Opendata, web and dolomites

Report

Teaser, summary, work performed and final results

Periodic Reporting for period 1 - TNT (Transformations with Neutrals and Turbo analyses)

Teaser

The cost of computing power is in excess of a hundred million euros per year for the LHC. Even with such vast sums, resources still are the limiting power to accuracy of the measurements. This is particularly true for the LHCb experiment, which collects more data than can be...

Summary

The cost of computing power is in excess of a hundred million euros per year for the LHC. Even with such vast sums, resources still are the limiting power to accuracy of the measurements. This is particularly true for the LHCb experiment, which collects more data than can be stored. Novel methods are therefore required in order to make best use of the available computing resources. As a result of this project, in real-time, analysts are able to save any required part of an interesting event. The project has also resulted in a new machine learning package, now freely available, called NNDrone. The package allows the conversion of artificial intelligence to a faster more standardised form. The advances of the project potentially have wide implications for society.

The TNT project has automated one of the most complex analyses currently performed in high energy physics, namely a decay time dependent matter-antimatter asymmetry violation measurement. The project also lays the foundations for new searches for physics beyond the standard theory that have the power to confirm previously measured LHCb hints of deviations from standard model predictions. If the discrepancy is confirmed, it will have profound implications on our understanding of the origins of our universe.

Work performed

Q3 2017:

One of the key aims of the project is to enable fully flexible data processing. What this means in reality is that for every collision that takes place in the detector, to be most efficient we want to only save the particles and sub-detector data that will be used for measurements.

A use case has also been implemented to select decay signatures of the form of the Bd->KS μμ Χ in real time, where X denotes more particles consistent with originating from the B decay vertex. This allows us to analyse one of the signature analyses of the TNT project, namely Bd->K*(->KS π0)μμ in real time.

Q4 2017:

For the creation of our machine learning classifiers, an analyst may have a preference that is not always supported for execution in real time. A toolkit was therefore created such that a drone artificial neural network could learn the properties of an analysts preferred neural network and then be used in the real-time framework.

In the paper, we created 2 datasets consisting of a beauty decay and a charm decay. We then trained a neural network with the SciKit-Learn package. A separate drone neural network was created and taught the properties of the SciKit-Learn neural network, demonstrating a machine learning from another machine.

Q1 2018:

An unexpected rerun of the offline step at the end of 2017 gave the opportunity to put in some improvements to the isolation of Bd->KS μμ Χ. While this rerun was taking place, time was available to perform a measurement of a so-called golden mode for LHCb: Bs->φφ.

Q2 2018:

In the HEPDrone project, together with Konstantin Gizdov, another example was prepared to go with the B physics example presented in Q4 2017 (work ongoing to make the tools more user friendly). In the new example, a simple jet classifier was made using Keras.

In preparation for 2018 data taking, we also put in place new neural network based classifiers to the real-time selection of LHCb making use of the HEPDrone tools to port the model to the LHCb production framework. The improved performance of the neural network allowed us to expand the mass range from just looking at the B meson window to a 15GeV wide window enabling sensitivity to axions.

Q3 2018:

With the use of a package known as snakemake, work began to automate the analysis pipeline for the analysis of the Bs→φφ decay (see Q1 2018). Q3 2018 also saw the Tau lepton workshop come to Amsterdam. My talk gave an update on the most recent results of lepton flavour non-universality, i. e. the amount at which decays to electrons (found in atoms around us) and the heavier versions (muons and tauons) differ from each other.

Q4 2018:

The snakemake package (see previous quarter), was used to expand the automated analysis chain from the 8 steps in the previous quarter, to 14 now. At the time of writing, the analysis steps from dataset preparation to final results and plots are automated, in addition to a portion of the systematic studies. Systematic studies are checks of how the result changes due to certain choices during the analysis. It is a broad term that encompasses many cross-checks.

The HEPDrone package has now been renamed to be NNDrone and has now been incorporated in the Scikit-HEP family of packages. The features of NNDrone have been used to select signatures of the X→γγ decay in real time, where X can be a Bs meson or a new undiscovered particle. A draft of a document known as a public note (freely available peer-reviewed CERN documentation of a feature or study) has been drafted and will be available shortly.

Q1 2019:

The automation has now been taken to a new level and has increased the number of steps to over 70.
Additionally, the public note envisioned in the previous quarter on the use of NNDrone to select signatures of the X→γγ decay in real time has been changed to be a publication, with the addition of sensitivity studies for various levels of LHCb data. For the discovery of a new undiscovered particle (known

Final results

Turbo 2.0
For the first time in high energy physics a fully flexible saving framework is available in the real-time analysis infrastructure. This project an essential component to allow all analyses to use the real-time model.

NNDrone
This project has enabled a novel conversion of neural networks from one format to another and has demonstrated machine to machine learning. As a result, the vast ecosystem of machine learning software can be exploited in real-time analysis of LHCb data. The success and novel nature of the package has led to its inclusion in the SciKit- HEP family of packages. The package has wider uses for any application of machine learning and is not limited to high energy physics.

Selection frameworks
This project has given LHCb sensitivity to ALPs in a region of phase space to which no other experiment has access.
This project has laid the foundations for new measurements of inclusive semi-leptonic decays that were previously unexplored and which are very exciting in the search for new physics.

Automation of a TD analysis
The project has not only provided the most accurate measurement of CP violation in one of LHCb’s most important channels, but has also introduced an unprecedented level of automation into one of the most complex analyses LHCb can perform.

Website & more info

More info: http://h2020-tnt.com.