Engineering Science & Technology

Explainable AI: ACombined XAI Framework for Interpreting Slice-Level Brain Tumour Classification Models

Patrick McGonigle, William Farrelly, Kevin Curran — Tue, 18 Aug 2026 00:00:00 +0800

This study investigates how multiple Explainable Artificial Intelligence (XAI) techniques can be combined to improve the interpretability of a deep learning model for brain tumour analysis in magnetic resonance imaging. A custom Convolutional Neural Network (CNN) was developed and trained on two-dimensional Fluid-Attenuated Inversion Recovery (FLAIR) slices derived from the BraTS 2021 dataset to perform slice-level binary classification, distinguishing slices that contain tumour tissue from slices that do not. It should be noted that non-tumour slices are not healthy controls, as they may originate from patients who have a brain tumour elsewhere in the volume. The improved model achieved a slice-level test accuracy of 91.24 percent; this figure reflects performance over individual 2D slices rather than patient-level or volume-level diagnosis. Three complementary XAI methods were applied: Gradient-weighted Class Activation Mapping (Grad-CAM), Layer-wise Relevance Propagation (LRP) and SHapley Additive exPlanations (SHAP). Grad-CAM highlighted broad spatial regions of interest, LRP provided pixel-level relevance and SHAP quantified the contribution of individual features. Presented together, these methods offered layered, qualitatively complementary explanations ranging from coarse regions of interest to fine pixel-level detail, including cases where slices captured only partial views of a tumour. The combined presentation provided a more rounded view of the model’s behaviour than any single method considered in isolation, although the study does not yet provide a quantitative measure of explanation faithfulness. The work illustrates the potential of integrated XAI techniques to support transparency in AI-assisted medical image analysis, while acknowledging that clinical usefulness would require multi-sequence and three-dimensional analysis, quantitative evaluation of the explanations, and assessment by radiologists. These remain directions for future work.

Design of Asymmetric Janus Aerogel Toward High-Performance Solar Evaporation

Fri, 31 Jul 2026 00:00:00 +0800

Interfacial solar evaporation technology enables high-efficiency evaporation with low energy consumption through solar-driven localized heating and has broad application prospects in seawater desalination and industrial wastewater treatment. As novel functional materials with asymmetric physicochemical characteristics, Janus aerogels provide a new approach for overcoming the performance limitations of traditional evaporators. In this work, a Janus aerogel was fabricated using Polyvinyl Alcohol/Cellulose Nanofiber (PVA/CNF) as the substrate and Carbon Nanotubes (CNT) as the photothermal layer. Owing to its asymmetric structure, which integrates hydrophilic water transport and hydrophobic photothermal characteristics, the resulting aerogel achieves efficient photothermal conversion and directional water delivery. Its maximum surface temperature reaches 64.4 °C with a light absorption efficiency of 99%, consistent with the finite element simulation results. Under 1 sun, the evaporation rate is 1.32 kg·m^-2·h^-1 and the corresponding evaporation efficiency is 88.37%, and the aerogel maintains stable evaporation performance over 7 cycles, with only minor salt precipitation on the surface after 2 h of irradiation. This developed material shows promising potential in seawater desalination and wastewater purification, offering a novel strategy for the rational design of multifunctional aerogel evaporators.

Numerical Investigation of Seismic Demands in Floor Diaphragms and Collectors

Alireza Sheykhaleslami, Shervin Maleki — Wed, 05 Aug 2026 00:00:00 +0800

This study investigates the seismic demand in diaphragms, chords, and collectors in typical floor systems used in steel structures using nonlinear Finite Element (FE) analyses. These systems include prefabricated one-way steel joist floor systems, open-web steel joists, and composite steel decks. The Lateral-Force-Resisting Systems (LFRSs) considered were Special Concentrically Braced Frames (SCBFs) and Special Moment Frames (SMFs) of varying heights. For each system, the effects of using shear connectors for collector beams, joist direction, geometric configuration, and number of stories on lateral force transfer were examined. Analyses were conducted on selected models. The study shows, in steel joist floor systems without shear connectors on collector beams, the diaphragm force transfer path is significantly influenced by joist orientation. Furthermore, using shear connectors on collectors bypasse joist participation in lateral load transfer, allowing the floor slab to transmit forces directly to the LFRS like a rigid diaphragm through the collectors. In composite steel decks, using shear connectors on collector beams was found to improve diaphragm ductility and cyclic performance. Overall, the study clarifies the true structural role of diaphragms, demonstrating that without proper analysis models, seismic load paths can become disrupted and traditional design assumptions may fail to ensure structural safety. The findings indicate that the introduction of diaphragm-specific response modification factors of one or less (Rs ≤ 1) results in more realistic diaphragm forces. Applying the outcomes of this research will align diaphragm design practice more closely with the actual behavior.

Uncertainty-Aware Semantic Segmentation via Decision Boundary Modeling and Confusion Zone Refinement

Antoni Mestre, Franccesco Malafarina — Mon, 10 Aug 2026 00:00:00 +0800

Deep learning has significantly improved the performance of semantic segmentation in high-resolution imagery and computer vision tasks. However, standard segmentation models treat all pixel-wise predictions equally, despite the fact that predictions close to the decision boundary are inherently uncertain. This limitation often leads to unstable classifications and increased false positives or false negatives in complex scenes. In this paper, we propose a practical uncertainty-aware semantic segmentation refinement framework that explicitly handles decisionboundary ambiguity through an empirical routing mechanism through a Confusion Zone mechanism. Instead of forcing a binary decision for all pixels, predictions whose probability scores fall within a predefined interval [τ_low, τ_high] are marked as uncertain and redirected to a secondary refinement module that exploits additional spatial features and local context before assigning a final label. The proposed two-stage architecture allows the segmentation model to treat ambiguous predictions differently from confident ones, improving robustness near object boundaries and visually complex regions. Experiments on benchmark remote sensing segmentation datasets demonstrate that incorporating the Confusion Zone mechanism improves segmentation quality, achieving up to 0.851 IoU with SegFormer and consistent gains of 2-3 percentage points across U-Net, DeepLabV3+, and SegFormer backbones. These results suggest that explicitly routing decision-boundary predictions for local refinement is a simple and practical strategy for improving the reliability of deep semantic segmentation systems.

Explainable Machine Learning and Necessary Condition Analysis for Product Sales Forecasting in Retail

Marcin Nowak, Paweł Kościelniak — Wed, 12 Aug 2026 00:00:00 +0800

Sales forecasting in retail and e-commerce-related environments supports key decisions concerning inventory management, promotion planning, pricing policy, and sales strategy optimization. The aim of this article is to develop a predictive framework for product sales forecasting using machine learning regression algorithms, Explainable Artificial Intelligence (AI) methods, and Necessary Condition Analysis (NCA). The study employed the Walmart M5 retail sales forecasting dataset, including daily product-level sales observations with calendar, product hierarchy, event-related, price-related, and historical demand features. The benchmark included a naive weekly model, Ridge regression, Random Forest, Extra Trees, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM). Predictive performance was evaluated using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), R², Weighted Mean Absolute Percentage Error (WMAPE), and Root Mean Squared Scaled Error (RMSSE) under rolling-origin validation and on a final 28-day test period. The best-performing model was interpreted using Permutation Feature Importance and SHapley Additive exPlanations (SHAP), while NCA was applied to identify threshold-like bottlenecks associated with high sales outcomes. The results show that Extra Trees achieved the best test performance, with RMSE = 1.6183, MAE = 0.9109, R² = 0.5852, WMAPE = 0.7708, and RMSSE = 0.7226. Permutation Feature Importance (PFI) and SHAP indicated that the most important predictors were rolling sales averages, product identity, weekend effects, SNAP-related information, and demand volatility. NCA further showed that selected rolling demand averages and volatility measures constituted necessary conditions for achieving high sales levels. The proposed Machine Learning (ML)-eXplainable Artificial Intelligence (XAI)-NCA framework therefore supports not only sales prediction but also the identification of interpretable demand thresholds relevant to retail decision-making.

Integrating Ethereum and SGX-TEE for Job Execution

Thu, 13 Aug 2026 00:00:00 +0800

This work proposes an architecture that integrates Ethereum blockchain technology and Trusted Execution Environments (TEEs), focusing on Intel Software Guard Extensions (SGX). The main motivation is to address existing challenges and limitations in on-chain job execution by smart contracts, proposing a reliable method that ensures the integrity and confidentiality of off-chain processed jobs. The proposed architecture is based on a strategic combination of the blockchain’s smart contract and the TEEs. The smart contract acts as a decentralized coordinator, responsible for job distribution, monitoring the state, and dynamically allocating resources among TEE machines. In this architecture, the SGX allows the creation of secure enclaves that serve as trusted environments where jobs can be processed in isolation, ensuring protection against external and internal attacks. A prototype implementation was developed to demonstrate the technical feasibility and potential benefits of the proposed architecture. Through experimental testing, limits and bottlenecks were observed for scenarios with varying numbers of TEE machines, revealing a direct proportional growth in response times according to the load, as well as the ability of horizontal scaling to increase system performance.

A Scan-Based Analysis of Internet-Exposed IoT Devices Using Shodan Data

Richelle Williams, Fernando Koch — Tue, 25 Aug 2026 00:00:00 +0800

Determining whether scan-observable network configurations capture meaningful security exposure characteristics across large-scale Internet of Thing (IoT) populations remains an open problem in measurement research. This study analyzes Internet-exposed IoT endpoints associated with the TR-069 protocol within a global population exceeding 37 million hosts, using a structured sample of 7,634 hosts derived from an initial collection of approximately 10,000 hosts obtained via Shodan Search and Shodan InternetDB. Data were collected during a fixed observation window in April 2026 and sampled across ten countries with the highest observed prevalence of TCP port 7547 exposure. Hosts were enriched with scan-derived metadata, including open ports, service indicators, CPE identifiers, vulnerability-associated fields, and network attribution metadata, enabling construction of a composite exposure proxy capturing externally observable service-surface characteristics. The analysis combines descriptive statistics, nonparametric hypothesis testing using the Kruskal–Wallis H test, and exploratory feature separability analysis to evaluate whether exposure patterns vary systematically across geographic regions and whether observable servicelevel characteristics differentiate between exposure groups. Results demonstrate statistically significant cross-country variation in exposure structure, reflected in differences in both average exposure levels and distributional characteristics. The exploratory feature separability analysis using logistic regression and random forest approaches yielded a balanced accuracy of 0.58, indicating that scan-derived service-level features capture only limited information related to the exposure proxy. These findings suggest that Internet-wide scan data can support systematic, population-level analysis of IoT exposure while also highlighting the limitations of port-based features as indicators of broader exposure characteristics. The study contributes a reproducible framework for large-scale exposure measurement using previously collected Internet-wide scan data, without requiring device interaction or exploit-based validation.

Engineering Secure AI Systems with Lattice-Based Cryptography and Large Language Models

Devharsh Trivedi — Thu, 20 Aug 2026 00:00:00 +0800

Deploying large language models in safety-critical domains while migrating to post-quantum cryptographic standards poses a compounded systems engineering challenge: infrastructure must become privacy-preserving, computa-tionally viable, and quantum-resistant at once. This review examines three frontiers where lattice-based cryptography and large language models converge, applying an evidence-grading scheme that separates established results from conjecture. The first frontier is fully homomorphic encryption for private model training. Ciphertext error does not provide differential privacy: it is not calibrated to query sensitivity, involves no gradient clipping, and is removed by decryption. The two mechanisms defend against disjoint adversaries and must be composed, not substituted. The second frontier is the hypothesis that language models can act as adaptive agents inside lattice reduction. No empirical result supports it. The review further shows that scheduling improvements cannot reduce core security estimates, since these depend on the minimum viable block size rather than the number of reduction tours, and supplies a falsifiable benchmark protocol in place of the parameter-selection guidance of earlier treatments. The third frontier is functional encryption for decentralized inference. Inner-product constructions admit exact reconstruction of hidden states by any node holding a spanning key set, requiring no cryptanalysis, so security rests on key allocation rather than lattice hardness. Attention is bilinear and lies outside the inner-product functionality class. Quantitatively, encrypted inference for a small encoder costs roughly five orders of magnitude more than plaintext, measured on a graphics processing unit; functional key material reaches four tebibytes per attention layer under the parameters the security proof requires, reducible by three orders of magnitude at a stated security cost; and memory capacity, not arithmetic throughput, is the binding constraint. Readiness levels for the three frontiers span one to three. Trusted execution environments remain the only approach in production today.