BEGIN:VCALENDAR VERSION:2.0 PRODID:-//CERN//INDICO//EN BEGIN:VEVENT SUMMARY:Noise suppression for Baikal-GVD using spectral properties of even t graphs DTSTART;VALUE=DATE-TIME:20261029T143500Z DTEND;VALUE=DATE-TIME:20261029T145000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6363@cern.ch DESCRIPTION:Speakers: Alexander Avrorin (INR RAS)\nBaikal-GVD neutrino tel escope is located in the Southern basin of Lake Baikal with the instrument ed depth ranging from 730 to 1275 m. The measured noise rates of the detec tor channels range from ten to hundreds of kilohertz\, depending on season and the depth of the installed photomultiplier. We present an evolution o f the method currently employed for muon track event reconstruction in Bai kal-GVD. Like in the previous approach\, it represents events as graphs\, with vertices corresponding to photomultiplier hits and edges correspondin g to a causality condition imposed by the Cherenkov cone. However\, instea d of identifying the vertices of fully connected subgraphs with Cherenkov photons\, we use the graph's spectral properties to identify noise hits an d provide an initial estimate of the muon direction. This increases collec tion efficiency for delayed photons from cascades along the track which be come dominant at higher energies. The proposed approach provides a 20% inc rease in Cherenkov hit collection efficiency for an astrophysical neutrino sample and a reduction in processing time\, while maintaining a noise rej ection rate comparable to that of the previous method.\n\nhttps://indico.p article.mephi.ru/event/626/contributions/6363/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6363/ END:VEVENT BEGIN:VEVENT SUMMARY:Machine learning for cosmic ray classification at the Carpet-3 exp eriment DTSTART;VALUE=DATE-TIME:20261029T132000Z DTEND;VALUE=DATE-TIME:20261029T133500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6206@cern.ch DESCRIPTION:Speakers: Nikita Pozdnukhov (INR RAS)\nWe discuss neural netwo rks for cosmic ray classification at the Carpet-3 experiment. We improve g amma ray sensitivity and obtain new limits on the isotropic diffuse gamma ray flux from Carpet-3 data\, improving previously obtained results. We al so study the possibility of measuring the cosmic ray mass composition at t he knee region at the Carpet-3 experiment\, comparing standard reconstruct ion methods to machine learning analysis. We address the domain shift betw een Monte Carlo simulations and real-world data by with domain adaptation and study its effects on the neural network performance.\n\nhttps://indico .particle.mephi.ru/event/626/contributions/6206/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6206/ END:VEVENT BEGIN:VEVENT SUMMARY:Reconstruction of neutrino-induced muon energy in Baikal-GVD with transformer neural networks DTSTART;VALUE=DATE-TIME:20261029T133500Z DTEND;VALUE=DATE-TIME:20261029T135000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6208@cern.ch DESCRIPTION:Speakers: Albert Matseiko (MIPT\, INR RAS)\nBaikal-GVD is a de ep underwater Cherenkov telescope at the bottom of Lake Baikal\, built to detect high-energy astrophysical neutrinos. The energy of the neutrino is one of the key parameters of a recorded event\, which makes its accurate m easurement an important task of the experiment.\n\nWe apply transformer ne ural networks to the reconstruction of the energy of a muon produced by a muon neutrino\, in events confined to a single cluster of the telescope. T he energy is reconstructed at the point of the muon track where it enters the sensitive volume of that cluster. The input of the network is the sequ ence of optical module hits recorded in the event\, and the network is tra ined on Monte Carlo simulation.\n\nTraining uses unsupervised domain adapt ation of the DANN type\, which prevents the network from distinguishing si mulation from experimental data. The loss function includes a negative log -likelihood term (beta-NLL)\, so that the network predicts not only the en ergy but also its own uncertainty for each event. Extension to the joint r econstruction of energy and muon track is planned.\n\nhttps://indico.parti cle.mephi.ru/event/626/contributions/6208/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6208/ END:VEVENT BEGIN:VEVENT SUMMARY:Forecasting the Onset Time of the Second Maximum of Proton Enhance ments Using Solar-Wind Kinematics and a Quantum Regression Neural Network (RQNN) DTSTART;VALUE=DATE-TIME:20261029T135000Z DTEND;VALUE=DATE-TIME:20261029T140500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6209@cern.ch DESCRIPTION:Speakers: Igor Davydenkov ()\nThe aim of this work is to forec ast the onset time of the second maximum of proton enhancements associated with the transport of particles by solar-wind structures. The "zeroth-ord er" forecast is based on the kinematic parameters of the solar wind measur ed at the L1 Lagrange point\, and is subsequently refined by machine-learn ing methods. The difficulty of this problem lies in the nonlinear\, non-st ationary relationships between proton-flux and solar-wind parameters\, on the one hand\, and the lack of a training base large enough to build a con ventional regression or neural-network model\, on the other. We therefore used a quantum regression model\, RQNN\, in which the input heliophysical parameters are encoded as phase shifts of the wavefunctions of emulated qu bits. As an example\, consider the prediction of the second maximum in the proton event of 10–11 May 2024\, which involved the catch-up (cannibali zation) of a slower CME (≈628 km/s\, launched on 8 May at 22:24 UT) by a faster one (≈1711 km/s\, launched on 9 May at 09:24 UT) at a distance o f ≈0.31 AU. Advection of the merged structure to Earth at the OMNI solar -wind speed (684 km/s) predicts the onset of the second proton-flux maximu m at about 09:33 UT on 11 May\, whereas the actual (GOES-cross-validated A RT-XC/Spektr-RG) maximum is observed at 07:33 UT on 11 May\, i.e.\, about 2 hours earlier. RQNN\, in turn\, predicts the second maximum at 06:58 UT on 11 May\, which is about 1 h 25 min (1.4 h) closer to the actual value\, reducing the residual from ≈2.0 h to ≈0.6 h.\n\nhttps://indico.partic le.mephi.ru/event/626/contributions/6209/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6209/ END:VEVENT BEGIN:VEVENT SUMMARY:Large language models in astrophysics and prospects for applying G igaChat to Baikal-GVD and TAIGA DTSTART;VALUE=DATE-TIME:20261029T150500Z DTEND;VALUE=DATE-TIME:20261029T152000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6212@cern.ch DESCRIPTION:Speakers: Ilya Plavunov ()\nWe review applications of large la nguage models in astrophysics and astroparticle physics\, including litera ture retrieval\, scientific programming\, observational data analysis\, an d computational experiments. We compare three complementary approaches: do main adaptation\, retrieval-augmented generation\, and agents that use ext ernal tools. Building on this comparison\, we propose an architecture for a research assistant for Baikal-GVD and TAIGA. Particular attention is giv en to GigaChat as a promising foundation for a Russian-language scientific assistant. Its Russian-language specialization and the availability of op en weights provide opportunities for adaptation to collaboration documenta tion and deployment within collaboration computing infrastructure. We prop ose three pilot use cases: extracting information from astronomical alerts \, reproducing a selected analysis\, and answering questions about technic al documentation. We formulate a protocol for domain-specific evaluation u sing AstroMLab-1 and collaboration tasks. Integration with scientific tool s\, computational reproducibility\, and validation of physical results are discussed as requirements for developing a GigaChat-based assistant.\n\nh ttps://indico.particle.mephi.ru/event/626/contributions/6212/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6212/ END:VEVENT BEGIN:VEVENT SUMMARY:Optimization of detector structure based on flow field numerical s imulation with the assistance of artificial intelligence DTSTART;VALUE=DATE-TIME:20261029T140500Z DTEND;VALUE=DATE-TIME:20261029T142000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6214@cern.ch DESCRIPTION:Speakers: Haoyu Wang ()\nIn order to improve the circulation p urification efficiency of the PandaX-xT dark matter detector and efficient ly achieve the ultra-trace background required by the detector through flo w field optimization\, this study conducts CFD numerical simulations of th e xenon flow field based on the structure of the PandaX-4T detector. The multi-variable simulation results are introduced into an artificial intell igence (AI) framework\, to establish a mapping relationship between the pa rameters and performance metrics for investigating the flow field characte ristics\, and completing the structural optimization of the detector. Fina lly\, the flow field simulation is updated based on the optimized paramete rs. The effectiveness of the optimization scheme in improving flow uniform ity and reducing impurity accumulation are analyzed\, which could provide guidance and optimization directions for the performance enhancement of la rge-scale liquid xenon dark matter detectors.\n\nhttps://indico.particle.m ephi.ru/event/626/contributions/6214/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6214/ END:VEVENT BEGIN:VEVENT SUMMARY:A machine learning method based on the extraction of essential fea tures for multimodal EAS data analysis. DTSTART;VALUE=DATE-TIME:20261029T142000Z DTEND;VALUE=DATE-TIME:20261029T143500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6216@cern.ch DESCRIPTION:Speakers: Александр Крюков (НИИЯФ МГУ)\ nTwo main types of ground-based experimental facilities are used to regist er EAS: atmospheric Cherenkov telescopes and arrays of Cherenkov detectors . The unique TAIGA scientific facility includes both types of installation s. The joint analysis of the data coming from these installations should i mprove the quality of the reconstructed EAS parameters compared to traditi onal methods that analyze the data separately. \n\nFor such a combined mul timodal analysis\, it is proposed to use a method based on extracting esse ntial features from experimental data using autoencoder type neural networ ks that do not require pre-marking the data for training. The main advanta ge of the proposed method over the traditional one is that data from vario us installations are naturally and easy aggregated at the level of essenti al features and further analyzed using classical deep neural networks to e xtract the physical parameters of the EAS.\n\nUsing the example of simulat ed Monte Carlo data for atmospheric Cherenkov telescopes and the Cherenkov detector array HiSCORE of the TAIGA installation\, it is shown that this approach allows obtaining the parameters of the EAS in some cases signific antly better than traditional analysis methods.\n\nhttps://indico.particle .mephi.ru/event/626/contributions/6216/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6216/ END:VEVENT BEGIN:VEVENT SUMMARY:Neural networks for data analysis in Baikal-GVD DTSTART;VALUE=DATE-TIME:20261029T145000Z DTEND;VALUE=DATE-TIME:20261029T150500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6217@cern.ch DESCRIPTION:Speakers: Ivan Kharuk (INR & MIPT)\nWe present an overview of the deep learning pipeline developed for data processing at the Baikal-GVD experiment. The pipeline covers the complete data analysis chain\, from n oise hit suppression to the reconstruction of neutrino arrival directions and energies. By capturing spatial and temporal correlations across optica l modules\, Transformer and Graph Neural Network (GNN) architectures consi stently outperform conventional algorithmic reconstruction across all task s. Furthermore\, we discuss architectural enhancements — such as master nodes\, fully connected output layers\, and loop structures — that boost overall reconstruction performance.\n\nhttps://indico.particle.mephi.ru/e vent/626/contributions/6217/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6217/ END:VEVENT BEGIN:VEVENT SUMMARY:Magnetic field approximation for the SPD detector DTSTART;VALUE=DATE-TIME:20261029T060000Z DTEND;VALUE=DATE-TIME:20261029T061500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6185@cern.ch DESCRIPTION:Speakers: Fomenok Dmitry (JINR)\nAccurate knowledge of the mag netic field distribution is essential for precise momentum reconstruction of charged particles in the SPD experiment at NICA. Traditional interpolat ion of pre-computed field maps becomes computationally expensive for high- resolution grids and may be inaccurate in regions with sparse measurements .\nA systematic comparative study of machine learning methods for approxim ating the three-dimensional magnetic field of the SPD detector is presente d. The study covers a broad range of approaches\, including neural network s and other algorithms. The magnetic field in this work was constructed us ing Monte Carlo simulation.\nThe results are presented and discussed\, wit h recommendations for choosing a field approximation method depending on t he available data and required precision.\n\nhttps://indico.particle.mephi .ru/event/626/contributions/6185/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6185/ END:VEVENT BEGIN:VEVENT SUMMARY:CEPC experiment potential for observation of rare Higgs boson deca y H->ZZ*->4nu at sqrt(s)=240 GeV DTSTART;VALUE=DATE-TIME:20261029T061500Z DTEND;VALUE=DATE-TIME:20261029T063000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6188@cern.ch DESCRIPTION:Speakers: Sergey Korpachev (LPI)\nThe CEPC working point at th e center of mass energy around 240 GeV corresponds to the largest cross se ction of the higgsstrahlung process and allows effective usage of the reco il mass technique for measurements of Higgs boson parameters. The interest ing task is the search for invisible decays of the Higgs boson: H->ZZ*->4n u. The dijet final states of the process e+e-->ZH\, where Z boson decays t o quarks\, have large branching fractions but quite large background contr ibutions. In this work\, a two-stage machine learning algorithm for separa tion of the invisible Higgs decay signal in dijet final states is presente d. At the first stage\, separate binary classifiers are trained on signal- background pairs for several dominating background channels. At the second stage\, a combined classifier is trained using the outputs of the first-s tage models as inputs. The study was done with full simulation and reconst ruction of the signal and background processes by the CEPC software. The p erformance of the two-stage approach is compared with conventional kinemat ic selections\, and its effect is discussed on the upper limit on the bran ching fraction of the Higgs boson invisible decay.\n\nhttps://indico.parti cle.mephi.ru/event/626/contributions/6188/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6188/ END:VEVENT BEGIN:VEVENT SUMMARY:Application of Machine Learning Methods for Event Reconstruction i n the SPD Range System DTSTART;VALUE=DATE-TIME:20261029T063000Z DTEND;VALUE=DATE-TIME:20261029T064500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6196@cern.ch DESCRIPTION:Speakers: Александр Осетров (JINR\, NRNU MEPhI )\nThe Joint Institute for Nuclear Research (JINR) is approaching the comm issioning and operational readiness of the NICA (Nuclotron-based Ion Colli der fAcility) mega-science accelerator complex. One of its future detector s is the SPD (Spin Physics Detector)\, designed to study the spin structur e of nucleons. The Range System (muon system) is an integral part of the d etector\, used for muon identification and coarse hadron calorimetry. This report is devoted to the simulation and reconstruction of events in the S PD Range System using the SpdRoot software framework. In particular\, it p rovides an overview of machine learning algorithms\, such as DBSCAN\, BDT\ , and CNN\, applied to clustering and particle classification tasks. These results provide the foundation for critical measurements\, such as charmo nium probing in the SPD experiment.\n\nhttps://indico.particle.mephi.ru/ev ent/626/contributions/6196/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6196/ END:VEVENT BEGIN:VEVENT SUMMARY:Physics-Guided Evaluation and Regularization of Diffusion Models f or Calorimeter Simulation DTSTART;VALUE=DATE-TIME:20261029T064500Z DTEND;VALUE=DATE-TIME:20261029T070000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6197@cern.ch DESCRIPTION:Speakers: Alexandr Dorosh (Laboratory of methods for Big Data Analysis at HSE)\nFast calorimeter simulation requires accurate generative models and meaningful\nmeasures of physical fidelity. We integrate physic s-supervised distributional evaluation and residual regularization into ou r conditional diffusion pipeline for the first time. Shower generation is conditioned on particle entry coordinates and momentum\, allowing the mode l to represent variations in detector response across incident-particle co nfigurations.\n\nThe evaluation uses a Frechet Physics Distance - FPD - co mputed in a representation learned by a network trained to predict particl e kinematics and shower-shape observables. Comparing the means and covaria nce matrices of reference and generated embeddings connects mathematical d istributional assessment with physically relevant features\, complementing PRD-based scores and KL-divergence diagnostics. Physics supervision guide s feature learning towards quantities relevant to shower characterization\ , while the covariance comparison incorporates correlations between learne d features. The diffusion model predicts clean shower maps in a logarithmi c energy representation\, with a timestep-weighted penalty on negative pre dictions added to its reconstruction objective. This introduces a soft non negativity constraint during training while retaining DDIM sampling withou t a physics solver.\n\nOn the validation sample\, the revised pipeline ach ieves an FPD of $0.058$ after\nclipping negative deposits\, compared with $0.889$ under the same evaluation\nprotocol. Both Energy and Physical PRD AUC scores exceed $0.999$\, with the latter\nevaluated on events with fini te shower-shape features. PRD is computed on raw outputs\, preserving a se parate assessment before clipping. Explicit audits additionally examine ne gative deposits\, sparsity\, and total deposited energy\, identifying disc repancies not fully reflected in aggregate\nscores. These checks help dist inguish agreement in the selected feature spaces from physical admissibili ty of individual generated showers. The resulting framework supports syste matic model comparison and targeted improvements to calorimeter simulation \, with controlled ablations needed to isolate the contribution of residua l regularization.\n\nhttps://indico.particle.mephi.ru/event/626/contributi ons/6197/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6197/ END:VEVENT BEGIN:VEVENT SUMMARY:Graph Attention Neural Network for Clustering Particle Tracks by E vents in the SPD experiment DTSTART;VALUE=DATE-TIME:20261029T070000Z DTEND;VALUE=DATE-TIME:20261029T071500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6199@cern.ch DESCRIPTION:Speakers: Savelii Omelianchuk ()\nA Graph Attention Neural Net work (GANN) for particle track clustering by events in time slices obtaine d at the Spin Physics Detector (SPD) of the Nuclotron-based Ion Collider f Acility (NICA) is presented. A novel approach is applied to time slice pro cessing at high particle multiplicity and pile-up. The model is trained us ing a hierarchical approach and evaluated on time slices generated by the Monte Carlo technique according to the SPD experiment configuration as clo se as possible at the moment. The model architecture includes a graph enco der and an edge classifier\, both of which use the message-passing princip le and iteratively structured approach. The GANN exhibits high performance in track clustering by events for the SPD simulation\, achieving 90% accu racy and 90% precision and recall.\n\nhttps://indico.particle.mephi.ru/eve nt/626/contributions/6199/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6199/ END:VEVENT BEGIN:VEVENT SUMMARY:Differentiable reconstruction of the AdS/QCD Warp Factor DTSTART;VALUE=DATE-TIME:20261029T071500Z DTEND;VALUE=DATE-TIME:20261029T073000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6200@cern.ch DESCRIPTION:Speakers: Anna Komleva (MIPT (NRU))\nHolographic QCD models de scribe strongly interacting matter through a five-dimensional gravitationa l background. In Einstein–Maxwell–Dilaton models\, the warp factor A(z ) deforms the AdS metric and affects QCD thermodynamics\, the phase struct ure\, and heavy-meson spectra. Such models are commonly calibrated using l attice-QCD data [1\, 2] in contrast with employment on the experimental da ta.\n\nIn this work\, we develop a differentiable-programming framework fo r the inverse reconstruction of the warp factor in AdS/QCD using PyTorch [ 3]. We consider parameterized families of phenomenological ansatz\, in par ticular combinations of logarithms of polynomials. Their parameters are op timized with gradient-based methods\, in particular Adam [4] and L-BFGS-B [5]\, while automatic differentiation provides gradients through the numer ical pipeline without finite-difference approximations.\n\nFor each parame ter set\, we construct the holographic background and calculate the phase- transition curve in the temperature–baryon-chemical-potential coordinate s. Its deviation from the chemical freeze-out line defines the thermodynam ic loss\; the method is tested using both an analytic freeze-out parameter ization and experimental data from the relativistic nucleus-nucleus collis ions [6]. Heavy-meson masses are calculated following [7] using the aforem entioned ansatz'. The constraints on the meson masses are included through gaussian process emulator differentiable surrogate [8] that approximate t he dependence of meson masses on the background parameters. The total loss combines information on the thermodynamics and meson spectroscopy along w ith regularization on the critical endpoint (CEP) position to achieve smoo th and physically admissible warp factors. The procedure reconstructs fami lies of warp factors compatible with chemical-freeze-out and spectroscopic constraints\, allowing us to assess the stability and non-uniqueness of t he inverse problem.\n\n## References\n\n[1] X. Chen and M. Huang\, “Flav or dependent critical endpoint from holographic QCD through machine learni ng\,” arXiv:2405.06179 [hep-ph] (2024).\n\n[2] M. T. Arun and R. Pal\, “Learning holographic QCD with unflavored meson spectra\,” J. High Ene rgy Phys. 06 (2026) 226\n\n[3] A. Paszke et al.\, “Automatic differentia tion in PyTorch\,” J. Mach. Learn. Res. 18 (2018) 1\, arXiv:1701.06757 [ cs.LG].\n\n[4] D. P. Kingma and J. Ba\, “Adam: A Method for Stochastic O ptimization\,” in 3rd International Conference on Learning Representatio ns (ICLR 2015)\, San Diego\, CA\, USA\, May 7-9\, 2015\, Conference Track Proceedings\, 2015.\n\n[5] D. C. Liu and J. Nocedal\, “On the Limited Me mory BFGS Method for Large Scale Optimization\,” in Mathematical Program ming\, vol. 45\, pp. 503–528\, 1989.\n\n[6] X. Luo and N. Xu\, “Search for the QCD critical point with fluctuations of conserved quantities in r elativistic heavy-ion collisions at RHIC: An overview\,” Nucl. Sci. Tech . 28 (2017) 112\n\n[7] B. Toniato\, D. Dudal\, S. Mahapatra\, R. da Rocha\ , and S. S. Jena\, “Holographic QCD model for heavy and exotic mesons at finite density: A self-consistent dynamical approach\,” Phys. Rev. D 11 1\, 126021 (2025)\n\n[8] M. C. Kennedy and A. O'Hagan\, “Bayesian calibr ation of computer models\,” J. R. Stat. Soc. Ser. B Stat. Methodol. 63 ( 2001)\, no. 3\, 425–464.\n\nhttps://indico.particle.mephi.ru/event/626/c ontributions/6200/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6200/ END:VEVENT BEGIN:VEVENT SUMMARY:Event-Adaptive FARICH Reconstruction under Unknown Dark-Count Rate s with a Track-Conditioned Transformer DTSTART;VALUE=DATE-TIME:20261029T073000Z DTEND;VALUE=DATE-TIME:20261029T074500Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6201@cern.ch DESCRIPTION:Speakers: Platon Rogozhin (HSE University)\nFocusing aerogel r ing-imaging Cherenkov (FARICH) detectors require the reconstruction of spa rse photon patterns in the presence of background caused by a photosensor dark count rate. We present the Track-Conditioned Ring Assignment Transf ormer (TCRAT)\, an attention-based model that simultaneously assigns photo n hits\, reconstructs the particle velocity\, and performs particle identi fication in a FARICH detector.\n\nEach event is represented as an unordere d set of pixelized hits. A track-conditioned query identifies hits compati ble with the charged-particle trajectory\, while a separate background que ry models unassigned hits. The method does not impose a circular or ellipt ical ring shape on the photodetector plane\, allowing it to account for an gle-dependent distortions caused by refraction at the aerogel-air boundary .\n\nThe method was evaluated using simulated single-particle events in th e FARICH detector concept for the future Super Charm-Tau Factory\, includi ng track and momentum measurement uncertainties and dark-count rates rangi ng from zero to 1 MHz/mm². When trained across the full background range without an explicit dark-count-rate input\, TCRAT maintained an RMS90 β r esolution of about $(4.6\\times10^{-4})$ and a balanced particle-identific ation accuracy of about 94% across the studied noise conditions. Providing the dark-count rate explicitly did not improve the reconstruction quality \, indicating that the model can infer the relevant background conditions directly from the event hit pattern.\n\nhttps://indico.particle.mephi.ru/e vent/626/contributions/6201/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6201/ END:VEVENT BEGIN:VEVENT SUMMARY:Deep Learning for Fast Simulation of the Microstrip GEM Detector i n the BM@N Experiment DTSTART;VALUE=DATE-TIME:20261029T074500Z DTEND;VALUE=DATE-TIME:20261029T080000Z DTSTAMP;VALUE=DATE-TIME:20261022T104618Z UID:indico-contribution-314-6202@cern.ch DESCRIPTION:Speakers: Dmitry Baranov (JINR)\nThe talk presents the first e xperience of applying deep neural networks to simulate the responses of mi crostrip GEM detectors used in the BM@N experiment at the NICA accelerator complex (JINR). The preparation of the training dataset based on data obt ained using classical Monte-Carlo simulation methods implemented in the BM NROOT software is discussed. The advantages of the neural network approach are considered\, particularly the reduction in detector response generati on time compared to traditional methods. The quality of the generated resp onses is also compared with reference simulations.\n\nhttps://indico.parti cle.mephi.ru/event/626/contributions/6202/ LOCATION:Sevastopol Hotel URL:https://indico.particle.mephi.ru/event/626/contributions/6202/ END:VEVENT END:VCALENDAR