Daily Summary for 2026-11-09

daily

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: It's the ninth of November, twenty twenty-six, and this is the day's research.

Mira: 18 new papers came out today.

Kai: I'm Kai.

Mira: And I'm Mira, here with Lev.

Kai: We'll take the day in one pass, then pull out the papers we're staying with.

Mira: Start us off with Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment.

The day's research: Kai: The work focuses on managing and reproducing complex physics analyses. We need to track how results come from initial data and code.

Mira: Celebi addresses this by separating analysis definitions from generated data products. It uses an Analysis Repository for preserved states and a Production Factory for execution history.

Lev: Preserving immutable object states links results clearly to their original settings, building on the modular workflow architecture described in CELEBI: Design and Realization of a Reproducible Physics Analysis Framework.

Kai: That organization is key because it allows tracking history across evolving analyses, connecting those preserved states to Yuki's execution infrastructure.

Mira: Regarding experimental results, progress exists in high-precision measurements of fundamental constants. BaBar measures the e+e-to pi+ pi- cross section using initial-state radiation to cancel systematic uncertainties.

Lev: That approach yielded a leading-order hadronic contribution to the muon magnetic anomaly of 517.8 plus or minus 2.4 stat plus or minus 3.3 syst in units of-10-10, which is significant.

Kai: On a practical level, researchers are refining data quality monitoring in facilities like ALICE using a detector trained on luminosity block residuals to find 43% of injected defects at a one percent false-positive rate.

Mira: That performance outperformed matched-filter scans described in Data-quality monitoring of luminosity blocks with ATLAS Open Data: an open benchmark and a learned within-run detector.

Lev: The most important development involves applying reinforcement learning to control systems like accelerator operations and data quality monitoring, shifting toward reward-guided policies.

Kai: That survey reviews how reinforcement learning agents learn policies through interaction with environments guided by scalar rewards in seven major HEP domains.

Mira: Today's papers include CELEBI: Design and Realization of a Reproducible Physics Analysis Framework presenting a framework for organizing analyses while separating definitions from data products.

Lev: Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment details a transformer-based algorithm for distinguishing prompt leptons from non-prompt ones in the ATLAS experiment.

Kai: Data-quality monitoring of luminosity blocks with ATLAS Open Data: an open benchmark and a learned within-run detector constructs a public benchmark to monitor data quality at the LHC using a learned detector.

Mira: Commissioning and Performance of the LHCb Upstream Tracker presents an overview of the new high-precision silicon strip detector installed in the LHCb experiment.

Lev: A New Precise Measurement of the e+e-to pi+ pi- Cross Section with the Initial-State Radiation Method at BaBar reports precise results for a cross section using an initial-state radiation method at BaBar.

Kai: Operational Challenges in Efficient and Sustainable Operation of the Event Processing Nodes GPU Farm at the ALICE Experiment provides a description of CERN's largest GPU computing facility.

Mira: Projected sensitivity of the detector to photophobic and fermion-coupled axion-like particles presents the projected sensitivity of the ANUBIS detector to searching for axion-like particles in different models.

Lev: LH mu C Concept: Energy-Frontier Antimuon-Proton Collisions at CERN proposes a concept for a TeV-scale lepton–hadron collider based on the HL-LHC to provide PDFs.

Kai: A High Performance Partial Wave Analysis Framework for Hadron Spectroscopy introduces a high performance framework to efficiently perform partial wave analysis of hadron spectroscopy data.

Mira: Reinforcement Learning in High-Energy Physics: A Survey from Accelerator Control to Theory Exploration reviews applications of reinforcement learning across seven major domains in high energy physics.

Lev: Search for Higgs boson production in association with another Higgs or an exotic scalar boson in the b b barW+W- single lepton final state at sqrt s = 13 TeV with the ATLAS experiment presents a search for Higgs boson pair production using ATLAS Run 2 data.

Kai: Search for dark matter axions at 1-2 micro-eV with a RADES-DarkQuantum haloscope at CERN describes a proposal to search for galactic axion and axion-like particle dark matter using a haloscope at CERN.

Mira: Amplitude Analysis of the Invariant-Mass Spectra of Centrally Exclusively Produced pi+ pi- and K+K- Pairs in pp Collisions at sqrt s =510 GeV presents an amplitude analysis of invariant-mass spectra from central exclusive collisions.

Lev: First broadband search for fixed-polarization dark photon dark matter with GigaBREAD performs a broadband search for a fixed-polarization dark photon using GigaBREAD data.

Kai: World average charged-current inclusive neutrino cross section from 6 to 200 GeV applies Methods of the Particle Data Group to produce a world average inclusive, charged-current neutrino cross section.

Mira: Measurements of isolated photon pair production in association with jets in pp collisions at sqrt s = 8 TeV with the ATLAS detector presents measurements of diphoton production in association with jets using ATLAS data at sqrt s = 8 TeV.

Lev: Measurement of the V cb element of the CKM matrix using t t bar events in pp collisions at sqrt s = 13 TeV performs a measurement of the Cabibbo - Kobayashi - Maskawa matrix element using top quark pair events.

Kai: Search for heavy neutral leptons in top quark decays with same-sign, same-flavor dilepton final states in proton-proton collisions at sqrt s = 13 TeV reports a search for a Majorana-type heavy neutral lepton in top quark decays.

Lucky paper: 2611.08194: Kai: Alright, we're looking at Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment today. This paper by the ATLAS Collaboration is about improving how they find prompt leptons.

Mira: So, this Prompt Lepton Isolation Tagger or PLIT algorithm uses a transformer neural network to distinguish those prompt leptons from non-prompt ones that come from hadron decays or misidentified particles.

Lev: It leverages information from associated charged-particle tracks combined with lepton properties to identify correlations that enhance discrimination.

Kai: What's the main performance claim here? How does it stack up against the existing methods they are comparing it to?

Mira: They evaluated PLIT using simulated top quark pair production events against existing lepton isolation algorithms, and it showed superior prompt lepton efficiency while keeping a high rejection of non-prompt leptons.

Lev: This algorithm is specifically designed to separate prompt leptons from those predominantly produced in semileptonic decays of heavy-flavor hadrons.

Kai: That sounds like a big deal for cleaner physics signals. How does the architecture itself help achieve this separation?

Mira: The architecture of PLIT is based on the GN2 transformer architecture used by ATLAS for heavy-flavor jet tagging, but they adapted it for lepton isolation.

Lev: The network processes kinematic and detector properties of the lepton alongside information from up to forty tracks within a cone around the candidate lepton with a delta R less than zero point four.

Kai: That's a lot of input data for one prediction; what is the output they are looking at?

Mira: The network outputs a prompt lepton probability, called p prompt, and then the PLIT discriminant is constructed as D PLIT equals p prompt divided by one minus p prompt.

Lev: They showed that for muons at low p T, the rejection is one point five to two times larger than the previous algorithm for any given efficiency.

Kai: A factor of two or three improvement in rejection sounds substantial across the board. What about electrons?

Mira: For electrons, they found a rejection gain of a factor of two or more across both p T ranges, reaching up to a factor of five for endcap electrons at low efficiency.

Lev: In terms of prompt lepton efficiency for the same level of rejection, PLIT achieves relative improvements: ten point zero percent for muons, fifty percent for barrel electrons, and one hundred percent for endcap electrons at low p T ranges.

Kai: That zero point zero percent improvement for endcap electrons at low p T is interesting. Where does the uncertainty in this measurement come from?

Mira: The dominant uncertainty stems from the modeling of hadronic activity in the simulation, specifically the difference between nominal Powheg plus Pythia and an alternative Sherpa sample.

Lev: All other systematic uncertainties, including template shape, QCD multijet background description, and fit range choice called lambda SC remain sub-dominant below zero point one percent over the full p T range.

Kai: So it seems the main hurdle right now is getting the simulation right to fully trust these results. What does this mean for those of us who have to run models?

Mira: It means we need very accurate modeling of hadronic activity in our simulations if we want to rely on these performance improvements.

Lev: The key thing is that PLIT doesn't rely on jet reconstruction or jet flavor-tagging tools, which simplifies the pipeline compared to older methods.

Lucky paper: 2611.07819: Kai: We're looking at Data-quality monitoring of luminosity blocks with ATLAS Open Data: an open benchmark and a learned within-run detector today, which is number two of five papers we're covering this day.

Mira: It seems like the paper tackles a real problem: how to decide which collision data gets used for physics analyses when the experiment's own monitoring data isn't public.

Lev: The paper constructs a public benchmark using thirty-nine thousand eight hundred twenty-eight luminosity blocks from fifty-five thousand six hundred twenty PHYSLITE files, where eight types of physics-motivated detector defects are injected at the event level in those same events.

Kai: So they aren't just looking at raw data; they're simulating real detector issues right into the event stream.

Mira: And to train the model, they use a rolling split protocol where a detector gets trained and calibrated on data recorded before the data it monitors, using earlier runs as reference sets.

Lev: The alarm threshold at a false-positive rate of alpha equals one percent is set by the split-conformal quantile of the scores from those clean reference luminosity blocks in every fold.

Kai: So they're setting that alarm level very carefully based on how well things look in the earlier runs.

Mira: In the learned detector, they compute features for every luminosity block, including "mean squared pull," and the largest positive and negative window sums at several widths.

Lev: The classifier itself is a histogram-based gradient-boosted decision-tree ensemble trained on those reference runs and synthetic defects introduced in those runs.

Kai: The paper shows that references built across different runs fail because "Histogram shapes differ far more between runs than within a run," with a median chi square divided by degrees of freedom of five point nine against one point one.

Mira: That difference in histogram shapes is what makes the comparison between different reference sets unreliable, which is why they focus on residuals with respect to each luminosity block's own run as the representation where defects become visible.

Lev: When they apply the learned classifier to these residuals, it finds forty-three percent of the luminosity blocks with injected defects at a one percent false-positive rate, compared to twenty-five percent for a matched-filter scan.

Kai: That forty-three percent detection efficiency is what they are highlighting when they talk about the temporal context configuration or CTX achieving that detection efficiency.

Mira: But the study points out limitations, noting that benign real-data anomalies inside certified runs and defects whose signature is weaker than the statistical precision of one luminosity block limit all detectors at a one percent false-alarm rate.

Lev: Specifically, tile-module failures and weaker electron or photon scale shifts are below the information content of object-level histograms at one hundred four events per luminosity block.

Kai: And they mention that the generic detector failure prior raises efficiency on unseen types from seventeen point nine percent to thirty-three percent, which is a notable jump for those rare cases.

Mira: So while it's good at finding known defects, the method has limits when dealing with anomalies that aren't exactly what it was trained on or when the anomaly is very subtle.

Lev: It shows how machine learning can improve filtering noisy data streams before they are used in critical physics analyses by finding those subtle differences in the residuals.

Lucky paper: 2611.08150: Kai: So we're looking at Commissioning and Performance of the LHCb Upstream Tracker, Lev, what is the main practical change this new hardware brings to reconstructing those charged particles?

Lev: It brings a new high-precision silicon strip detector installed in the LHCb experiment as part of the overall LHCb Upgrade I project. This detector plays a crucial role in reconstructing charged particles through a fully software-based trigger.

Mira: What does this hardware actually look like? I want to understand how it helps with that reconstruction you mentioned.

Lev: The Upstream Tracker consists of four planes of silicon strip detectors, oriented at zero degrees for UTaX, negative five degrees for UTaU, positive five degrees for UTbV, and zero degrees again for UTbX. They are built from vertically-mounted double-sided staves.

Kai: That orientation sounds specific to optimizing resolution in the bending xz plane of the magnet. What about how they read out those sensors?

Lev: The sensors are read out by one hundred twenty-eight channel ASICs implemented in one hundred thirty nm CMOS technology, which we call SALT. This SALT ASIC uses a differential low-power six bit Successive Approximation Register ADC digitizes signals at forty MHz.

Mira: And what is the digital backend doing with that data stream? I need to know how they are handling the raw signals before reconstruction starts.

Lev: The digital back end of the SALT ASIC includes a digital signal processing for pedestal subtraction, common mode subtraction CMS, and zero suppression. They also perform calibration procedures to optimize configurable parameters and monitor the optimal working point and electronic noise.

Kai: That sounds like a lot of real-time processing power needed for high luminosity runs. How does this affect the track reconstruction process itself?

Lev: The LHCb tracking system reconstructs several classes of tracks according to subdetectors, as illustrated in Figure six. The three principal track types containing Upstream Tracker hits are long, downstream, and upstream tracks.

Mira: Long and downstream tracks sound like they come from different parts of the detector setup. Can you clarify how those two types are reconstructed?

Lev: Long tracks traverse the full tracking system and generally have hits in the VELO, Upstream Tracker, and SciFi detectors. Downstream tracks are reconstructed from seven SciFi seeds by extrapolating track segments to the Upstream Tracker.

Kai: Extrapolating track segments sounds like a tricky step where errors can build up quickly. How accurate is that momentum resolution after the High Level Trigger two reconstruction of downstream tracks?

Lev: The momentum resolution achieved after the High Level Trigger two reconstruction of downstream tracks is comparable to that achieved in Run two which is approximately one percent.

Mira: So it's matching previous performance while adding this new detector layer. What about the efficiency of this Upstream Tracker across different physics variables?

Lev: The Upstream Tracker efficiency versus transverse momentum p T and pseudorapidity eta shows an about two percent decrease near eta equals two point eight, which is due to a set of locally inefficient regions in the Upstream Tracker.

Kai: A two percent drop near that specific pseudorapidity is something we need to keep an eye on when planning future runs. What made the inclusion of the Upstream Tracker essential for reaching the design luminosity?

Lev: The inclusion of the Upstream Tracker in the High Level Trigger one pattern recognition proved essential for LHCb to reach its design luminosity of two point zero times ten cubed cm minus two s minus one.

Mira: So it was a necessary component to hit that target. Does this detector also help with monitoring stability during operation?

Lev: The Low Voltage system supplies regulated low voltage power to the hybrids at one point two volt and the DCBs at one point five volt and two point five volt, serving a total of one thousand four hundred eight channels via two hundred sixty-eight identical eight channel low voltage regulator boards.

Kai: That level of precise power regulation sounds critical for maintaining detector stability under load. What kind of safety systems are in place for the hardware itself?

Lev: The Upstream Tracker safety system, which we call DSS, runs autonomously and is built around industrial PLCs connected to numerous independent sensors. A single event upset SEU events in either the baseline DAC or pedestal registers are expected to occur at a rate of about three.

Mira: So it's a robust system designed for long-term operation under heavy data load. What would be the cost or complexity if we tried to run this setup on a much smaller, lower budget computing farm?

Lev: The paper doesn't discuss the cost directly, but it emphasizes that the digital signal processing in the SALT ASIC handles common mode subtraction and zero suppression for noise management.

Lucky paper: 2611.06586: Kai: We're looking at Operational Challenges in Efficient and Sustainable Operation of the Event Processing Nodes GPU Farm at the ALICE Experiment today.

Mira: This paper talks about ALICE’s Event Processing Nodes, which is CERN’s largest GPU computing facility by capacity, with three hundred fifty GPU-based nodes and two thousand eight hundred GPUs.

Lev: It successfully enabled real-time processing of heavy-ion collisions at fifty kHz during LHC Run three.

Kai: That farm is central to the Online–Offline framework, which unifies data acquisition and offline reconstruction through continuous readout.

Mira: The operational core is the EPN farm, with approximately three hundred fifty nodes interconnected through one hundred gigabits per second InfiniBand and one gigabit per second Ethernet for management services.

Lev: The principal reconstruction workload runs in the GPU barrel-processing chain, and the reconstructed data goes to an entropy encoder which produces CTFs.

Kai: GPU acceleration substantially reduces the required server count and energy consumption, achieving a power usage effect of one point zero five to one point one five for cooling and infrastructure overheads.

Mira: The system uses three hundred fifty nodes with two thousand eight hundred AMD Instinct MI50 GPUs, where a single MI50 GPU replaces roughly eighty AMD Rome CPU cores in synchronous processing.

Lev: They achieved a PUE of one point zero five to one point one five, and only five to fifteen percent of the total power was used for cooling and other infrastructure overheads.

Kai: The operational sustainability aspect involves using modular, energy-efficient infrastructure with adiabatic air cooling and on-site reverse osmosis water treatment.

Mira: The modular container-based design allows for incremental expansion and includes IT and cooling containers, a dedicated power container with dual two point five megawatt transformers, a reverse osmosis plant, and spare equipment facilities.

Lev: During Run three the ALICE Event Processing Node farm demonstrated reliable synchronous processing under the demanding conditions of Pb–Pb data taking.

Kai: In two thousand twenty-four and two thousand twenty-five the instantaneous luminosity was successfully leveled at fifty kHz, confirming that the online computing infrastructure could sustain those expected interaction rates.

Mira: The successful operation of GPU-based synchronous reconstruction was essential to ALICE Run three enabling online calibration, reconstruction, and compression at the nominal fifty kHz Pb–Pb interaction rate and approximately seven hundred gigabytes per second input to the reconstruction farm.

Lev: The EPN farm produces compressed Time Frames or CTFs, which are the only online data products retained for later processing.

Kai: These CTFs and calibration data are transferred to the central one hundred fifty petabyte disk-storage system, which provides three hundred sixty gigabytes per second throughput with approximately twenty five percent redundancy.

Mira: The raw Time Frames cannot be permanently stored on the farm itself.

Lev: The EPN Data Distribution system is responsible for collecting detector sub-Time Frames or sTFs, transporting them to the EPN farm, and assembling complete Time Frames or TFs.

Kai: At the First-Level Processing servers, incoming detector raw data are aggregated into sub-Time Frames by the sTFBuilder which organizes those data blocks from the acquisition logic over multiple LHC orbits.

Mira: That description of how they handle the data flow shows how they manage that continuous stream of raw Time Frames transferred to the EPN farm for online processing in Operational Challenges in Efficient and Sustainable Operation of the Event Processing Nodes GPU Farm at the ALICE Experiment.

Lev: What I don't get is how they handle those raw TFs that can't be permanently stored, since they are used for online processing.

Kai: That’s because they are only retained as CTFs, and the raw TFs cannot be permanently stored on the farm itself.

Mira: It sounds like a very tightly controlled pipeline where everything has to be compressed into those Time Frames before it moves to long-term storage or analysis.

Lev: The PUE of one point zero five to one point one five, while good for energy use, does that mean the cooling infrastructure is the main controllable variable there?

Kai: It shows how much GPU acceleration reduces the required server count by a factor of eight and halves the energy consumption compared to a CPU-only implementation.

Mira: The modular container-based design enabling incremental expansion is important for future scaling, though it requires careful management of those IT and cooling containers.

Lev: I wonder if that modularity helps when you think about extending the infrastructure through Run four which is mentioned in the paper.

Kai: It suggests a long-term vision where the IT infrastructure life can be extended through Run four using that modular approach for expansion.

Lucky paper: 2611.01117: Kai: We're looking at the LH mu C Concept: Energy-Frontier Antimuon-Proton Collisions at CERN today, which is paper number five of five for this discussion.

Mira: This concept is presented because TeV-scale lepton–hadron colliders are needed to provide PDFs for an adequate interpretation of the HL-LHC data, as well as future machines like the FCC-hh and SppC.

Lev: Lepton–hadron collisions have played an exceptional role in revealing the internal structure of matter, especially in the region of small Bjorken x below ten GeV squared that HERA didn't cover.

Kai: Muon–proton colliders can reach much higher energies, and this paper looks at how muon–proton colliders could provide those PDFs for interpreting the HL-LHC data.

Mira: The LH mu C antimuon complex has six stages: a three GeV, two MW rapid-cycling synchrotron operating at fifty Hz.

Lev: Then there's a pion-production ring where each proton bunch crosses a ten millimeter graphite target forty times, which yields zero point zero one six pions per proton.

Kai: The next part is the ultra-cold mu plus source where those pions are stopped in a multi-layer tungsten target and the thermal muonium is laser ionized.

Mira: That's followed by a linac re-accelerating them to two hundred twelve MeV with a normalized emittance epsilon n equal to four micrometer, in trains of forty bunches of four point five times ten to the ten plus mu plus particles.

Lev: The booster then accelerates them up to one TeV in about one millisecond before they collide with LHC protons at the main ring's interaction point.

Kai: So it’s basically a complex six stage machine that aims to get muon plus particles colliding with LHC protons at high energy.

Mira: And the paper notes that this option awaits the solution of the muon cooling problem, which is expected in the 2030s.

Lev: The concept relies on an ultra-cold antimuon option where you create a low emittance mu plus beam using ultra-cold muons, and that yield is about one hundred five times higher than at the present J-PARC MLF beamline.

More episodes

← Home