All issues

The paper presents a cascade model of a physically supported neural network designed to predict pressure drop in three-phase flow (oil, gas, water) in a pipe section with various angles of inclination. To overcome the constraints of existing empirical correlations and computation-intensive numerical modeling methods, we propose an architecture that decomposes the problem into three sequential physically interpretable subtasks: regression prediction of the fluid hold-up coefficient, fluid flow regime classification, and pressure gradient evaluation. Each subtask is solved by a separate fully connected neural network, the output of which is passed to the next model in the cascade. Training and testing of the proposed architecture was performed on an extensive synthetic dataset (8 · 10⁷ records) generated using a semi-empirical model. Verification is performed on independent experimental data. A comparative analysis with a single fully connected (non-cascade) neural network is made, and the sensitivity of the models is examined using Sobol and Borgonovo methods. The cascade model demonstrates superior accuracy and ensures high interpretability of results by providing intermediate physical parameters (fluid hold-up coefficient, flow regime). The developed model has low computational complexity, which allows it to be used in real-time systems and digital twins of hydraulic systems in the oil and gas industry.

The SAS program is the primary data processing tool for the YuMO small-angle neutron scattering spectrometer. The paper presents a retrospective analysis of its two-decade evolution, from a Fortran prototype to a modern software system. The analysis focuses on the architectural decisions that have ensured the program’s long-term viability and its ability to adapt to instrument upgrades.

The core solution was a modular architecture that abstracts the detector system. This enabled the seamless integration of data from two scattering detectors and, later, from a position-sensitive detector. A strict processing pipeline and a unified internal data representation formed the basis for physically grounded algorithms, including weighted merging of spectra, resolution-aware smoothing, and built-in statistical quality control. The program’s interfaces—a command line for batch processing and a graphical user interface for interactive work—are built upon a single computational core, ensuring result consistency and flexibility in use.

Long-term operation has confirmed that the underlying architectural principles naturally align with the key characteristics of international software quality standards, particularly those critical for long-term sustainability. Therefore, the development and evolution of SAS demonstrates a universal set of architectural principles that can serve as a foundation for building sustainable scientific software in related fields of experimental physics.

This paper addresses the problem of detecting small objects in visible-spectrum aerial imagery. High background variability and weak feature saliency make small object detection a non-trivial task. Under such conditions, classical computer vision algorithms based on hand-crafted descriptors exhibit low efficiency, prompting a shift towards deep neural network architectures, which demonstrate superior generalization capability and robustness to false positives. We selected a one-stage approach based on a neural network predictive model as the primary object detection method. Also, several neural network architectures belonging to this class were analyzed, outlining their advantages and disadvantages. As the baseline detector, we adopted YOLO version 11 and incorporated a multi-scale feature aggregation module (which combines information from neural network layers operating at different scales) and a dimension-aware selective integration module (which automatically determines the dimension (channel, height, or width) along which to process features and fuses them selectively). These modifications aim to both enhance computational efficiency, enabling deployment of neural network models onboard aerial vehicles for real-time image analysis, and improve small object detection accuracy. Given the complexity of image annotation, we used an open synthetic image database containing approximately 4 000 images for training and testing (with a 90/10 split, respectively). We extend the training set using various random augmentation techniques. To evaluate the performance of the resulting predictive models, we employed mean Average Precision across all classes, using both a fixed 50% intersectionover- union threshold and a varying threshold from 50% to 95%. Overfitting was monitored by analyzing loss curves during training process. The proposed modifications to the YOLO architecture reduced image processing time by a factor of two while maintaining detection accuracy.

The short-term prediction of road traffic condition is one of the main tasks of transportation modelling. The main purpose of which are traffic control, reporting of accidents, avoiding traffic jams due to knowledge of traffic flow and subsequent transportation planning. A number of solutions exist — both model-driven and data driven had proven to be successful in capturing the dynamics of traffic flow. Nevertheless, most space-time models suffer from high mathematical complexity and low efficiency. Artificial Neural Networks, one of the prominent datadriven approaches, show promising performance in modelling the complexity of traffic flow. We present a neural network architecture for traffic flow prediction on a real-world road network graph. The model is based on the combination of a recurrent neural network and graph convolutional neural network. Where a recurrent neural network is used to model temporal dependencies, and a convolutional neural network is responsible for extracting spatial features from traffic. To make multiple few steps ahead predictions, the encoder-decoder architecture is used, which allows to reduce noise propagation due to inexact predictions. To model the complexity of traffic flow, we employ multilayered architecture. Deeper neural networks are more difficult to train. To speed up the training process, we use skip-connections between each layer, so that each layer teaches only the residual function with respect to the previous layer outputs. The resulting neural network was trained on raw data from traffic flow detectors from the US highway system with a resolution of 5 minutes. 3 metrics: mean absolute error, mean relative error, mean-square error were used to estimate the quality of the prediction. It was found that for all metrics the proposed model achieved lower prediction error than previously published models, such as Vector Auto Regression, LSTM and Graph Convolution GRU.

The reconstruction of charged particle trajectories in tracking detectors is a key problem in the analysis of experimental data for high energy and nuclear physics.

The amount of data in modern experiments is so large that classical tracking methods such as Kalman filter can not process them fast enough. To solve this problem, we have developed two neural network algorithms of track recognition, based on deep learning architectures, for local (track by track) and global (all tracks in an event) tracking in the GEM tracker of the BM@N experiment at JINR (Dubna). The advantage of deep neural networks is the ability to detect hidden nonlinear dependencies in data and the capability of parallel execution of underlying linear algebra operations.

In this work we generalize these algorithms to the cylindrical GEM inner tracker of BESIII experiment. The neural network model RDGraphNet for global track finding, based on the reverse directed graph, has been successfully adapted. After training on Monte Carlo data, testing showed encouraging results: recall of 98% and precision of 86% for track finding.

The local neural network model TrackNETv2 was also adapted to BESIII CGEM successfully. Since the tracker has only three detecting layers, an additional neuro-classifier to filter out false tracks have been introduced. Preliminary tests demonstrated the recall value at the first stage of 99%. After applying the neuro-classifier, the precision was 77% with a slight decrease of the recall to 94%. This result can be improved after the further model optimization.

We suggested some approaches for solving image processing tasks in the decision support system (DSS) of the Center for Automated Recording of Administrative Offenses of the Road Traffic (CARAO). The main task of this system is to assist the operator in obtaining accurate information about the vehicle registration plate and the vehicle brand/model based on images obtained from the photo and video recording systems. We suggested the approach for vehicle registration plate recognition and brand/model classification on the images based on modern neural network models. LPRNet neural network model supplemented by Spatial Transformer Layer was used to recognize the vehicle registration plate. The ResNeXt-101-32x8d neural network model was used to classify for vehicle brand/model. We suggested the approach to construct the training set for the neural network of vehicle registration plate recognition. The approach is based on computer vision methods and machine learning algorithms. The SIFT algorithm was used to detect and describe local features on images with the vehicle registration plate. DBSCAN clustering was used to detect and delete outliers in such local features. The accuracy of vehicle registration plate recognition was 96% on the testing set. We suggested the approach to improve the efficiency of using the ResNeXt-101-32x8d model at additional training and classification stages. The approach is based on the new architecture of convolutional neural networks with “freezing” weight coefficients of convolutional layers, an additional convolutional layer for parallelizing the classification process, and a set of binary classifiers at the output. This approach significantly reduced the time of additional training of neural network when new vehicle brand/model classification was needed. The final accuracy of vehicle brand/model classification was 99% on the testing set. The proposed approaches were tested and implemented in the DSS of the CARAO of the Republic of Tatarstan.

The article presents a technology for building clinical decision support systems (CDSS) based on service containers using Docker and a web interface that runs directly in the browser without installing specialized software on workstation of a clinician. A modular architecture is proposed in which each application module is packaged as an independent service container combining a lightweight web server, a user interface, and computational components for medical image processing. Communication between the browser and the server side is implemented via a persistent bidirectional WebSocket connection with binary message serialization (MessagePack), which provides low latency and efficient transfer of large data. For local storage of images and analysis of results, browser facilities (IndexedDB with the Dexie.js wrapper) are used to speed up repeated data access. Three-dimensional visualization and basic operations with DICOM data are implemented with Three.js and AMI.js: this toolchain supports the integration of interactive elements arising from the task context (annotations, landmarks, markers, 3D models) into volumetric medical images.

Server components and functional modules are assembled as a set of interacting containers managed by Docker. The paper discusses the choice of base images, approaches to minimizing containers down to runtime-only executables without external utilities, and the organization of multi-stage builds with a dedicated build container. It describes a hub service that launches application containers on user request, performs request proxying, manages sessions, and switches a container from shared to exclusive mode at the start of computations. Examples of application modules are provided (fractional flow reserve estimation, quantitative flow ratio computation, aortic valve closure modeling), along with the integration of a React-based interface with a three-dimensional scene, a versioning policy, automated reproducibility checks, and the deployment procedure on the target platform.

It is demonstrated that containerization ensures portability and reproducibility of the software environment, dependency isolation and scalability, while the browser-based interface provides accessibility, reduced infrastructure requirements, and interactive real-time visualization of medical data. Technical limitations are noted (dependence on versions of visualization libraries and data formats) together with practical mitigation measures.

Aberrant alternative splicing of the CD44 gene drives colorectal cancer progression and facilitates the emergence of cancer stem cells. Although biomedical research recognizes this transmembrane glycoprotein as a major catalyst of malignancy, deciphering its multi-isoform regulatory networks remains a complex analytical challenge. To address this knowledge gap, this study presents a machine learning framework designed to decode these biological mechanisms. The author constructed a neural network regressor based on beta regression to model bounded isoform proportions. This computational architecture jointly estimates both the mean and the precision parameters of the underlying probability distribution. Furthermore, the system employs elastic net regularization to perform quantitative feature selection from highdimensional molecular expression data.

The investigation evaluates the proposed framework using gene expression profiles from colorectal cancer patients. The primary objective involves identifying specific ribonucleic acid-binding proteins acting as regulatory splicing factors. The experimental design contrasts two distinct mathematical modeling strategies. The first configuration incorporates an independent ”one-vs-all” approach that treats each transcript variant as an isolated regression target. The second formulation utilizes a structured ”isoform tree” method that directly mirrors hierarchical exon inclusion relationships. Validation experiments on synthetically generated datasets confirmed the mathematical integrity of the network. The model recovered true distribution parameters with precision and exhibited no systematic bias. Comprehensive empirical comparisons subsequently demonstrated that the independent ”one-vs-all” layout consistently outperforms the hierarchical tree configuration in predictive stability and accuracy.

The computational analysis maps the regulatory landscape of the CD44 gene. The framework validates several established splicing factors while uncovering new candidate proteins, including ACO1, NUDT21, and AGO2. Based on these statistical associations, the paper introduces a biological hypothesis. This concept functionally connects intracellular iron metabolism via the ACO1 protein with the shifting balance of CD44 variants. These discoveries provide deeper insights into oncogenic splicing regulation. Ultimately, they highlight molecular targets for future therapeutic interventions aimed at suppressing the cancer stem cell phenotype.

Reliable plant disease diagnosis requires not only high classification accuracy, but also stable generalization and interpretable decision-making. Although attention mechanisms are useful to these deep learning models, the performance also depends on where and how they are integrated into the network architecture. This study presents a design-optimized YOLO11m-based classification framework that systematically investigates the impact of Convolutional Block Attention Module (CBAM) injection at different architectural levels for plant disease diagnosis. We perform a comparative modelcontrolled analysis of three model architectures: (i) the baseline YOLO11m-Cls architecture lacking attention, (ii) only adding the backbone block along with CBAM and (iii) a hybrid architecture that includes reduced backbone attention combined with CBAM added at classification head. All models are trained and tested under the same experimental settings using a largescale dataset with around 90 000 images of 38 types of plant diseases. Experimental results clearly show that the uniform injection of CBAM into the backbone reduces stability but causes generalization to worsen with a higher validation loss and significantly lower Top-1 accuracy (≈ 90.5%), while the hybrid attention design balances stability and discrimination, with Top-1 accuracy up to 99.71%, Top-5 accuracy up to 99.99% and near-baseline validation behavior respectively Grad-CAMbased interpretability analysis also demonstrates that the hybrid model generates enhanced and biologically interpretable activation maps, which are more disease-specific with less distraction from background. Notably, the aim of our work is not for achieving superior performance over all classifiers but instead only to provide design-level evidence on how placing attention modulates rigidity and interpretability in YOLO-based classification models. The results provide practical architectural considerations for building dependable and interpretable AI systems in agriculture.

The numerical methods used to simulate the evolution of hydrodynamic systems require the considerable use of computational resources thus limiting the number of possible simulations. The data-driven simulation technique is one promising approach to the development of heuristic models, which may speed up the study of such models. In this approach, machine learning methods are used to tune the weights of an artificial neural network that predicts the state of a physical system at a given point in time based on initial conditions. This article describes an original neural network architecture and a novel multi-stage training procedure which create a heuristic model of a two-phase flow in a heterogeneous porous medium. The neural network-based model predicts the states of the grid cells at an arbitrary timestep (within the known constraints), taking in only the initial conditions: the properties of the heterogeneous permeability of the medium and the location of sources and sinks. The proposed model requires orders of magnitude less processor time in comparison with the classical numerical method, which served as a criterion for evaluating the effectiveness of the trained model. The proposed architecture includes a number of subnets trained in various combinations on several datasets. The techniques of adversarial training and weight transfer are utilized.