<?xml version="1.0" encoding="utf-8"?>
    <rss version="2.0">
      <channel xmlns:content="http://purl.org/rss/1.0/modules/content/">
        <title>Frontiers in Artificial Intelligence | Machine Learning section | New and Recent Articles</title>
        <link>https://www.frontiersin.org/journals/artificial-intelligence/sections/machine-learning</link>
        <description>RSS Feed for Machine Learning section in the Frontiers in Artificial Intelligence journal | New and Recent Articles</description>
        <language>en-us</language>
        <generator>Frontiers Feed Generator,version:1</generator>
        <pubDate>2026-10-23T17:16:52.232+00:00</pubDate>
        <ttl>60</ttl>
        <item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1884880</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1884880</link>
        <title><![CDATA[SpectralGuard: a Spectrally grounded λ-SHAP framework for XAI validity certification in pharmaceutical supply chain networks]]></title>
        <pubdate>2026-10-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ntebogang Dinah Moroke</author><author>Mmapula Merriam Lebele</author>
        <description><![CDATA[Permutation-based explainability methods such as SHAP are central to regulatory compliance in AI-governed pharmaceutical supply chains across Sub-Saharan Africa, yet their validity depends on a structural property that existing frameworks do not monitor: the algebraic connectivity of the supply hypergraph, quantified by the Fiedler eigenvalue λ2. We prove that λ2 → 0 is a sufficient condition for permutation-based XAI failure (Theorem 3.5): no increase in sampling effort can recover validity once network connectivity collapses. We introduce SpectralGuard, demonstrated on a calibrated synthetic DGP driven by real Eskom load-shedding and SARB ZAR/USD records (n = 3, 227 business days, 2014–2026); all spectral results are from the synthetic hypergraph. λ-SHAP shrinks attributions toward a fluid-regime prior using λ2(t) as the shrinkage weight, maintaining auditability at an explicit bias–variance cost (Remark 3.12). A probabilistic validity certificate V(t) combines four spectral metrics with SHA-256 tamper-evidence. An oracle experiment (B = 500, Seed = 2,026) demonstrates 70.4–89.2% attribution MSE reduction across governance regimes. The spectral gate is the decisive ablation component. In the synthetic DGP, early-warning detection leads classical baselines by 19 business days (λc alert preceding the λm crossing). Field validation on real CCMDD procurement data is planned for 2027 (PLMIF-2). A nine-step DHIS2 integration pathway supports SDG 3, 9, 13, and 16.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1840945</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1840945</link>
        <title><![CDATA[Maximum entropy relaxation of multi-way cardinality constraints for synthetic population generation]]></title>
        <pubdate>2026-10-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>François Pachet</author><author>Jean-Daniel Zucker</author>
        <description><![CDATA[Generating synthetic populations from aggregate statistics is a core component of microsimulation, agent-based modeling, policy analysis, and privacy-preserving data release. Beyond classical census marginals, many applications require matching heterogeneous unary, binary, and more generally multi-way constraints derived from surveys, expert knowledge, or automatically extracted descriptions. Constructing populations that satisfy such multi-way constraints simultaneously poses a significant computational challenge. We consider populations where each individual is described by categorical attributes and the target is a collection of global frequency constraints over attribute combinations. Exact formulations scale poorly as the number and arity of constraints increase, especially when the constraints are numerous and overlapping. We propose a maximum-entropy relaxation of this problem. Multi-way cardinality constraints are matched in expectation rather than exactly, yielding an exponential-family distribution over complete population assignments and a convex optimization problem over Lagrange multipliers. The convex program defines the theoretical MaxEnt solution; in practice we approximate it with a fixed-budget MCMC moment-matching procedure, and all reported results measure the empirical performance of this finite-budget approximation. We evaluate the approach on NPORS-derived scaling benchmarks with four to 40 attributes, and on a controlled synthetic benchmark family generated independently of NPORS, and compare it primarily against generalized raking. The results show that MaxEnt becomes increasingly advantageous as the number of attributes and ternary interactions grows, while raking remains competitive on smaller, lower-arity instances.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1973276</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1973276</link>
        <title><![CDATA[An energy-aware diffusion graph neural network based on intuitionistic fuzzy credibility graphs for research-paper classification]]></title>
        <pubdate>2026-10-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Sadhu Varun</author><author>Shaik Sharief Basha</author>
        <description><![CDATA[The classification of research papers from different academic data is still challenging due to the presence of semantic ambiguity, lack of information, fluctuating credibility, and complex linkages across research articles. Existing fuzzy graph neural networks depend too much on membership information and cannot truly consider the intuitionistic uncertainty, trust dependability, and structural importance of nodes at the graph propagation level. To overcome these constraints, this paper presents an intuitionistic fuzzy graph neural network architecture that combines trust-aware intuitionistic representations and a unique node Laplacian energy mechanism. The suggested Node Laplacian Energy is derived from the graph Laplacian's row-wise norm and is used as an adaptive node-weighting metric to assess each research paper's local structural contribution. Furthermore, a diffusion-based propagation technique is used to enhance information transmission across semantically linked publications and capture larger graph relationships. The suggested system is tested for research-paper categorization with semantic similarity features and intuitionistic fuzzy credibility information. Its performance is compared to known machine learning and graph neural network baselines utilizing accuracy, precision, recall, and F1-score. The findings show that including uncertainty-aware representations, structural node weighting, and diffusion-based graph propagation can improve discriminative academic content categorization. The suggested paradigm offers a unified way to model semantic linkages, uncertainty, believability, and graph structural relevance in research-paper categorization.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1965356</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1965356</link>
        <title><![CDATA[Proximal Policy Optimization-guided novelty-aware multi-strategy particle swarm optimization for adaptive traffic signal optimization]]></title>
        <pubdate>2026-10-23T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Manikandan R.</author><author>Arun Pandian J.</author>
        <description><![CDATA[IntroductionTraffic signal timing is a non-linear, stochastic and time-varying optimization problem. Particle Swarm Optimization (PSO), Genetic Algorithm (GA) and Ant Colony Optimization (ACO) run one fixed search strategy for the whole run, which often leads to early convergence on a poor timing plan. This study proposes PPO-NAMSPSO, which changes the search strategy during the run.MethodsA Proximal Policy Optimization (PPO) agent observes the state of a Novelty-Aware Multi-Strategy PSO (NAMSPSO) and, at each iteration, selects one of six operators: standard PSO update only, novelty-guided exploration, elite-guided update, mutation, restart, or local search. Candidates are ranked on waiting time, queue length and behavioral novelty. An elite pool guides the swarm, and stagnation recovery reinitializes the worst half of the particles after four iterations without improvement. We tested the framework in the Simulation of Urban Mobility (SUMO) on a synthetic 3 × 3 grid and on an OpenStreetMap-derived network with 25 signalized intersections, at medium and high demand. PSO, GA, ACO and fixed-time control were compared, and each optimizer was tuned with Optuna under the same budget. Each optimizer was trained with five seeds and evaluated over 100 runs.ResultsAgainst PSO, average waiting time fell from 105.52 s to 96.97 s (8.10%) on the grid and from 48.88 s to 33.07 s (32.35%) on the OSM network, where its standard deviation fell from 17.42 s to 3.54 s. Under high demand on the grid, waiting time was 207.72 s against 315.91 s for PSO. PPO-NAMSPSO had the lowest waiting time of the five methods in all three reported comparisons. On the OSM network, ANOVA and Kruskal-Wallis tests found differences among methods (p < 0.001), and Holm-corrected Welch tests against each method gave p ≤ 1.17 × 10−5, with Cohen's d from 0.71 to 2.33. Removing the PPO agent raised mean waiting time from 60.38 s to 64.91 s under high demand.DiscussionThe gain was larger on the OSM network than on the grid. The agent used all six operators, and the elite-guided update was chosen most often (20.61% of selections). Demand was synthetic, so validation on field data remains open.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1958096</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1958096</link>
        <title><![CDATA[A systematic review of DoS and DDoS attack detection in cloud environments: evolutionary trends and the emerging role of meta-learning for cross-domain adaptability]]></title>
        <pubdate>2026-10-21T00:00:00Z</pubdate>
        <category>Systematic Review</category>
        <author>Amit Tiwari</author><author>C. Navaneethan</author>
        <description><![CDATA[Denial of Service (DoS) and Distributed Denial of Service (DDoS) attacks continue to be one of the most persistent and rapidly evolving threats to current network infrastructures as they are becoming more sophisticated and increasingly disrupting the availability, reliability and security of cloud-based systems, Software-Defined Networks (SDNs), Internet of Things (IoT) ecosystems and next generation communication environments. This paper provides a systematic review of the literature on DoS and DDoS attack detection and mitigation, covering the various stages of evolutions in the field from early statistical and anomaly-based methods to adaptive machine learning and deep learning, as well as domain specific detection methods. The review, which is based on a thorough examination of 120 papers (114 primary studies and 6 supplementary studies), provides a synthesis of the latest advances in network and application-layer detection, protocol-specific detection frameworks, cloud-based intrusion detection systems, SDN-enabled programmable detection, IoT/Fog/Edge architectures, ensemble and hybrid models, transfer learning, and blockchain-based security mechanisms. A systematic review of benchmark datasets, feature engineering techniques and comparative performance shows that current deep learning systems attain almost perfect classification rates in experimental environments, but fundamental limitations in their generalizability, computational efficiency, preservation of privacy and explainability remain a constraint for their deployment in real-world environments. Importantly, this is a lot of work, but less than 6 % of the reviewed studies (seven out of 120) use meta-learning, transfer learning or few-shot approaches explicitly, which reflects the ongoing lack of capability in the field of adaptive generalization and a deficiency in addressing zero-day attacks and domain shift in dynamic cloud environments. We propose the emerging paradigm of meta-learning as the most promising one for next-generation cloud-based intrusion detection, which can be adapted to new domains with very few labeled examples — directly overcoming the three key limitations of existing approaches, namely the zero-day detection gap, domain shift limitations and label scarcity. This review offers a structured basis for researchers and practitioners designing explainable, privacy-preserving and computationally efficient detection systems and highlights the critical need for moving along the path of designing cloud-based intrusion detection systems based on meta-learning.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1918025</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1918025</link>
        <title><![CDATA[ClimateEdge: a lightweight weather-augmented LSTM framework for real-time demand response in climate-stressed smart grids]]></title>
        <pubdate>2026-10-21T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Harsha Kumar A</author><author>Venkatesan Murugan</author><author>Dhanalakshmi R</author><author>Sahaya Beni Prathiba</author><author>Suria R. Asai</author>
        <description><![CDATA[Extreme weather events driven by climate change impose non-linear load surges on electricity grids at the moments when accurate forecasting is most critical and cloud-based inference latency is unacceptable. This paper presents ClimateEdge, a lightweight framework that integrates weather-augmented feature construction, a Climate-Enhanced Long Short-Term Memory (CE-LSTM) architecture with temporal attention, and an INT8 post-training quantization pipeline to enable real-time demand response inference on an NVIDIA Jetson Nano edge device. The framework is evaluated on two grid datasets spanning contrasting climate scenarios: the California Independent System Operator (CAISO) summer heatwave corpus, covering June 2020 to September 2022 at hourly resolution and containing 20,422 records with 1,228 extreme-heat events, and the ISO New England (ISONE) winter cold-snap corpus, covering July 2023 to February 2024 and containing 5,813 records with 210 extreme-cold events. In both cases the CE-LSTM outperforms five baseline models including a full-size iTransformer, achieving a Mean Absolute Error (MAE) of 471.8 MW and a Mean Absolute Percentage Error (MAPE) of 1.55% on the CAISO extreme-weather partition, representing a 21.1% improvement over the next-best baseline, and an MAE of 329.6 MW and MAPE of 2.31% on the ISONE extreme-weather partition, a 19.4% improvement. The INT8-quantized CE-LSTM runs at 9.2 ms per inference step with a runtime memory footprint of 8.1 MB and a power draw of 3.4 W on the Jetson Nano, achieving a 3.65 × latency speedup over the FP32 baseline at a cost of less than 1% relative MAE increase. Ablation studies confirm that the weather-augmented features account for the largest share of the accuracy improvement, and that temporal attention up-weights climate-stressed timesteps during both heatwave and cold-snap events.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1931803</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1931803</link>
        <title><![CDATA[QDentalNet: a geometry-aware hierarchical mesh learning and quantum-enhanced network for automated dental occlusion classification]]></title>
        <pubdate>2026-10-21T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Ramakrishnan Ilanghovan</author><author>S. Alden Jenish</author><author>R. Karthik</author><author>K. Suganthi</author><author>R. Praveen</author>
        <description><![CDATA[ObjectivesThe accurate analysis of dental occlusion is important to orthodontic diagnosis and occlusal classification. The existing assessment procedures largely depend on clinician-driven evaluation of intraoral scans, which introduces subjectivity and limits scalability in routine diagnostic workflows. To address these challenges, this study proposes an automated dental occlusion classification architecture that integrates geometry-aware mesh learning with a quantum-enhanced network.MethodsThe proposed QDentalNet employs a hierarchical mesh-processing backbone designed to capture both local and global geometric characteristics of dental structures. Rotation-aware feature encoding and kernel-based geometric correlation are incorporated through the Face Rotation Convolution (FRC) and Face Kernel Correlation (FKC) modules. The network also adopts a learning strategy with task towers and quantum-enhanced classification heads. This design helps the prediction of multiple clinically relevant occlusal classes. Additionally, auxiliary features via upper and lower contact point estimation enable efficient feature learning. The study also presents a novel composite loss function that enhances the discriminative capability of the network by addressing class imbalance, anatomical consistency, and efficient feature learning.ResultsThe proposed network achieved an average classification accuracy of 80.46% across all occlusal tasks. The obtained performance was also contextualized against previously reported results from existing 3D dental classification approaches as reference benchmarks.ConclusionTo the best of our knowledge, this is one of the initial studies to incorporate geometry-aware features with quantum-based decision-making for dental occlusion classification, providing a proof-of-concept for automated occlusal assessment.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1920823</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1920823</link>
        <title><![CDATA[Bayesian network-based risk stratification of allergic sensitization using NHANES population data]]></title>
        <pubdate>2026-10-21T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Assem Shayakhmetova</author><author>Nurbolat Tasbolatuly</author><author>Anar Sultangaziyeva</author><author>Marat Shurenov</author><author>Aigul Bissarinova</author><author>Nurbol Kurishbay</author>
        <description><![CDATA[Allergic sensitization is commonly modeled as a binary classification outcome, an approach that discards the probabilistic structure linking demographic, immunological, and clinical variables and offers limited support for individualized risk communication. This work develops and evaluates a Bayesian network (BN) for probabilistic stratification of laboratory-defined serological IgE sensitization, fitted on harmonized data from the National Health and Nutrition Examination Survey (NHANES) 2005–2006 cycle (analytic cohort n = 8,285). The network integrates demographic covariates (age, sex), immunological markers (total serum IgE, eosinophil percentage), and clinical indicators (physician-diagnosed asthma, self-reported allergic symptoms) into a directed acyclic graph whose conditional probability tables were estimated by maximum-likelihood estimation with Laplace smoothing. Serological IgE sensitization, defined as at least one allergen-specific IgE ≥ 0.35 kUA/L across a 19-allergen panel, was present in 47.1% of the cohort. A five-fold cross-validated comparison of BN, logistic regression (LR), and random forest (RF) on identical predictor sets showed indistinguishable discrimination: AUC 0.808/0.809/0.808 (IgE + eosinophils) and 0.820/0.824/0.822 (IgE + eosinophils + age + sex). Individual-level calibration was excellent: Brier score = 0.171, calibration slope = 0.9998 (95% CI 0.9563–1.0434), intercept = −0.0001 (95% CI − 0.0522 to 0.0520). The three posterior probability strata (low < 0.40, intermediate 0.40–0.70, high > 0.70) correspond structurally to the three total IgE tertiles; adding eosinophil category improves continuous discrimination by ΔAUC = +0.010 but reclassifies zero participants between strata (categorical NRI = 0). Bidirectional inference—updating sensitization probability from a confirmed asthma diagnosis—raised the posterior from 0.472 to 0.651 (ΔAUC +0.006, asthma subsample n = 8,275), confirming a structural advantage unavailable from unidirectional classifiers. Observed serological sensitization rates across strata were 13.8% (low), 46.1% (intermediate), and 81.6% (high). The framework is offered as a transparent, auditable methodological demonstration of probabilistic modeling of serological IgE sensitization. It does not model clinical allergy and should not be used to guide clinical decisions without prospective validation. Its principal advantages over discriminative classifiers are structural: bidirectional inference, modular auditability of conditional probability tables, and native uncertainty propagation.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1951063</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1951063</link>
        <title><![CDATA[Comparative analysis of GAN-based augmentation pipelines for Parkinson's hypokinetic dysarthria severity detection using deep learning]]></title>
        <pubdate>2026-10-20T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Aurobindo S</author><author>Prakash Ramachandran</author><author>Rajesh Kumar M</author>
        <description><![CDATA[Class imbalance is a common problem in speech-based severity detection, which can result in biased models and poor recognition of minority severity classes. Generative Adversarial Networks (GANs) have been widely used for data augmentation, but very few studies have comparatively evaluated GAN-based augmentation at different stages of the speech-processing pipeline for assessing Parkinson's hypokinetic dysarthria severity. This paper compares GAN-generated speech (GGS) and GAN-generated image (GGI) augmentation methods for addressing class imbalance in low-frequency cepstrogram-based severity detection using the PC-GITA dataset. GAN-based augmentation was used to generate balanced training datasets using two different augmentation pipelines. The original speech and synthetic speech generated using GGS were converted into low-frequency cepstrogram representations, whereas GGI directly generated low-frequency cepstrogram images. These representations were used as inputs to a CNN trained and evaluated under Speaker-independent (SI) and Speaker-dependent (SD) conditions. In the speaker-independent evaluation, repeated five-fold speaker-stratified cross-validation with five independent random seeds yielded a mean accuracy of 56.47% ± 3.40% for GGS compared to 55.47% ± 3.35% for GGI and 53.13% ± 3.26% for the original-dataset baseline. GGS showed higher mean accuracy, precision, and specificity, whereas GGI showed higher mean recall and F1-score. In the speaker-dependent evaluation, GGI achieved the highest mean accuracy of 76.48% ± 4.09% compared to 73.04% ± 3.63% for GGS and 73.04% ± 3.25% for the original-dataset baseline. However, the speaker-dependent results represent an optimistic upper-bound evaluation because speaker identity was shared between the training and evaluation partitions. In general, the two GAN-based augmentation pipelines showed potential to address class imbalance under the conditions studied, while their relative performance depended on the evaluation protocol and the classification measure.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1975692</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1975692</link>
        <title><![CDATA[The rise of super-emotional intelligence]]></title>
        <pubdate>2026-10-20T00:00:00Z</pubdate>
        <category>Perspective</category>
        <author>Sophie Bonnaire</author><author>Farah I. Corona-Strauss</author><author>Daniel J. Strauss</author>
        <description><![CDATA[The debate about the future of artificial intelligence (AI) runs along a single axis: cognitive capability and its projected culmination in superintelligence. We argue that this framing leaves a complementary axis unexamined and propose the goal of Super-Emotional Intelligence (SEI): systems that perceive, model, regulate, and respond to affect better than skilled human experts, while remaining human-centered. SEI is a performance construct realized by a functional architecture, not a claim about machine experience, and we give measurable criteria for each of its four capacities. Today's affect-aware AI, from multimodal emotion recognition to empathic chatbots, recognizes, or simulates emotion at or below human level and is judged against human ceilings. Three things would set SEI apart: perception that adds peripheral and contactless physiological channels, none of them emotion-specific, but some, like the vestigial ear muscles, resolving affective processing at the timescale of the events that evoke it; a generative model of affect grounded in active inference rather than pattern matching; and functional allostatic self-regulation. The decisive gap in current systems is architectural. A sensed state conditions a prompt, but nothing carries a state estimate across turns, models the effect of the system's own responses, or selects actions under an explicit objective. We specify such a closed loop, state the conditions under which it is stable and steerable rather than manipulative, propose a preregistered validation plan whose failure would falsify our claims, and treat governance as a design constraint of the loop rather than an afterthought.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1952539</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1952539</link>
        <title><![CDATA[Physics-inspired feature engineering and explainable machine learning for tensile strength prediction of CNT–Si₃N₄ reinforced FWTPET joints]]></title>
        <pubdate>2026-10-20T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Krishnan Venkatachalam</author><author>Senthil Kumaran Selvaraj</author><author>R. Lokeshkumar</author><author>G. Sundaramali</author>
        <description><![CDATA[IntroductionThe prediction of the tensile strength of carbon nanotube (CNT)–silicon nitride (Si₃N₄) reinforced friction welding of tube-to-tube plate using an external tool (FWTPET) joints is crucial for improving joint performance.MethodsA physics-inspired machine learning framework was developed using 25 experimentally derived observations to investigate the relationships between welding parameters, engineered process descriptors, and tensile strength. Physically motivated descriptors were constructed to represent thermal input, reinforcement interaction, material flow, and dispersion effects. Hyperparameter optimization was performed for multiple regression models to improve predictive performance.ResultsThe XGBoost model showed the best performance, with an MAE of 5.70, an RMSE of 7.74, and an R² of 0.9772. SHAP analysis identified CNT content and the dispersion-related descriptor as the most influential predictors. Ablation analysis showed that the combination of the original experimental process variables provided greater predictability than the engineered variables alone or their combination, while the physics-inspired process variables provided further physically interpretable representations of the welding process.DiscussionThe results indicate that small-data machine learning, domain-based feature engineering, and explainable analysis are potential techniques for predicting the tensile strength of FWTPET joints.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1960066</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1960066</link>
        <title><![CDATA[Data-based surrogate modeling and performance prediction of MEMS touch mode capacitive sensing using deep multilayer perceptron networks]]></title>
        <pubdate>2026-10-19T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Pushparaj E</author><author>Sumit Kumar Jindal</author>
        <description><![CDATA[IntroductionThis work presents a structurally optimized rectangular diaphragm that operates in touch mode as the basis for a high-performance micro-electro-mechanical system (MEMS) capacitive pressure sensor (CPS). Rectangular diaphragms offer significant advantages in terms of structural control, constant stress handling and directional extension of the contact region.MethodsThe proposed framework encompasses structural design, analytical modeling, numerical computation, finite element analysis (FEA) and Artificial Intelligence (AI)-based prediction. In this design, greater directional control results in a smoother and more reliable change in capacitance with pressure. The rectangular diaphragm design can be changed to smooth the transition between normal and touch modes and produce more consistent linearity and sensitivity. Numerical computation, FEA simulation and deep neural network (DNN) surrogate model for predicting the performance characteristics of a novel MEMS capacitive pressure sensor. A comprehensive dataset is constructed by combining Numerical computation and FEA-simulated results across a wide range of diaphragm geometries and applied pressure conditions. A DNN surrogate model is subsequently developed and trained on this dataset to predict the performance characteristics of the proposed MEMS capacitive pressure sensor.ResultsThe model achieves an overall mean absolute percentage error (MAPE) of 0.22 % and R2 above 0.999 across all four outputs, a substantial speed improvement, with prediction time reduced to a matter of seconds compared to the tens of min required for direct FEA simulation.DiscussionThe results show that the proposed rectangular diaphragm effectively controls the touch region. It also improves capacitance and sensitivity. The DNN surrogate model predicts sensor performance precisely and efficiently. It enables rapid design-space exploration for MEMS capacitive pressure sensors.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1935983</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1935983</link>
        <title><![CDATA[Two-stage transformer encoder and attention-guided GRU-based residual correction for short-term load forecasting]]></title>
        <pubdate>2026-10-19T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Radhika Chandrasekaran</author><author>Senthil Kumar Paramasivan</author>
        <description><![CDATA[Accurate short-term load forecasting (STLF) plays a vital role in reliable power system planning and operations. The increasing electricity demand is driven by electrification, digitalization, population, and economic growth, which necessitates the need for accurate short-term load forecasting to ensure the real-time balance between supply and demand. Traditional forecasting approaches struggle to capture long-range temporal dependencies, and correcting the forecasting errors remains challenging due to increasing demand and dynamic load behavior. To address this limitation, this study proposes a two-stage forecasting architecture (transformer encoder and attention-guided GRU-based residual corrector [TE-AG-GRU-RC]) for short-term load forecasting. In the first stage, a transformer encoder (TE) processes historical load and exogenous features to generate the base forecast and capture long-range temporal dependencies from historical load sequences. In the second stage, an attention-guided GRU residual correction mechanism (AG-GRU-RC) captures short-term dynamics and refines the initial prediction with the temporal information derived from the transformer encoder. The final forecast is attained by combining the predicted residual with the base forecast, enabling the first stage to model dominant load patterns, while the second stage corrects the remaining errors. The proposed framework is evaluated using real-world multivariate hourly time series datasets from two geographically distinct regions. The forecasting performance of the model is evaluated against several baseline models, such as recurrent neural network (RNN), long short-term memory (LSTM), Gated Recurrent Unit (GRU), temporal fusion transformer (TFT), and ResidualLSTMPlus, to demonstrate its effectiveness. The proposed model attains mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R2) values of 0.0297, 0.0014, 0.0376, 3.58, and 0.879, respectively, on the Tamil Nadu test data, and 0.0334, 0.0023, 0.0482, 4.99, and 0.790, respectively, on the Panama test data. These findings demonstrate the generalizability of the proposed model across both datasets with low prediction error and improved forecasting accuracy.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1960601</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1960601</link>
        <title><![CDATA[ReqTestAgent: execution-grounded language-model agents for acceptance testing under requirement ambiguity]]></title>
        <pubdate>2026-10-19T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Yunjiao Shi</author><author>Rui Song</author>
        <description><![CDATA[IntroductionLarge language models (LLMs) can translate ambiguous requirements into executable tests whose assertions conflict with the accepted behavior. Evaluating such tests requires distinguishing behavioral compatibility from fault-detection strength.MethodsWe present ReqTestAgent, an execution-grounded workflow that connects requirement interpretation, testing-knowledge retrieval, test generation, and feedback from canonical execution and mutation testing. We evaluate 100 requirement instances formed from 25 Orchid/BigCodeBench tasks, each with lexical, semantic, syntactic, and vagueness variants, using a reference implementation available to the evaluator. Comparisons include three generation baselines, their verifier-enhanced counterparts, and component ablations.ResultsThe generation baselines fail the canonical implementation on 51%–68% of instances, whereas all verifier-enhanced methods achieve 100% canonical pass. ReqTestAgent achieves 92% qualified suites and a macro-averaged effective mutation score of 0.678, compared with 0.655–0.670 for the verifier-enhanced baselines. Removing verification reduces its effective score to 0.216. Across instance–mutant evaluations, ReqTestAgent kills 492 of 756 mutants, a pooled rate of 0.651; this rate weights instances by their mutant counts. A comparative execution trace shows how explicit domain assertions detect faults missed by the verifier-enhanced baselines.DiscussionExecution feedback provides the principal validity improvement in this controlled setting. The workflow links interpretation records to executable checks, while the evaluation establishes canonical compatibility as a prerequisite for interpreting mutation-based test strength.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1950536</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1950536</link>
        <title><![CDATA[AFI-ID: adaptive feature intelligence for intrusion detection]]></title>
        <pubdate>2026-10-19T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>S. Antony Joseph Raj</author><author>M. Madiajagan</author>
        <description><![CDATA[The increasing sophistication of cyber-attacks necessitates intrusion detection systems (IDS) that are both accurate and computationally efficient, particularly in high-dimensional and class-imbalanced environments. Although ensemble-based IDS models enhance predictive performance, most existing approaches treat feature selection and ensemble aggregation as independent stages, limiting integrated optimization between feature representation and decision-making. To overcome these limitations, this paper proposes Adaptive Feature Intelligence for Intrusion Detection (AFI-ID), an intrusion detection framework that integrates hybrid adaptive feature intelligence with attention-based deep stacking. The proposed AFI module combines statistical relevance measures—Mutual Information and Chi-square with model-based importance derived from XGBoost and Extra Trees through a weighted fusion strategy. This strategy enables the selection of a compact and discriminative feature subset while preserving sensitivity to minority attack classes. At the ensemble level, heterogeneous base learners generate structured meta-features, including log-probability representations, prediction confidence, and entropy. These meta-features are processed by an attention-based Bidirectional Long Short-Term Memory (BiLSTM) meta-classifier, which dynamically weights classifier contributions to achieve context-aware decision fusion. Experimental evaluation on the UNSW-NB15 and CICIDS2017 datasets under binary and multiclass settings demonstrates that AFI-ID achieves strong performance relative to the evaluated baseline classifiers under the experimental protocols considered in this study. AFI-ID achieves 93.31% accuracy with a Receiver Operating Characteristic—Area Under Curve (ROC-AUC) of 0.9849 in binary classification and 79.95% accuracy with a ROC-AUC of 0.9845 in multiclass classification on UNSW-NB15. On CICIDS2017, AFI-ID achieves 98.33% accuracy with a ROC-AUC of 0.9976 in binary classification and 93.54% accuracy with a ROC-AUC of 0.999 in multiclass classification. In repeated UNSW-NB15 binary experiments across five random seeds, AFI-ID achieved a mean accuracy of 93.62%, with paired t-tests indicating statistically significant improvements over Logistic Regression (p = 1.08 × 10−10), Extra Trees (p = 1.38 × 10−5), and XGBoost (p = 3.69 × 10−5). These results demonstrate the effectiveness of the proposed adaptive feature selection and attention-based decision fusion strategy under high-dimensional and class-imbalanced network traffic conditions.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1977034</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1977034</link>
        <title><![CDATA[Pre-arrival OCC decision support for aircraft turnaround coordination: a human-in-the-loop architecture and scenario-based evaluation]]></title>
        <pubdate>2026-10-19T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Parth Yogeshbhai Purohit</author><author>Thomas Feuerle</author><author>Peter Hecker</author>
        <description><![CDATA[Aircraft turnaround coordination depends on maintenance, catering, passenger-assistance, gate, and bottleneck information that may become available before arrival but is often handled through fragmented workflows. This paper presents a human-in-the-loop Operations Control Center (OCC) decision-support architecture for pre-arrival aircraft turnaround coordination. The architecture integrates component-aware maintenance risk indicators, service-area inventory readiness, process-class bottleneck forecasting, passenger-assistance readiness logic, and a transparent rule-based fusion layer that converts heterogeneous signals into readiness scores, alert levels, and operator-facing action queues. The evaluation is deliberately framed as controlled synthetic scenario analysis rather than airline deployment validation. Existing module-level holdout results are combined with a deterministic scenario sweep of 1,000 synthetic inbound cases across six operational regimes, monotonicity and sanity checks, and ablation and baseline comparisons. The scenario sweep produced mean readiness 47.4, median readiness 50.6, 370 Critical cases, 284 Priority cases, and 188 Routine cases. Ablations show that removing maintenance influence increased mean readiness by 13.7 points, while removing bottleneck influence increased mean readiness by 10.0 points, indicating that the fusion layer materially changes prioritization under controlled stress conditions. Two maintenance-related monotonicity checks did not strictly pass because small local readiness increases occurred at low malfunction levels before urgency thresholds changed; these failures are treated as limitations of the current model-fusion interaction rather than hidden. The work contributes a reproducible OCC decision-support architecture and scenario-evaluation package, while explicitly preserving the limitations that no airline operational data, OCC user study, deployment, certification, or measured operational savings are claimed.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1943572</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1943572</link>
        <title><![CDATA[A global machine learning framework for classifying national CO2 emission levels: responsible and explainable artificial intelligence for sustainable climate governance]]></title>
        <pubdate>2026-10-16T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Muhammad Ali Khawaja</author><author>Muhammad Farooq Shaikh</author>
        <description><![CDATA[We investigate whether trade openness, governance, environmental taxation and selected macroeconomic indicators can classify the contemporaneous CO2 emission level of countries observed over 2000 to 2023, using data from the World Development Indicators and the OECD. CO2 emissions per capita are divided into three ordinal levels, low, medium and high, using tertile cut points estimated on the training data only within every validation split. Four classifiers, Random Forest, XGBoost, a support vector machine and a multinomial logistic regression, are tuned and compared alongside two models designed specifically for ordinal outcomes, an ordered logistic regression and an ordinal decomposition of XGBoost, under four validation designs: a country grouped holdout and country grouped 5 fold cross validation, in which no country appears in both the training and test sets; a chronological holdout; a 1 year ahead forecasting task; and a random stratified country year split, retained only as a reference point. Under country grouped cross validation, no classifier achieves a decisively higher accuracy than the others: nested cross validation gives a mean accuracy of 0.644 for XGBoost, 0.648 for Random Forest, 0.671 for the support vector machine, 0.674 for the multinomial logit, 0.676 for the ordered logit and 0.637 for the ordinal XGBoost model, with overlapping confidence intervals throughout. This is markedly lower than the 0.91 accuracy obtained under a random country year split with the same tuned pipeline, a gap of 12.5 percentage points for XGBoost, confirming that a substantial part of the high accuracy obtainable under random splitting reflects the model recognizing previously seen country profiles rather than generalizing to unseen countries. Accuracy falls less sharply under the chronological holdout and the 1 year ahead forecasting task, in the region of 0.65 to 0.74, indicating that the classifiers generalize better across time within a country than across countries. Permutation importance, SHAP and LIME consistently identify government effectiveness, urban population growth and environmental tax revenue as the features most strongly associated with model predictions, with a moderate contribution from trade openness and the smallest from GDP growth, a ranking that is stable across importance methods, learning algorithms and cross validation folds. These variables are predictors of the model's output, not estimated causal determinants of emissions, and the results are best read as evidence for climate risk monitoring and further policy evaluation rather than as a causal policy assessment.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1884749</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1884749</link>
        <title><![CDATA[Conditional diffusion-based inverse kinematics for redundant 7-DOF manipulators with differentiable refinement]]></title>
        <pubdate>2026-10-16T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Lesia Hentosh</author><author>Yaroslav Hladun</author><author>Ivan Tsmots</author>
        <description><![CDATA[BackgroundInverse kinematics of redundant manipulators is a multi-solution problem, since one end-effector pose may correspond to multiple admissible joint configurations. Classical numerical solvers and regression-based neural models usually return a single joint configuration and therefore do not explicitly represent the inverse-kinematic solution manifold. Existing generative solvers based on normalizing flows produce multiple solutions, but they require invertible network architectures and do not include an explicit task-space correction stage. This study aimed to develop a hybrid generative-kinematic method for inverse kinematics of a redundant 7-DOF manipulator that can generate multiple admissible configurations for a given target pose.MethodsA hybrid method, DiffusionIK, was developed by combining conditional diffusion-based generation, classifier-free guidance, DDIM sampling, and differentiable forward-kinematics refinement. A ResMLP denoiser is trained in the normalized joint space using procedurally generated training pairs obtained from the forward kinematics of the xArm manipulator. For each target pose, the diffusion model generated multiple candidate joint configurations, and the refinement stage subsequently reduced the task-space residual for each candidate. Numerical experiments were performed on 500 reachable target poses with 50 generated candidates per pose. Performance was evaluated using position and orientation success thresholds of 1 mm and 1°, respectively, together with the preservation of joint-space variability after refinement.ResultsDiffusionIK reached 99.40% success after 200 refinement steps and 100.00% after 500 steps under 1 mm and 1° thresholds, with a mean joint-space variability of 9.77 rad. In a unified benchmark against damped least squares, KDL, direct regression, and a normalizing-flow baseline, DiffusionIK with refinement achieved 99.80% success on 500 unseen poses and 98.40% near singularities, matching the strongest numerical solver while returning 50 diverse solutions per pose.ConclusionThe proposed DiffusionIK method provides an effective formulation of redundant inverse kinematics as conditional generation followed by local kinematic correction. By combining diffusion-based candidate generation with differentiable forward-kinematics refinement, the method represents the multi-solution nature of redundant manipulators and achieves high task-space accuracy while preserving configuration diversity. Unlike normalizing-flow-based solvers, DiffusionIK imposes no invertibility constraints on the denoiser and combines generation with differentiable forward-kinematics refinement, which provides sub-millimetre and sub-degree accuracy together with configuration diversity.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1887024</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1887024</link>
        <title><![CDATA[A machine learning framework for stage-structured predator–prey systems with fear effect: Sparse Identification, Gaussian Process Regression, and bifurcation analysis]]></title>
        <pubdate>2026-10-16T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>G. Ramraj</author><author>T. Poornima</author>
        <description><![CDATA[This paper presents a comprehensive machine learning framework for the analysis of a stage-structured predator–prey system incorporating the fear effect, intra-specific competition among juvenile predators, and maturation delay. The three-species model, comprising prey (N), juvenile predators (P1), and mature predators (P2), is governed by a non-linear ordinary differential equation system in which the prey growth rate is modulated by a fear factor 1/(1+kP2). We first establish the existence and local stability of the coexistence equilibrium E* = (16.02, 16.09, 32.19) and rigorously identify a supercritical Hopf bifurcation at a critical fear intensity kcrit ≈ 0.067, beyond which the equilibrium becomes unstable and a stable limit cycle emerges. Five machine learning methods are deployed and benchmarked: (i) Sparse Identification of Non-linear Dynamics (SINDy) with a generic polynomial library, which recovers the predator equations exactly (R2 = 1.0) and approximates the prey equation (R2 = 0.9997); (ii) Dynamic Mode Decomposition (DMD), whose linear assumption yields poor reconstruction (R2 = −2.885 for P2), demonstrating the inadequacy of linear methods for this non-linear system; (iii) Gaussian Process Regression (GPR) with a periodic–RBF composite kernel, achieving RMSE values of 1.92, 1.03, and 0.84 from sparse noisy observations; (iv) a physics-constrained function approximation for recovering population trajectories from incomplete data; and (v) a polynomial surrogate ODE model for derivative field approximation. Additionally, the effect of maturation delay τ on system stability is examined through delay differential equation analysis with a fourth-order Runge–Kutta integrator, revealing delay-induced destabilization and transitions to limit cycle dynamics at τ ≥ 10. All numerical results are independently verified and all performance metrics are computed from actual analyses. The framework provides a rigorous, reproducible methodology for applying modern data-driven methods to stage-structured ecological models.]]></description>
      </item><item>
        <guid isPermaLink="true">https://www.frontiersin.org/articles/10.3389/frai.2026.1944175</guid>
        <link>https://www.frontiersin.org/articles/10.3389/frai.2026.1944175</link>
        <title><![CDATA[PharmAI: domain-specific QA generation by adaptive learning and fine-tuning of small language models]]></title>
        <pubdate>2026-10-15T00:00:00Z</pubdate>
        <category>Original Research</category>
        <author>Shenbagarajan Anantharajan</author><author>Saranraj Loganathan</author><author>Rajkumaran Soundarraj</author><author>E Emerson Nithiyaraj</author>
        <description><![CDATA[BackgroundDomain-specific medical question-answering (QA) models face bottlenecks due to scarce labeled data, specialized terminology, and constrained computational hardware.ObjectivesThe study aims to develop an efficient pharmaceutical question-answering (QA) system that can be trained and deployed on consumer-grade hardware using Small Language Models (SLMs).Materials and methodsThe pipeline processes four pharmaceutical textbooks containing 3,977 total pages, producing 7,370 semantically similar text chunks and 2,144 synthetic QA pairs under four questioning categories: definition, mechanism, comparison, and clinical application. Three Small Language Models (SLMs), namely SmolLM2-1.7B-Instruct, Qwen2.5–1.5B-Instruct, and Llama-3.2–1B-Instruct, are fine-tuned under the same QLoRA (Quantized Low-Rank Adaptation) configurations and then evaluated with retrieval metrics (Recall@K, Mean Reciprocal Rank) and generation metrics (ROUGE-1/2/L, Semantic Similarity, BERTScore). Architectural decisions including NF4 4-bit quantization, CPU-based FAISS indexing, and completion-only loss masking are implemented due to hardware constraints and validated via ablation analysis.ResultsSmolLM2-1.7B-Instruct, Qwen2.5–1.5B-Instruct, and Llama-3.2–1B-Instruct demonstrated identical retrieval performance, achieving a Recall@1 of 0.517, Recall@3 of 0.732, Recall@5 of 0.797, and an MRR of 0.628. Similarly, the generation performance evaluation revealed that Qwen2.5–1.5B FT outperformed the other fine-tuned SLMs, achieving the highest scores across all metrics (ROUGE-1 = 0.47, ROUGE-2 = 0.25, ROUGE-L = 0.38, and BERTScore = 0.72). In comparison, SmolLM2-1.7B FT and Llama-3.2–1B FT exhibited comparable performance, with ROUGE-1 scores of approximately 0.35, ROUGE-2 scores of 0.16–0.17, ROUGE-L scores of 0.25–0.26, and BERTScores of 0.66, indicating lower lexical and semantic alignment with the reference outputs than Qwen2.5–1.5B FT.ConclusionWe evaluate retrieval and generation performance using comprehensive metric suites, incorporating a category-wise breakdown that reveals comparison questions to be the most challenging for single-passage retrieval architectures.]]></description>
      </item>
      </channel>
    </rss>