https://miscj.aut.ac.ir/ AUT Journal of Modeling and Simulation en daily 1 Thu, 01 Oct 2026 00:00:00 +0330 Thu, 01 Oct 2026 00:00:00 +0330 https://miscj.aut.ac.ir/article_5918.html AbstractBiomedical Natural Language Inference (BioNLI) is a core task in biomedical NLP that looks to identify whether the biomedical premise entails the hypothesis from the premise. Prompt-based methods are gaining traction as one of the simplest and fastest approaches for effectively using large language models (LLMs) for these sorts of inference tasks without the need for complex, time-consuming fine-tuning of the model. However, the inherent difficulty of biomedical NLI problems represents a significant challenge for prompt engineering given the heavy use of terminology specific to the biomedical domain. During in-context prompting or using predefined examples for zero-shot or few-shot prompting, there is often a lack of contextual information or generalizability to the heterogeneity found in biomedical texts for determining entailment decisions. In this work, we present a comprehensive evaluation of a variety of prompting methods (zero-shot, few-shot static, few-shot dynamic, Chain-of-Thought, self-consistency, and Tree-of-Thought) with two LLMs, DeepSeek-R1-Distill-Qwen-14B and LLaMA-3.1-8B-Instruct, from the prompt-engineering perspective. We applied these methods to the BioNLI dataset and reported on key evaluation metrics across all methods. Our results show that dynamic contextual in-context prompting, together with structured reasoning, produces high-quality inference in our context. Between all of the models and configurations, few-shot ToT prompting using the DeepSeek model produced the best results, scoring a macro-F1 score of 71.05, even outperforming retrieval-augmented models reported on in prior studies. These findings show that prompt engineering alone can handle complex biomedical reasoning effectively, without needing retrieval or full fine-tuning. https://miscj.aut.ac.ir/article_6116.html Masked face segmentation plays a crucial role in face recognition, security, and medical applications, yet remains challenging due to occlusions, varying lighting conditions, and diverse mask patterns. To address these issues, we propose MBFANet, a lightweight and efficient deep learning model that integrates multi-branch feature fusion and attention mechanisms for real-time segmentation. Our architecture is built upon RepVGG, an efficient convolutional backbone, and is enhanced by two key modules: Channel Aggregation and Coordinate Refinement (CACR) for improving spatial and channel-wise feature representation, and the Atrous Pyramid Attention Module (APAM) for capturing multi-scale contextual information. Additionally, a Multi-Branch Feature Fusion (MBFF) module is employed in the decoder to refine segmentation outputs. To optimize segmentation performance, we introduce a hybrid loss function that combines Dice Loss and Focal Loss, effectively addressing class imbalance while improving boundary precision. Experimental evaluations on the Masked Face Segmentation Dataset (MFSD) demonstrate that MBFANet achieves superior segmentation accuracy, outperforming existing models in IoU, F1-score, recall, and precision while maintaining a low parameter count (10M) and high inference speed (153 FPS). Ablation studies confirm the effectiveness of the proposed modules, showing consistent improvements across all evaluation metrics. With its balance between accuracy, efficiency, and computational cost, MBFANet presents a highly effective solution for real-time masked face segmentation, making it well-suited for practical deployment in security, biometric authentication, and mobile applications. https://miscj.aut.ac.ir/article_6110.html Time delays in data transmission are a critical factor that significantly influence the stability and performance of pocket-based control systems, especially under realistic network conditions where delays are unpredictable and variable. These delays, stemming from communication constraints, often include challenges such as packet disorder and loss, which can compromise control accuracy and system robustness. To address this issue, this paper introduces a novel round-trip delay compensation strategy that explicitly considers both feedback and feedforward delays. These delays are modeled as stochastic but constrained entities, accurately capturing the uncertainties present in real-world network environments. The impact of these delays on system stability is rigorously analyzed using spectral radius methods, providing a quantitative assessment of stability margins and robustness. Additionally, the effects of key parameters such as sampling time and the design of controllers based on linear quadratic regulation are investigated through comprehensive simulations, highlighting their influence on system stability and performance. Practical validation is performed on a pocket-based inverted pendulum system, demonstrating that the proposed approach reduces overshoot by approximately 20%, shortens the settling time by about 25%, and decreases peak motor torque by 23%. These results underscore the enhanced robustness and reliability achieved through the proposed delay compensation, making it a practical and effective solution for networked control systems operating under variable communication delays and packet irregularities. https://miscj.aut.ac.ir/article_6203.html Compared with conventional radiation techniques, proton therapy provides highly conformal dose distributions while reducing irradiation of surrounding normal tissues. Combining proton beams with different energies generates an extended Spread-Out Bragg Peak (SOBP) to cover the target. Passive scattering uses a range modulator wheel (RMW) to adjust proton energy and broaden the Bragg peak. In this study, a proton therapy nozzle comprising an RMW, compensator, and collimators was simulated for ocular tumor treatment using the GATE Monte Carlo code. A three-dimensional computational eye phantom including the lens, aqueous humor, cornea, choroid, retina, sclera, vitreous body, optic nerve, and tumor was developed to evaluate ocular dose distributions. The designed RMW generated a relatively uniform SOBP with a target length of 3.0 mm, a distal dose gradient (DDF) of 0.9 mm, and a d90 of 21.5 mm. The tumor dose was normalized to 100%, while relative doses were 52% for the lens, 63% for the aqueous humor, 62% for the cornea, 60% for the retina, 51% for the choroid, 44% for the sclera, 67% for the vitreous body, and 1% for the optic nerve. The results indicate that the RMW and compensator produced the intended SOBP and localized dose distribution within the modeled tumor geometry. These findings demonstrate the potential of the proposed nozzle configuration for achieving target-specific dose distributions while limiting irradiation of critical ocular structures. The developed computational framework provides a useful platform for investigating ocular proton-nozzle design and treatment-planning strategies. https://miscj.aut.ac.ir/article_6114.html This paper introduces Sentence Splitter, a self-supervised framework built upon a T5-based encoder--decoder architecture for uncovering the latent factual structure of natural language sentences. The proposed method identifies the semantic boundary between a descriptive prefix (head) and its factual completion (tail) by formulating sentence splitting as a discrete segmentation problem, where a sentence of length $N$ admits $N$ possible split points but only one recovers the intended head--tail structure. Rather than explicitly searching over all candidate boundaries, the model learns to recover the factual completion through probabilistic sequence generation. To eliminate the need for manual annotation, symbolic head--tail pairs are first verbalized into natural-language templates that provide supervision for training the Sentence Splitter. The trained splitter is then applied to raw text to extract aligned prefix--tail pairs, which are subsequently used to train a generative model that proposes additional plausible completions through a lightweight bootstrapping process. This unified pipeline provides a scalable and structure-aware approach to constructing self-supervised training data while bridging symbolic knowledge and natural language. Experiments on both structured and naturally occurring text demonstrate that the proposed splitter generalizes beyond synthetic templates and that the resulting structure-aware supervision consistently improves downstream performance on knowledge graph completion and commonsense question answering, highlighting the effectiveness of recovering latent factual structure for knowledge-centric NLP. https://miscj.aut.ac.ir/article_6195.html This paper develops an exploration-aware nonlinear model predictive control framework for safe online learning of unknown nonlinear residual dynamics. A recursive sparse Gaussian process with a fixed inducing set is used to update the residual model from closed-loop measurements while keeping the online estimator dimension independent of the accumulated data. A deterministic bounded-noise sparse-GP error certificate is expressed using recursively available quantities and converted into a nested residual envelope that remains nonexpanding under successive noisy updates. This envelope is used to construct a causally updated rigid tube whose nominal center follows only the known dynamics and whose radius remains fixed during each NMPC optimization. A certificate-aware exploration objective combines mutual information with the predicted change in the deterministic certificate, thereby favoring measurements that are both statistically informative and beneficial for uncertainty reduction. Exploration is performed over the same robust feasible set as a baseline regulation controller and is constrained by a prescribed performance-degradation budget. The resulting scheme guarantees recursive feasibility and hard state and input constraint satisfaction under every bounded-noise RSGP update without model-update rejection or post-update feasibility tests. Numerical studies demonstrate improved model accuracy and sample efficiency while preserving hard constraints, with moderate computational overhead during safe closed-loop operation. https://miscj.aut.ac.ir/article_6163.html Unmanned Aerial Vehicles (UAVs) equipped with environmental sensors and imaging systems provide an efficient platform for precision agriculture and early crop disease detection. This study presents a UAV-based intelligent framework for the identification of Solanum lycopersicum (tomato) leaf diseases by integrating RGB image analysis with real-time environmental sensing. A lightweight six-arm UAV was developed using a Raspberry Pi 3 controller, GPS, radio telemetry, RGB camera, and onboard sensors for temperature, humidity, rainfall, and gas concentration monitoring. The acquired RGB images and environmental sensor data were used to construct a multimodal dataset for disease classification. Three classification models, namely 2D Convolutional Neural Network (2D-CNN), Support Vector Machine (SVM), and Random Forest (RF), were trained and evaluated using a common testing dataset. The 2D-CNN achieved an accuracy of 69.10%, whereas the SVM improved the classification performance to 93.40% with a 35.17% relative improvement in accuracy over the 2D-CNN. The Random Forest classifier achieved the highest performance with an accuracy of 98.61%, precision of 99.40%, recall of 98.24%, and F1-score of 98.82%, representing a 42.71% improvement in accuracy compared with the 2D-CNN and a 5.58% improvement over the SVM. The results demonstrate that integrating UAV-based RGB imaging with environmental sensing significantly improves disease identification under field conditions. The proposed framework provides a reliable and cost-effective solution for early crop health assessment, reduces manual inspection effort, and supports precision agriculture through data-driven decision making. https://miscj.aut.ac.ir/article_6152.html Customer churn prediction is critical for businesses, yet most prior work relies on structured data and makes limited use of textual reviews, leaving low-resource languages such as Persian underexplored. Text-based churn models require large labeled corpora, which are often expensive and time-consuming to produce. To address this, we propose a practical method that combines least-confidence active learning, explainable AI (LIME) to assist expert annotators, and iterative semi-supervised labeling (self-training, label propagation, and a hybrid method). To our best knowledge there is no study on using active learning, semi-supervised learning and explainable AI to label textual data for churn prediction. Beginning with an expert-verified seed of 2,004 balanced Persian reviews, we trained LSTM/GRU classifiers on ParsBERT embeddings, augmented training sets via active sampling (100 samples per fold), and expanded labels to 10,500 Digikala reviews over 12 semi-supervised stages. Active learning increased average accuracy from 82.5% to 84.7%, and self-training produced the best sustained gains (mean peak accuracies 86%), outperforming label propagation and the hybrid approach. Moreover, the explanations generated by LIME were positively evaluated by three domain experts, with mean ratings exceeding 4 for two experts and above 3.9 for the third, reflecting consistently high satisfaction. In addition, expert agreement increased by up to 8%, reaching 0.91 after reviewing the explanations. The proposed framework offers a reproducible solution for churn prediction in low-resource text domains by introducing the first labeled Persian review dataset for churn prediction and a unified framework integrating active learning, semi-supervised learning, and explainable AI. https://miscj.aut.ac.ir/article_6166.html This study presents the Fishing Boat Group (FBG), a new metaheuristic optimization method based on how fishing boats at sea work together when they hunt for fish. This paper’s main contribution is the development of an algorithm that, through minimal only one parameter tuning, achieves superior computational efficiency for high-dimensional problems while consistently demonstrating robust performance and competitive fitness values across multimodal benchmarks. FBG categorizes boats into three groups (A, B, and C) to search regions near the best boat (G_best), the virtual median boat, and unexplored areas. To check how well it works, FBG was tested against six well-known and newer metaheuristic algorithms using ten multimodal benchmark functions and two real-world engineering design problems. The experimental results show that FBG did better than the other methods in 8 of the 10 tests, suggesting it stays pretty robust and performs consistently across different cases. In practical applications, the algorithm achieved superior mean fitness values for the Three-bar truss (263.896) and Tension/compression spring (0.012989313) designs. Most notably, FBG maintained the shortest execution times across all benchmark functions, with an overall average of 0.68 seconds. These results suggest that FBG could be a strong option for tackling high-dimensional optimization problems, especially in situations where you need both accurate results and fast computation. https://miscj.aut.ac.ir/article_6153.html Gamma-ray shielding materials play a crucial role in nuclear technology, medical physics, and radiation protection, where efficient attenuation of ionizing radiation is required. In this study, the gamma-ray shielding performance of several polymer-based materials, namely polyurethane (PUR), polypropylene (PP), polytetrafluoroethylene (PTFE), polyvinylidene fluoride (PVDF), and polyvinylidene fluoride–hexafluoropropylene (PVDF-HFP), together with silicon carbide (SiC), included as a non-polymeric ceramic reference material, was systematically investigated and compared with conventional lead (Pb) using the MCNPX Monte Carlo code. The shielding effectiveness of each material was evaluated in terms of key radiation interaction parameters, including the mass attenuation coefficient (MAC), half-value layer (HVL), and mean free path (MFP), over a wide photon energy range of 122, 356, 511, 662, 1170, 1275, and 1330 keV, corresponding to commonly used radioactive sources in medical and industrial applications. The obtained results indicate that lead exhibits the highest gamma-ray attenuation capability across the entire energy spectrum, as expected due to its high atomic number and density. Among the investigated materials, SiC demonstrates superior shielding performance compared to the polymers, followed by PTFE, particularly in low and intermediate-energy regions. These materials also show comparatively lower HVL values and shorter mean free paths, indicating improved radiation attenuation efficiency. Overall, the findings suggest that SiC and PTFE-based polymer systems can serve as promising lightweight and non-toxic alternatives to lead for gamma-ray shielding applications, especially in low to medium-energy radiation environments. https://miscj.aut.ac.ir/article_6150.html Online question-and-answer (Q&A) repositories have become vital resources for knowledge acquisition. In specialized domains such as Islamic jurisprudence, these repositories present additional complexity: short, concise questions are paired with lengthy, context-dependent answers written in highly domain-specific language. This structure makes automatic and interpretable topic discovery both crucial and challenging. Conventional topic modeling and recent embedding-based clustering approaches have improved semantic representation, yet they continue to suffer from cluster instability, poor label quality, and a lack of reproducible evaluation standards, limiting their practical reliability and interpretability. To overcome these issues, this study introduces HybriDToD, a hybrid deep topic discovery framework that combines the scalability of unsupervised learning with the semantic precision of Large Language Models (LLMs). The core innovation lies in integrating two novel modules, Topic Auditor and Interactor, that together establish a feedback-driven evaluation loop. The Topic Auditor employs an LLM-as-a-Judge mechanism to assess topic quality across six rubrics, providing structured, human-like evaluation without extensive manual effort, while the Interactor incorporates minimal expert feedback to refine and validate the automatically generated topics. Experiments on the ISLAMQA dataset demonstrate that HybriDToD achieves superior coherence, interpretability, and stability. This research contributes a scalable and explainable framework for organizing and analyzing large religious or educational Q&A archives, advancing automated topic modeling toward greater transparency, accountability, and domain adaptability. https://miscj.aut.ac.ir/article_6151.html Spelling correction is a challenge in the field of natural language processing. The objective of spelling correction tasks is to recognize and rectify spelling errors automatically. The development of applications that can diagnose and correct Persian spelling errors has become more important in order to improve the quality of text. Typographical Error Type Detection in Persian is a relatively understudied area. Therefore, this paper presents an approach for detecting typographical errors in Persian texts. Our work includes the presentation of a publicly available dataset called FarsTypo, which comprises 3.4 million words arranged in chronological order and tagged with their corresponding part-of-speech. These words cover a wide range of topics and linguistic styles. We develop an algorithm designed to apply Persian-specific errors to a subset of these words. By leveraging FarsTypo, we establish a baseline and conduct a comparison of various methodologies employing different architectures. We propose a Deep Sequential Neural Network for token-level error type classification. The model incorporates three key components: joint word-level and character-level embeddings to handle both in-vocabulary and out-of-vocabulary words, bidirectional LSTM layers to capture contextual dependencies in text, and a TimeDistributed classifier for predicting 51 error categories. Experimental results demonstrate that the proposed model achieves 97.62% accuracy, 98.83% precision, and 98.61% recall on the constructed dataset. In addition, the model shows higher inference efficiency compared to industrial spell-checking systems. The results indicate that separating error detection from correction and leveraging combined word–character sequential modeling improves typographical error type detection.