Differences

This shows you the differences between two versions of the page.

--- dlcp2026:program [02/07/2026 15:41] – [45. Вычисление моментов как способ параметризации 2D-карт фотолюминесценции при решении обратной задачи спектроскопии] admin
+++ dlcp2026:program [05/07/2026 18:51] (current) – [7. Shower Core Reconstruction in the HiSCORE Experiment using Neural Networks fed by Autoencoder-Derived Essential Features] admin
@@ Line 1: / Line 1: @@
 {{ :dlcp2026:dlcp26-logo.png?200|}}
-====== Program (Draft) ======
+====== Program (final) ======
-//23.06.2025//
+//05.06.2025//
-**The final list of accepted reports will be publish later.**
 <color /orange>The first author is the presenter.</color>
-//If someone did not find themselves in the list, please inform us by email [[dlcp@sinp.msu.ru]]//
+{{ :dlcp2026:dlcp2026_program.pdf |DLCP2026 program}} (final).
-{{ :dlcp2026:dlcp2026_program.pdf |DLCP2026 program}} (draft).
 ===== Section 1. Machine Learning in Fundamental Physics =====
@@ Line 56: / Line 52: @@
 We develop a Lorentz-equivariant Geometric Algebra Transformer (LGaTr) as the encoder of a foundational model for top-quark physics. Each hard event is encoded as a multivector in the spacetime algebra Cl(1,3): tokens carry their raw 4-momentum in the grade-1 channel and invariant flavour/charge tags in grade-0, while the geometric product makes higher grades — decay-plane bivectors, 3-body trivectors, and the CP-odd pseudoscalar — emerge automatically. Since every grade transforms under the same rotor sandwich, the encoder is Lorentz-equivariant by construction, so invariant masses and angles are exact rather than learned. By the Cayley–Menger collapse GA generates no new invariant scalars beyond the 1-bit CP-odd pseudoscalar sign; its value is structural — enforcing equivariance and reducing multi-step covariant constructions such as top-spin correlations to a single operation. On a `pp → tWb` proof-of-concept the GATr-lite encoder reaches higher AUC compared to a hand-crafted-feature baseline with ~3 times fewer parameters.
+==== 66. Высокопроизводительное хранилище по запросу для поддержки инференса – преимущество подхода дизаггрегированной компонуемой инфраструктуры. ====
+//**А.А. Московский** \\ группа компаний РСК.//
+//Приглашенный доклад//
 ==== 27. Application of machine learning for the analysis of four-jet final states in the CEPC experiment ====
 //**A.Staritsyna**(1), M.Chadeeva(2) \\ (1)Moscow Institute of Physics and Technology, (2) Lebedev Physical Institute of the Russian Academy of Sciences //
@@ Line 102: / Line 103: @@
 and plans. Physics of Atomic Nuclei. — 2021 — Vol. 84, no. 6 — P. 966–974.
-/**
+==== 7. Neural Network-Based Shower Core Reconstruction in the HiSCORE Experiment using Autoencoder-Derived Essential Features ====
-==== 1. To be annoncemented ====
+//**Ю.Ю. Дубенская**(1), А.П. Крюков(1), П.А. Волчугов(1), Е.О. Гресь(1,2), А.П. Демичев(1), Д.П. Журов(1,2), С.П. Поляков(1), Е.Б. Постников(1), А.Ю. Разумов(1) \\ (1) НИИЯФ МГУ, Москва, (2) НИИПФ ИГУ, Иркутск //
+The goal of this work is to investigate how to prepare and process HiSCORE data so as to minimize losses, while also bringing it into a form suitable for combining with data from other detector types of the TAIGA project. Our main idea is to use a neural network that reconstructs the core position using essential features (i.e., a compact representation of the data) as input. The essential features are not related to the dimensions and format of the data, so the essential features for different types of detectors can be easily combined. As essential features we propose to use a latent space of an autoencoder (AE) trained on HiSCORE data. We demonstrate that the latent space of a properly trained autoencoder can be used to reconstruct the core position of the EAS, and that adding additional pre-computed angle parameters to the input further improves the reconstruction accuracy. The standard deviation of the core position reconstruction error for the proposed method is 15-17 meters, which is comparable with the results obtained by other methods.
+==== 1. Invertible Neural Networks and the Possibility of their Applications to Cosmic Rays Data Analysis ====
 //**А.П. Крюков**(1), П.А. Волчугов(1), Е.О. Гресь(1,2), А.П. Демичев(1), Ю.Ю. Дубенская(1), Д.П. Журов(1,2), С.П. Поляков(1), Е.Б. Постников(1), А.Ю. Разумов(1) \\ (1) НИИЯФ МГУ, Москва, (2) НИИПФ ИГУ, Иркутск //
-**/
 ----
@@ Line 201: / Line 206: @@
 Инженерное проектирование представляет собой задачу оптимизации с ограничениями черного ящика - каждая возможная конфигурация должна удовлетворять жёстким физическим ограничениям, а испытание конфигураций обходится дорого. Мы сравниваем взаимодействие машинного обучения и механического проектирования на основе четырех классических задач оптимизации с ограничениями: проектированию редуктора, пружины растяжения-сжатия, сосуда под давлением, и сварной балки. В рамках единого подхода, основанном на Байесовской оптимизации, мы сравниваем несколько стратегий выбора кандидатов, и предлагаем схему на основе динамики Ланжевена. Оценивая все четыре задачи по общим показателям — сходимости наилучшей допустимой стоимости, доле допустимых решений и времени до первого допустимого решения, — мы показываем, как алгоритмы семплирования позволяют получать лучшие решения по сравнению со стандартными подходами поиска оптимума.
+==== 45. Вычисление моментов как способ параметризации 2D-карт фотолюминесценции при решении обратной задачи спектроскопии ====
+//**Г.А.Куприянов**(1, 2), И.В.Исаев(2), К.А.Лаптинский(1, 2), С.А.Доленко(1, 2), Т.А.Доленко(1) \\ (1) Физический факультет Московского государственного университета имени М.В. Ломоносова, (2) Научно-исследовательский институт ядерной физики им. Д.В.Скобельцына//
+Фотолюминесцентные наносенсоры на основе углеродных точек являются перспективным инструментом для экспрессного контроля ионного состава водных сред. При решении многопараметрической обратной задачи спектроскопии важную роль играет представление 2D-карт фотолюминесценции (ФЛ), поскольку исходная матрица возбуждения–испускания содержит большое число взаимно коррелированных признаков. В связи с этим актуальна разработка способов параметризации, позволяющих уменьшить размерность входного пространства без потери информации, существенной для определения концентраций ионов. В работе исследовано представление 2D-карт фотолюминесценции через разложение по базисам Тейлора, Лежандра, Фурье и Цернике. Такой подход можно рассматривать как вычисление набора моментов карты ФЛ, причем число признаков регулируется максимальным порядком разложения. Использование моментов Цернике ранее показало эффективность при анализе 2D-карт ФЛ природных вод и биологических объектов [1, 2]; в настоящей работе проведено сравнение нескольких функциональных базисов применительно к обратной задаче определения концентраций ионов металлов. Исследование выполнено на 7-параметрической базе данных, содержащей 7813 2D-карт ФЛ водных растворов углеродных точек с катионами Ni2+, Cu2+, Co2+, Pb2+, Al3+, Cr3+ и анионом NO3-. Полученные коэффициенты разложения использовались как входные признаки для моделей машинного обучения: линейной регрессии, случайного леса, градиентного бустинга, многослойного персептрона; в качестве референсного подхода рассматривалась сверточная нейронная сеть, обучаемая непосредственно на 2D-картах ФЛ. Показано, что параметризация 2D-карт ФЛ на основе функциональных базисов позволяет достигать точности на уровне сверточной нейронной сети при использовании существенно меньшего числа входных признаков: достаточно 78 коэффициентов разложения и простой модели линейной регрессии. Наиболее эффективными оказались базисы Лежандра и Фурье, что связано с их ортогональностью и определением на всей области 2D-карты ФЛ. Таким образом, разложение 2D-карт фотолюминесценции по функциональным базисам является эффективным способом сжатия и структурирования спектральной информации при решении обратных задач фотолюминесцентной спектроскопии. Предложенный подход особенно полезен для задач с ограниченным объемом экспериментальных данных, поскольку позволяет уменьшить размерность входного пространства, сохранить физически значимые особенности формы 2D-карт ФЛ и повысить устойчивость моделей машинного обучения. Исследование выполнено за счет гранта Российского научного фонда № 22-12-00138-П, https://rscf.ru/project/22-12-00138/. Г.А.Куприянов является стипендиатом Фонда развития теоретической физики и математики «Базис». [1] Wang X., Li B.Q., Zhai H.L. et al. An efficient approach to the quantitative analysis of humic acid in natural water using Zernike moments // Food Chemistry. 2016. V. 190. P. 1033–1039. https://doi.org/10.1016/j.foodchem.2015.06.074 [2] Chen Y., Chen T., Duan W. et al. Rapid measurement of brown tide algae using Zernike moments and ensemble learning based on excitation-emission matrix fluorescence // Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy. 2023. V. 294. Art. 122547. https://doi.org/10.1016/j.saa.2023.122547
 ==== 4. Топовые вычислительные системы для суперкомпьютинга и ИИ ====
@@ Line 273: / Line 284: @@
 ==== 44. Texture Based Explainable Machine Learning for Automated Glacial Crevasse Detection in GPR Radargrams ====
-//**Ripul Ghosh**(1,2), Ashish Lawaniya(1), Anuj Kumar Mishra(1,2), Sergey Popov(3,4), Marina Kashkevich(3), Margarita Stepanova(5), Siddhartha Sarkar(1,2), Satish Kumar(1,2) \\ (1) Centre of Excellence for Intelligent Sensors and Systems (iSenS), India, CSIR-Central Scientific Instruments Organisation, India, (2) Academy of Scientific and Innovative Research (AcSIR), India, (3) Institute of Earth Sciences, Saint Petersburg State University, Russia, (4) Gramberg All-Russian Scientific Research Institute for Geology and Mineral Resources of the World Ocean (VNIIOkeangeologia), Russia, (5) The Faculty of Physics, Saint Petersburg State University, Russia//
+//**Ripul Ghosh**(1,2), Ashish Lawaniya(1), Anuj Kumar Mishra(1,2), Sergey Popov(3,4), Marina Kashkevich(3_, Margarita Stepanova(5), Satish Kumar(1,2) \\
+(1)Centre of Excellence for Intelligent Sensors and Systems (iSenS), CSIR-Central Scientific Instruments Organisation, India \\
-Glacial crevasses across Antarctica’s ice sheets pose risks to human activities and transportation in polar regions. Reliable crevasse detection is essential for safe scientific expeditions and surveys. Though ground penetrating radar (GPR) is widely used for crevasse detection, manual interpretation of radargrams remains time-intensive and dependent on operator expertise. Recently, advances in computer vision and artificial intelligence have enabled automated detection of these crevasses. Although deep learning approaches have shown promising performance, lightweight machine learning frameworks have received relatively less attention for automated crevasse detection. This work integrates an explainable machine learning framework for the automated crevasse detection in GPR images using the texture-based feature extraction techniques. A dataset consisting of 474 real and synthetic radargram image frames representing hyperbolic subsurface reflections and background radar signatures was prepared. For the improvement of the spatial localization and increase sample diversity, the frames were obtained through a sliding window 512 × 512 pixel with an overlap of 20% in both dimensions. Three handcrafted texture feature extraction approaches namely Gabor filtering, histogram of oriented gradients (HOG) and local binary pattern (LBP) were explored to characterise the radargram images. Gabor filtering was used to capture multi-scale texture orientation with the frequency responses that closely align and represent the wave propagation characteristics in the radar images. HOG descriptors were used to represent local edge orientation patterns associated with hyperbolic boundaries. LBP descriptors were extracted for the characterisation of local texture micro-patterns and intensity transitions which are present in radar reflections. Further, these extracted feature representations were evaluated using the four different lightweight machine learning classifiers such as shallow artificial neural network (ANN), support vector machine (SVM), random forest (RF) and extreme gradient boosting (XGBoost). To improve interpretability and understand feature contributions, SHapley Additive exPlanations (SHAP) was integrated as an explainable artificial intelligence (XAI) framework. SHAP analysis enabled the identification of the different dominant texture features which influence the classification decisions. This interpretability component helps in the improvement of the model transparency. Thus, supports trustworthiness in safety-critical applications where automated decisions directly influence operational planning. Model performance was carried out under the three fold cross-validation to ensure robustness. Evaluation metrics contain accuracy, precision, recall, F1-score and AUC curve to comprehensively measure classification performance. Experimental results indicate that the Gabor based texture descriptors consistently outperformed both the HOG as well as LBP feature representations, indicating that frequency and orientation-sensitive texture analysis is usually well suited to characterise the GPR hyperbolic reflections. Among all the other evaluated machine learning models, XGBoost achieved the best overall classification performance, attaining an decrease in the false negative rate (FNR) with 42.3 %, increase in the recall and F1-Score by 2.8 % and 1.2 % respectively when compared to the RF classifier. The findings of this study demonstrate that handcrafted texture descriptors combined with lightweight machine learning classifiers provide a computationally efficient and highly interpretable alternative to computationally expensive deep learning frameworks for automated crevasse detection in GPR radargrams.
+(2)Academy of Scientific and Innovative Research (AcSIR), India \\
+(3)Institute of Earth Sciences, Saint Petersburg State University, Russia \\
+(4)Gramberg All-Russian Scientific Research Institute for Geology and Mineral Resources of the World Ocean (VNIIOkeangeologia), Russia \\
+(5)The Faculty of Physics, Saint Petersburg State University, Russia//
+Glacial crevasses across Antarctica’s ice sheets pose risks to human activities and transportation in polar regions. Reliable crevasse detection is essential for safe scientific expeditions and surveys. Though ground penetrating radar (GPR) is widely used for crevasse detection, manual interpretation of radargrams remains time-intensive and dependent on operator expertise. Recently, advances in computer vision and artificial intelligence have enabled automated detection of these crevasses. Although deep learning approaches have shown promising performance, lightweight machine learning frameworks have received relatively less attention for automated crevasse detection. This work integrates an explainable machine learning framework for the automated crevasse detection in GPR images using the texture-based feature extraction techniques. A dataset consisting of 503 real and synthetic radargram image frames representing hyperbolic subsurface reflections and background radar signatures was prepared. For the improvement of the spatial localization and increase sample diversity, the frames were obtained through a sliding window 512 × 512 pixel with an overlap of 50% in both dimensions.  Three handcrafted texture feature extraction approaches namely Gabor filtering, histogram of oriented gradients (HOG) and local binary pattern (LBP)  were explored to characterise the radargram images. Gabor filtering was used to capture multi-scale texture orientation with the frequency responses that closely align and represent the wave propagation characteristics in the radar images. HOG descriptors were used to represent local edge orientation patterns associated with hyperbolic boundaries. LBP descriptors were extracted for the characterisation of local texture micro-patterns and intensity transitions which are present in radar reflections. Further, these extracted feature representations were evaluated using the four different lightweight machine learning classifiers such as shallow artificial neural network (ANN), support vector machine (SVM), random forest (RF) and extreme gradient boosting (XGBoost).
+To improve interpretability and understand feature contributions, SHapley Additive exPlanations (SHAP) was integrated as an explainable artificial intelligence (XAI) framework. SHAP analysis enabled the identification of the different dominant texture features which influence the classification decisions. This interpretability component helps in the improvement of the model transparency. Thus, supports trustworthiness in safety-critical applications where automated decisions directly influence operational planning.
 ==== 48. Параметризация субмезомасштабного вертикального потока плавучести с помощью свёрточной нейронной сети на основе крупномасштабных характеристик перемешанного слоя ====