assessment) as a show case. Besides producing the first results on this front, and deploying in production the first Deep Learning algorithm in the CMS experiment, we established a new technique for data quality monitoring that is now extensively generalized by a dedicated team of people, appointed by the management of the CMS experiment.
Once this first phase was completed, we focused our attention on developing Deep Learning models for event reconstruction (particle identification, energy measurement, etc.). To this purpose
- we explored many network architectures, giving an important contribution to establish Graph Neural Networks as the most natural choice for High Energy Physics.
- we considered several additional problems (event generation, pileup subtraction), beyond the usual HEP ML tasks
- we worked on develop tools to train and deploy these models in the infrastructure of a typical LHC experiments
In particular, we worked on two libraries:
- NNLO, a library to perform large-scale training on GPU clusters
- hls4ml, a library to deploy Neural Networks and Boosted Decision Trees on FPGA cards
Both could have impact beyond high energy physics, facilitating the work of other researchers, startups, and private companies in their R&D.
hls4ml was the topic of a POC project submitted to the ERC, with the idea of generalizing its features beyond HEP and making it a valuable tool for Neural Network edge computing and ASIC design.
An important part of the activity was devoted to develop unsupervised data analysis techniques, to extend the reach of new physics searches at the LHC beyond the list of new physics scenarios usually considered. By learning from data (e.g. through anomaly detection techniques), we wanted to go beyond the domain-specific cultural prejudice on how new physics should look like. Doing so, we are establishing an alternative path towards discoveries, which would complement the established techniques. Given how the LHC data taking is structured (and in particular, the presence of a real-time filtering system that selects only a small fraction of events for further analysis), it is essential to run these anomaly detection algorithms in real time. This is another reason why we were motivated to push for the development of hls4ml. The worked was successfully carried up to the promised prototype. Being ahead of scheduled, we could have actually deployed the prototype in the real data-taking system. Unfortunately the interruption of the LHC physics program in July 23 forced us to delay this exploitation step to next year. In 2024, the work done during mPP will start producing benefit for the CMS collaboration. Meanwhile, two papers documenting our studies on the real CMS data are in internal review. Due to the time of such a review, they will only be submitted to journal in 2024.
The work on generative models produced (in particular, the proof-of-concepts described in our publication) were the basis of a structured effort by the CMS collaboration to endorse these methods to reduce the computational cost of detector simulation for the High-Luminosity LHC. This plan is currently under discussion in the periodic review of the experiment activity, organized by CERN with a committee of external referees. The fact that the future of the experiment is being planned, based on what the mPP project studied, testifies the high quality of the final delivery of the project.
The last part of the period was also devoted to work on Quantum Machine Learning. We were interested to explore to which extent the emerging quantum technologies could be exploited on the kind of problems investigated by the mPP team =during the project. We used the anomaly detection problem as a benchmark and made a quantitative comparison study, including tests performed on an actual quantum computer (access privileges provided by CERN via IBM). The work was documented in a paper to be presented at the QTML conference (Nov '23) and submitted to Nature Communications in Physics.
Once this first phase was completed, we focused our attention on developing Deep Learning models for event reconstruction (particle identification, energy measurement, etc.). To this purpose
- we explored many network architectures, giving an important contribution to establish Graph Neural Networks as the most natural choice for High Energy Physics.
- we considered several additional problems (event generation, pileup subtraction), beyond the usual HEP ML tasks
- we worked on develop tools to train and deploy these models in the infrastructure of a typical LHC experiments
In particular, we worked on two libraries:
- NNLO, a library to perform large-scale training on GPU clusters
- hls4ml, a library to deploy Neural Networks and Boosted Decision Trees on FPGA cards
Both could have impact beyond high energy physics, facilitating the work of other researchers, startups, and private companies in their R&D.
hls4ml was the topic of a POC project submitted to the ERC, with the idea of generalizing its features beyond HEP and making it a valuable tool for Neural Network edge computing and ASIC design.
An important part of the activity was devoted to develop unsupervised data analysis techniques, to extend the reach of new physics searches at the LHC beyond the list of new physics scenarios usually considered. By learning from data (e.g. through anomaly detection techniques), we wanted to go beyond the domain-specific cultural prejudice on how new physics should look like. Doing so, we are establishing an alternative path towards discoveries, which would complement the established techniques. Given how the LHC data taking is structured (and in particular, the presence of a real-time filtering system that selects only a small fraction of events for further analysis), it is essential to run these anomaly detection algorithms in real time. This is another reason why we were motivated to push for the development of hls4ml. The worked was successfully carried up to the promised prototype. Being ahead of scheduled, we could have actually deployed the prototype in the real data-taking system. Unfortunately the interruption of the LHC physics program in July 23 forced us to delay this exploitation step to next year. In 2024, the work done during mPP will start producing benefit for the CMS collaboration. Meanwhile, two papers documenting our studies on the real CMS data are in internal review. Due to the time of such a review, they will only be submitted to journal in 2024.
The work on generative models produced (in particular, the proof-of-concepts described in our publication) were the basis of a structured effort by the CMS collaboration to endorse these methods to reduce the computational cost of detector simulation for the High-Luminosity LHC. This plan is currently under discussion in the periodic review of the experiment activity, organized by CERN with a committee of external referees. The fact that the future of the experiment is being planned, based on what the mPP project studied, testifies the high quality of the final delivery of the project.
The last part of the period was also devoted to work on Quantum Machine Learning. We were interested to explore to which extent the emerging quantum technologies could be exploited on the kind of problems investigated by the mPP team =during the project. We used the anomaly detection problem as a benchmark and made a quantitative comparison study, including tests performed on an actual quantum computer (access privileges provided by CERN via IBM). The work was documented in a paper to be presented at the QTML conference (Nov '23) and submitted to Nature Communications in Physics.
The project contributed to establish Deep Learning (DL) as a production-ready tool for High Energy Physics (HEP) experiments. When the project was approved, several proof-of-principle studies existed, discussing the potential of DL for HEP. On the other hand, no practical use of DL happened by then. Since the project started, the mPP team contributed to promote DL for HEP in several ways:
- By producing proof-of-principle studies showing novel applications of DL for HEP.
- By developing dedicated software libraries for DL training (NNLO) and deployment on HEP-specific computing environments (hls4ml).
- By investigating cutting-edge DP concepts (e.g. generative models, unsupervised learning, graph networks) on specific HEP problems.
- By extending the use of DL applications beyond the domain of data analysis, proposing it as a tool for even generation, data quality monitoring, trigger, etc.
By the end of the project, we developed applications based on DL to be used in real applications specific to the CMS experiment at the LH. By documenting this experience, we contributed to make DL a standard tool for HEP, so that the next generation of HEP experiments will be designed to maximally exploit the power of DL (e.g. with specific choices on detector geometry, computing infrastructure, electronic design, etc.). In the years to come, several CMS publication will exploit the outcome of the mPP project in exploiting LHC data for scientific publications.
- By producing proof-of-principle studies showing novel applications of DL for HEP.
- By developing dedicated software libraries for DL training (NNLO) and deployment on HEP-specific computing environments (hls4ml).
- By investigating cutting-edge DP concepts (e.g. generative models, unsupervised learning, graph networks) on specific HEP problems.
- By extending the use of DL applications beyond the domain of data analysis, proposing it as a tool for even generation, data quality monitoring, trigger, etc.
By the end of the project, we developed applications based on DL to be used in real applications specific to the CMS experiment at the LH. By documenting this experience, we contributed to make DL a standard tool for HEP, so that the next generation of HEP experiments will be designed to maximally exploit the power of DL (e.g. with specific choices on detector geometry, computing infrastructure, electronic design, etc.). In the years to come, several CMS publication will exploit the outcome of the mPP project in exploiting LHC data for scientific publications.