Daily Summary for 2026-11-15
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 fifteenth of November, twenty twenty-six, and this is the day's research.
Mira: 9 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 Search for Dark Photon Dark Matter at 74--159 mu eV/ c squared Using Broadband Cryogenic Receivers.
The day's research: Kai: The search for dark photon dark matter at very low energy levels is significant work.
Mira: It probes physics beyond the Standard Model by looking for hypothetical massive vector bosons.
Kai: Using cryogenic receivers, researchers looked for signals between seventy-four and one hundred fifty micro electron volts per c squared.
Mira: They found no statistically significant signal and set upper limits on the kinetic mixing parameter chi less than three point one times ten to the power twelve at a ninety-five percent confidence level over most of that search range.
Kai: This result improves upon previous limits by more than an order of magnitude in much of that energy window.
Mira: It provides direct constraints two orders of magnitude below cosmological bounds.
Lev: A partial wave analysis observed the first evidence for a resonance with quantum numbers J P equal to one minus one in chi c1 to K plus K minus eta decays.
Kai: This observation had an eleven point zero sigma statistical significance.
Mira: Which allowed researchers to measure the mass and width of this K 1 star at eighteen hundred sixty-nine MeV per c squared.
Lev: There is work on improving how top quark pairs are reconstructed using machine learning methods applied to CMS data.
Kai: Specifically, a residual multilayer perceptron and a conditional normalizing flow were developed to enhance the resolution by approximately thirty percent compared to previous methods.
Mira: This shows that neural networks can learn genuine physics-driven corrections even when dealing with intrinsic limitations imposed by neutrino approximations.
Lev: The most significant finding relates to how machine learning can help separate neutrino and antineutrino events in a large liquid scintillator detector.
Kai: Without a magnetic field, these detectors struggle to directly distinguish between them for probing CP violation.
Mira: This work investigates using logistic regression, XGBoost, and multilayer perceptron classifiers on variables like inelasticity, neutron multiplicity, and charged lepton energy to achieve this separation.
Lev: The XGBoost classifier performed quite well reaching an area under the receiver operating characteristic curve of about zero point eight three for both neutrino and antineutrino samples.
Kai: This means the model is reasonably good at telling them apart based on the input features used during training.
Mira: Under a square root approximation, this ML-based separation boosts peak CP violation sensitivity from one point four nine sigma to one point nine six sigma for normal mass ordering and from one point one four sigma to one point six one sigma for inverted mass ordering.
Lev: Search for Dark Photon Dark Matter at 74--159 mu eV/ c squared Using Broadband Cryogenic Receivers Dark photons are hypothetical massive vector bosons that couple to ordinary photons through kinetic mixing and this search found no significant signal.
Kai: Observation of a strange J P=1- resonance K+K-eta decays A partial wave analysis observed a new resonance in kaon eta decays with high statistical significance.
Mira: Beam-beam effects on measurements at electron-positron colliders: LEP neutrino counting and Higgs factories This paper investigates how beam interactions affect precision measurements at colliders and provides corrected luminosity predictions.
Lev: AlphaHits: Full-Event Tracking and Reconstruction Towards a Hit-Level Foundation Model This work introduces a Transformer model that reconstructs charged particles directly from detector hits for scalable physics analysis at the LHC.
Kai: Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data This paper develops machine learning methods to improve the reconstruction of top quark pairs by better handling the missing neutrino information.
Mira: Measurements of C!P violation with charm hadrons at Belle and Belle II Recent searches for CP violation in charm decays showed that all reported asymmetries are consistent with CP symmetry within current precision.
Lev: Sensitivity to Sterile Neutrinos with Elastic Neutrino-Electron and Highly Elastic Neutrino-Nucleus Scattering in the DUNE Near Detector This study explores how a near detector can be used to search for sterile neutrinos in neutrino oscillations.
Kai: Measurement of the branching fractions for the decays chi cJ to phi omega eta and chi cJ to phi omega eta Using BESIII data this paper reports the first observation and measurement of branching fractions for specific dark photon decays.
Mira: Machine learning assisted nu/nu bar discrimination in atmospheric neutrino events for CPV sensitivity in a 20-kton liquid scintillator detector This research shows that machine learning can help separate neutrino and antineutrino events to improve CP violation sensitivity in large detectors.
Lev: Measurement of the branching fractions for the decays chi cJ to phi omega eta and chi cJ to phi omega eta Using BESIII data this paper reports the first observation and measurement of branching fractions for specific dark photon decays.
Lucky paper: 2611.12774: Kai: So, we're moving into a paper titled Search for Dark Photon Dark Matter at seventy-four--one hundred fifty-nine mu eV/ c squared Using Broadband Cryogenic Receivers.
Mira: This research looks for dark photon dark matter that converts into photons at a metal surface based on kinetic mixing.
Kai: They used cryogenic K- and Ka-band receivers with a broadband spectrometer to cover the range of seventy-four to one hundred fifty micro electron volts per c squared.
Lev: The paper details the signal power received by the antenna as PDP equals chi squared times rho times Aeff alpha squared.
Mira: And they found no statistically significant signal in that entire search window, setting upper limits on chi less than three point one times ten to the power twelve at a ninety-five percent confidence level.
Kai: That result is quite interesting because it improves previous limits by more than an order of magnitude over most of that energy window.
Lev: Specifically, they provide the first direct limits over most of one hundred ten to one hundred fifty-nine micro electron volts per c squared, which is approximately two orders of magnitude below cosmological bounds.
Mira: The authors also mentioned systematic uncertainties, saying that the gain uncertainties contributed five point one percent and two point eight percent spectively to the uncertainty in chi.
Kai: And they factored in the uncertainty from local dark matter density as three point nine percent to that overall uncertainty in chi.
Lev: They calculated a smallest local p value of two point four three times ten to the negative seven at nu zero equal to three seventy-five point eight four nine three gigahertz.
Mira: Which corresponds to a global p value of zero point three eight, suggesting no significant excess was found.
Kai: So, looking at the methodology, they increased the integration time per frequency bin by a factor of two thousand for the same observing time and frequency coverage with that broadband spectrometer.
Lev: That change in integration time helps them get better data compared to previous searches using a spectrum analyzer.
Mira: It's also important to know that they excluded frequency intervals contaminated by identified spurious signals based on comparisons with fifty omega termination spectra and measurements at different LO frequencies.
Kai: That shows they were careful about filtering out noise or artifacts in the data set before drawing their final conclusions.
Lev: The core idea here is that if dark photon dark matter constitutes all local dark matter, this search constrains kinetic mixing chi using cryogenic receivers across a very wide energy range.
Mira: It gives us a direct constraint at scales far below what cosmology predicts for dark matter.
Kai: For someone just listening, what does setting an upper limit on kinetic mixing mean in plain terms?
Lev: It means that if dark photon dark matter is the source of local dark matter, the kinetic mixing parameter chi cannot be larger than three point one times ten to the power twelve with ninety-five percent confidence.
Mira: That’s a very strong constraint because it pushes that limit down by more than an order of magnitude compared to earlier studies in that energy window.
Kai: It's good they were able to cover such a broad range, from seventy-four up to one hundred fifty micro electron volts per c squared.
Lev: The authors also mentioned they assumed dark photon dark matter constitutes all local dark matter when setting those limits.
Mira: That assumption is key, because if the actual composition of local dark matter is different, the constraint on chi changes significantly.
Kai: If we only listen to this, it suggests that for the energy range they explored, there isn't a signal coming through these specific cryogenic receivers.
Lev: Exactly. No statistically significant signal was found in that search across seventy-four to one hundred fifty-nine micro electron volts per c squared.
Mira: It’s important to remember this is an upper limit, not a discovery of dark matter itself.
Kai: Right, so the paper confirms that based on their method and data, they haven't seen evidence for this specific dark photon interaction in that energy region.
Lucky paper: 2611.13222: Kai: We're looking at AlphaHits now, which is this paper about building a hit-level foundation model for LHC physics.
Mira: It uses a Transformer based approach, and it seems to be focusing on scalable full event reconstruction while also doing multi-task reconstruction and classification from low-level detector information.
Lev: The authors are using the TrackFormer and HiggsFormer architectures as a building block for AlphaHits, which is pretrained on charged particle tracking but then fine-tuned for multi-task tasks.
Kai: They mention that AlphaHits operates directly on detector hits, processing events containing up to one hundred fifty thousand hits under HL-LHC pileup conditions.
Mira: And they claim a TrackML score of zero point nine five nine when evaluated inclusively over all reconstructable tracks, which they say is the highest reported score for full-detector inclusive charged particle reconstruction so far.
Lev: They describe a two stage training process where Stage one trains the backbone as a tracker from scratch using only tracking and hit particle heads, and then Stage two adds seven other heads with fresh weights to train them jointly.
Kai: What about those auxiliary terms in the loss function? How does that prevent the model from collapsing?
Mira: They add fixed-weight auxiliary terms outside of the main objective to ensure a near zero scaffold loss cannot collapse it, which is pretty specific.
Lev: The main tasks for Stage two include event class, HT, E miss T, leptons, jets set with twelve jet queries and flavor prediction.
Kai: That sounds like a lot of things being learned at once. What's the performance comparison against baselines like the Particle Transformer?
Mira: AlphaHits achieves an AUC of zero point nine four four for Higgs event classification using silicon hits alone, which is comparable to what a Particle Transformer baseline reaches when using fully reconstructed inputs.
Lev: They also found that AlphaHits finds all six GenJets with the right flavor when using only hits, matching the Delphes track jet reference with seventy seven percent b-efficiency.
Kai: That's pretty strong for finding jets from raw hits, but what about the one shot model? What score does it get without the full fine-tuning?
Mira: The one shot model keeps the tracking head but reaches a TrackML score of zero point eight zero six with equal token decode and a double majority efficiency of zero point eight nine three for pT greater than zero point one five GeV.
Lev: They also compared it to the AlphaHits-TrackFormer Tracker, noting that the backbone is smaller, with twelve layers instead of eighteen layers, and nine point seven million parameters instead of nineteen point four million parameters.
Kai: So the complexity is reduced while still getting a good score? What about how they handled the input representation?
Mira: The hits are represented as short vectors of their position and local geometry, embedded by one linear layer into two hundred fifty six dimensional vectors and passed through a stack of Transformer encoder layers.
Lev: They also add learned embeddings of the detector volume layer and module, along with eight cluster shape features for every input hit before it goes through the models.
Kai: That means it's not just looking at where the hit is, but also what kind of detector region it's in and its local geometry.
Mira: They then predict for every hit its particle’s kinematics, production vertex displacement, and charge next to the track token using these representations.
Lev: The hierarchical reconstruction pipeline uses a learned Router that first assigns detector hits to overlapping logical groups before the full event tracker predicts track parameters for every hit.
Kai: And what does this router do exactly? Is it just grouping them or something more complex?
Mira: The Router is a small MLP that processes hits independently, and it predicts membership in two of four by four by thirty two groups based on hit radius, track pseudorapidity, and track azimuth.
Lev: This allows the full event attention to be computationally tractable and introduces a learned locality structure determining which detector regions can communicate directly.
Kai: So it's essentially creating a sparse attention structure across the hits based on where they are in space.
Mira: Exactly, and then a compact local Refiner resolves any combinatorial ambiguities that the router leaves behind.
Lev: The overall objective for the global stage is a multi-task combination of discrete track parameter prediction, hit classification, continuous track parameter regression, charge prediction, and routing supervision.
Kai: That's quite comprehensive for one model to handle all those things at once.
Mira: And they use AdamW optimizer with an initial learning rate of three times ten to the negative four and weight decay of one times ten to the negative four with cosine learning rate annealing for fifty epochs.
Lev: The training was done on an NVIDIA H200 card and took roughly one week for the best model, which is quite fast given the scale.
Kai: So this paper suggests that scaling sparse neural networks can learn physically informative representations directly from detector measurements instead of relying solely on explicit reconstruction pipelines.
Mira: That's what they are pointing towards, suggesting that end-to-end optimization of physics analyses directly from detector measurements could improve sensitivity beyond conventional reconstruction pipelines.
Lucky paper: 2611.09797: Kai: We're looking at Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data today.
Mira: The challenge there is reconstructing the top quark pair kinematics because you have two neutrinos that are completely undetected, which makes things really tricky.
Lev: This paper develops improved kinematic reconstruction methods, including machine learning approaches, to enhance the resolution of the reconstructed top quark pair invariant mass.
Kai: They compare their new methods against existing techniques like full and loose kinematic reconstruction, or FKR and LKR.
Mira: The authors introduce LKRv3 which uses an approximate equality of invariant masses and rapidities between the dilepton and neutrino-pair systems to improve on the older LKRv3.
Lev: The ML approaches they developed, specifically a residual multilayer perceptron and a conditional normalizing flow, improve resolution by approximately thirty percent compared to LKRv3 for the invariant mass.
Kai: That thirty percent improvement is significant because it shows where the machine learning actually finds physics-driven corrections instead of just fitting detector noise.
Mira: The paper mentions that these ML methods also improve the transverse momentum by twenty percent and rapidity by fifteen percent, and they show similar performance when trained on the same input information.
Lev: They are trained using an L1 loss, which is the mean absolute error, because it estimates the median of the targets for given inputs and gives a smaller bias.
Kai: So they are focusing on getting that median value right rather than just minimizing squared errors, which makes sense when dealing with these kinds of systematic uncertainties.
Mira: The generator-level analysis is interesting because they train the networks on true leptons, b quarks, and neutrino pair transverse momentum without using any CSV discriminants.
Lev: That generator-level result shows that the neural networks reduce the uncertainty in the invariant mass to fifty GeV, which is a factor of one point eight smaller than what analytical algorithms achieve alone.
Kai: So it seems like most of what was limiting them before came from how they approximated those neutrinos, not just detector resolution itself.
Mira: The paper also tests that the neural network learns a genuine kinematic correction to that neutrino-pair approximation, which is a key finding for this paper on Kinematic Reconstruction of Top-Quark Pairs Using Analytical and Machine-Learning Methods with CMS Open Data.
Lev: The study uses twenty-four variables as input for the neural networks, including the four-momenta of the leptons and jets selected by LKRv3, plus two CSV b-tagging discriminants.
Kai: That means they are using a lot of information to feed into this model, and I wonder how computationally expensive it is to run those networks in real time for analysis.
Mira: The summary says the neural network is approximately one order of magnitude less computationally expensive than the multilayer perceptron they developed earlier.
Lev: They also augment their training events by random rotation around the beam axis to increase diversity without changing the regression targets, which helps make sure it generalizes well.
Kai: So they are not just overfitting to specific event shapes, but trying to learn a general correction that applies across different data samples.
Mira: The paper concludes that the achievable resolution is limited primarily by the information available in the reconstructed event and the intrinsic limitations of the neutrino-pair approximation, rather than by the capacity of the neural network.
Lev: And they confirm this limitation in their generator-level study, showing that it's really about what you can reconstruct from what you see.
Lucky paper: 2611.11474: Kai: We're looking at paper number four of four today, and it focuses on using machine learning to help separate neutrino and antineutrino events in a large liquid scintillator detector to improve CP violation sensitivity.
Mira: It’s interesting because without a magnetic field, these detectors can’t see the difference between neutrinos and antineutrinos directly for probing CP violation.
Lev: The authors are using the XGBoost classifier to separate neutrino and antineutrino events based on features like inelasticity, neutron multiplicity, and charged lepton energy.
Kai: They reported an AUC of about zero point eight three for both flavor samples, which means the model is reasonably good at telling them apart based on what they fed it during training.
Mira: And that separation actually boosts the peak CP violation sensitivity from one point four nine sigma to one point nine six sigma for normal mass ordering.
Lev: They also showed that under inverted mass ordering, the peak sensitivity goes from one point one four sigma up to one point six one sigma with this machine learning assistance.
Kai: So, what’s the actual improvement they are claiming here?
Mira: The ML assisted method improves the peak CP violation sensitivity by using a Poisson likelihood with systematic nuisance parameters and marginalization over five oscillation parameters within their respective three sigma ranges.
Lev: In the ideal separation scenario, they predict peak CPV sensitivity values of six point nine eight for normal mass ordering and five point four three for inverted mass ordering.
Kai: That’s a big jump from the charge blind baseline they were comparing it to earlier in their study.
Mira: But the authors have a caveat here, saying these preliminary results are based on detector response approximated by parameterized smearing and a simplified neutron tagging model.
Lev: They acknowledge that more realistic modeling of neutron detection, event reconstruction, backgrounds, and correlated systematics is required to assess the robustness of this approach.
Kai: So it’s a promising direction for improving CPV sensitivity even with charge blind detectors but the accuracy depends heavily on how well you model those detector effects.
Mira: Exactly. The study shows that ML based statistical discrimination is a promising way to improve CPV sensitivity of a large liquid scintillator detector.
Lev: It’s important to remember the specific features they used; neutron multiplicity was the dominant feature, contributing about seventy-seven percent and seventy-nine percent for the muon and electron flavor samples respectively.
Kai: That tells us that what is happening in the way neutrons are produced in those collisions is really driving this classification.
Mira: It suggests that understanding event characteristics beyond just energy or direction is crucial when you're working with detectors that lack magnetic fields.
More episodes
- Daily Summary for 2026-11-19
- 2512.21290-Model-independent ZH production cross section at FCC-ee
- 2610.21982-First Observation of the Radiative Decay D 0 to gamma omega
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- Daily Summary for 2026-11-18
- 2512.08065-Searches for Light Dark Matter and Evidence of Coherent Elastic Neutrino-Nucleus Scattering of Solar Neutrinos with the LUX-ZEPLIN (LZ) Experiment
- Daily Summary for 2026-11-18
- 2508.18022-Precision Measurements of the Electroweak Mixing Angle in the Region of the Z pole at the Large Hadron Collider
- 2512.23992-Prospect for measurement of CP-violating parameters of B s 0 to phi gamma at the Tera Z factory