Abstract: Ear biometrics has emerged as a promising alternative in biometric recognition systems, offering robustness in unconstrained environments where traditional modalities such as face recognition may fail on its own, but can be enhanced by ear. Ear segmentation, in particular, plays a crucial role in downstream recognition by isolating discriminative ear regions and reducing background interference. However, existing approaches to ear detection and segmentation are commonly susceptible to severe occlusions, ear accessories, and variable illumination, and their performance deteriorates on images captured in the wild. To address these limitations, we introduce a tailored ear-segmentation architecture based on a U-Net with a ResNet-50 encoder. Trained and validated on the Annotated Web Ears (AWE) dataset, our method achieves a mean Intersection over Union (IoU) of 77.1% and a pixel-wise accuracy of 99.7%, outperforming the Convolutional Encoder--Decoder (CED) baseline. We further evaluate on the EarSegDB-25 dataset, where our approach attains a test-set IoU of 94.76%, significantly surpassing previous ear segmentation methods based on the original U-Net architecture. High pixel-wise accuracy across methods is largely attributable to background dominance; in contrast, the improved IoU achieved by our approach more accurately reflects gains in ear region segmentation performance. Leveraging a ResNet-50 encoder, our model demonstrates robust performance under occlusion and illumination challenges, achieving state-of-the-art results on AWE and EarSegDB-25 and showing strong potential for biometric applications in unconstrained environments.

Abstract: Based on outer raceway control theory and considering the effects of elastic deformation, centrifugal force, and gyroscopic moment between the rolling elements and raceways, a geometric and force analysis of angular contact ball bearings is conducted. A five-degree-of-freedom theoretical model capable of accounting for the combined action of radial force and moment is established. The accuracy of the model is verified through numerical calculations and experimental results from existing literature. Upon validation of the theoretical model, a modified Archard model is employed to develop a wear volume model for the bearing raceways. The influence of both single and combined loads on sliding wear in the bearing raceways is systematically analyzed.

Abstract:

Methane production from Brosimum alicastrum seed coat was evaluated using a logistic model through three alkaline concentrations (0.19 M, 0.26 M, and 0.28 M) and three enzymatic activity levels (3000 U mL^-1, 5000 U mL^-1, and 7000 U mL^-1) as pretreatments. Laccase was produced through submerged fermentation using T. hirsuta Bm-2 fungi, while NaOH served as the alkaline agent. Enzymatic pretreatment resulted in the highest specific CH₄ yield (427.43±2.28 mL CH₄/g VS_added), surpassing both alkaline pretreatment (235.61 ± 9.19 mL CH₄/g VS_added) and the control (102.54 ± 5.55 mL CH₄/g VS_added). Kinetic analysis of CH₄ production indicated that cumulative CH₄ production reached its stationary phase within 30 days of digestion. Moreover, enzymatic pretreatment exhibited the highest CH₄ formation rate (0.15–0.17 h^-1), except for the control, which had a slightly higher rate (0.21 h^-1). The kinetic analysis revealed that the enzymatic pretreatment significantly improved the hydrolysis stage of Ramon's seed coat, promoting higher cumulative CH₄ production and leading to an increased specific CH₄ yield.

Abstract: Hydrogen (H2) is regarded as a promising option for sustainable energy systems; however, its large-scale use in electricity supply remains limited. This study develops a stochastic network optimization model to examine the applicability of H2-based electricity generation. The proposed Hydrogen Supply Chain (HSC) model evaluates cost and emission performance under uncertainty by considering disaster conditions, transmission losses, depreciation, and the time value of money. The Marmara Region of Türkiye is divided into 24 grid nodes, and a single-period model for 2023 is solved using Mixed-Integer Linear Programming (MILP). The HSC is allowed to meet 10–40% of electricity demand and to replace collapsed grid lines by supplying critical public centers (CPCs) during disasters. The results show that the HSC can meet about 25% of demand, although at costs higher than power grid (PG) electricity, while keeping emissions near zero. The model is then extended to a multi-period structure (2023–2053) and solved by Variable Neighborhood Search (VNS). Over time, H2 costs decline, and its share rises from 19% to 35%. These findings suggest that H2 can support long-term sustainability, resilience, and energy security.

Abstract: Floating belt conveyor routes, consisting of serially arranged belt conveyors, the end parts of which are mechanically attached to floating bodies, are designed for the continuous transport of extracted granular materials from the water. The paper deals with the analytical determination of the position of the centre of gravity of the buoyancy force, the coordinates of which change depending on the longitudinal deflection of the floating body from the equilibrium state, which acts as a supporting element of individual conveyor belts. The analysis of the individual phases of deflection of the floating body, consisting of a pair of floats with a circular cross-section, shows that the complete submergence of one of the floats occurs at a higher value of the angle of inclination in the case when the floats are initially submerged under the surface to exactly half of their diameter. On the realized experimental device the buoyancy force was detected using strain gauges during the deflection of the floating body from the equilibrium position for three defined levels of immersion. The floating body of the experimental device consists of a pair of floats with a circular cross-section with a diameter of 80 mm. The output is a structured methodological procedure for determining the position of the centre of gravity of the displacement (centre of buoyancy) of a floating body when it deviates from the equilibrium position and a methodology for calculating the stability arm, which is a key parameter for assessing the buoyancy and stability of the body. On the basis of the laboratory measurements, the magnitude of the buoyancy force can be quantified as a function of the immersion depth of the floating body. It was found that the buoyancy force remains constant when the body deflects only when the immersion corresponds to half the diameter of a float with a circular cross-section. If the depth of the immersion is less than the radius of the float, the buoyancy force increases during deflection; on the contrary, if the immersion is greater than the radius of the float, the buoyancy force decreases.

Abstract: The vibroacoustic simulation of geared drivetrains has become increasingly important as electrified powertrains expose tonal gear noise and high-frequency structure-borne excitation more clearly than conventional internal-combustion vehicles. In this context, software choice is no longer a secondary implementation detail but a central engineering decision, because different platforms emphasize different parts of the excitation–transfer–radiation chain. This review therefore examines gearbox and geared-drivetrain NVH simulation from a software-specific perspective rather than a purely phenomenon-based one. The article critically compares dedicated gearbox CAE tools, general multibody dynamics platforms, integrated multiphysics and structural–acoustic finite-element environments, and early-stage 1D system simulation tools. The comparison covers major software ecosystems including KISSsoft/KISSsys, Romax Suite, SMT MASTA/DRIVA, MSC Adams, AVL EXCITE, RecurDyn/DriveTrain, Siemens Simcenter 3D Motion / Transmission Builder / Acoustics, SIMULIA Simpack, Ansys Motion with Mechanical/Acoustics and Motor-CAD, COMSOL Multiphysics, GT-SUITE, and Simcenter Amesim, while also considering relevant recent module extensions and workflow updates. The review shows that the current software landscape is structured around four main methodological layers: dedicated gearbox analysis tools that are strongest in gear-contact modeling and microgeometry iteration; high-fidelity multibody platforms that are strongest in system-level dynamic response and transmission-path representation; integrated structural–acoustic environments that provide the deepest access to housing vibration and radiated-noise prediction; and 1D or multidomain system tools that are most efficient for early concept evaluation and architecture-level trade-off studies. Recent developments since 2023 indicate a clear shift toward tighter support for electrified drivetrain NVH, measured manufacturing deviations, optimization workflows, and faster acoustic prediction, including reduced-order or embedded acoustic methods. At the same time, major gaps remain. Open literature still contains relatively few independent studies that validate the full chain from tooth contact and transmission error through dynamic transfer paths to housing vibration and radiated sound within a single commercial workflow. Likewise, interoperability for measured flank topography, wear-driven NVH evolution, and fully validated electro-magnetic–mechanical–acoustic simulation remains limited and uneven across platforms. For this reason, the review argues that current software ecosystems are best understood not as universally proven end-to-end solutions, but as partially overlapping toolchains with different strengths, evidence levels, and practical compromises.

Abstract: To address the challenge of separating fine-grained apatite from layered silicate gangue minerals (chlorite and biotite) in medium-low grade collophanite ores, this study systematically investigated the effect of carboxymethyl cellulose sodium (CMC-Na) as a selective depressant on flotation behavior of different particle size fractions and its underlying mechanism. Pure mineral and artificial mixed ore flotation experiments demonstrated that at pH 9 and collector dosage of 5 kg/t, CMC-Na enabled selective separation of apatite from gangue minerals, with optimal dosage showing significant particle size effects: for the -0.5+0.074 mm fraction, effective separation was achieved with collector alone; for the -0.074+0.023 mm fraction, the optimal CMC-Na dosage was 10~100 mg/L, yielding 87% apatite recovery for pure minerals and 41.8% recovery with 23.7% P₂O₅ grade for mixed ores; for the -0.023 mm fine fraction, the optimal dosage was 30~300 mg/L, achieving 24.8% recovery and 13.2% grade. Mechanism studies revealed that CMC-Na significantly enhanced the hydrophilicity of chlorite and biotite, enlarging their surface property differences with apatite. FTIR and XPS analyses indicated that CMC-Na adsorbed on biotite via ion exchange with interlayer K⁺ and coordination with octahedral Fe²⁺/Mg²⁺, and on chlorite through chemical coordination with octahedral Mg²⁺, whereas only weak physical adsorption occurred on apatite surface Ca²⁺. The adsorption strength followed the order: biotite > chlorite > apatite. This study provides an effective reagent scheme and theoretical basis for flotation separation of fine-grained phosphate ores.

Abstract: Accurate evaluation of indoor daylighting performance is essential for improving visual comfort and reducing lighting energy use in office buildings. However, simulation-based daylighting analysis is often too time-consuming to support rapid comparison of multiple design options in early-stage design. To address this issue, this study proposes MTL-Light, an explainable chained multi-task learning framework for fast daylighting performance prediction in typical office units. A parametric simulation dataset was constructed, and multiple representative daylighting indicators were extracted from the spatial distribution of daylight factors on the work plane. MTL-Light was then developed to jointly predict these indicators by modeling their interdependencies within a lightweight multi-task learning architecture. In addition, SHAP was employed to interpret the prediction results by quantifying the marginal contributions of geometric design variables. The results show that, compared with single-task models, MTL-Light achieves higher accuracy and more stable performance across multiple indicators, particularly for metrics sensitive to spatial distribution. Moreover, it reduces daylighting evaluation from minute-level simulation to millisecond-level inference. The interpretability analysis further indicates that room depth and window geometry dominate daylighting performance, while different indicators exhibit different sensitivities to geometric variables.

Abstract: As the era of Industry 4.0 (i.e., the fourth-generation industrial revolution) develops, machine tools in particular are becoming interconnected, forming a collaborative community in smart factories. “Smart manufacturing” is becoming the norm, in a world where intelligent machines, systems, and networks are able to exchange information between them and respond independently, with autonomy to information, with the goal to manage industrial production processes. An important challenge is the transformation of traditional machines into intelligent machines, respectively intelligent or smart spindle. The purpose of this paper is to analyze the spindle using the intelligent models in in-situ conditions. The historical evolution, recent challenges and future trends of machine tool spindles were analyzed, noting that further development would be necessary to enable sensor/AI module integration to make the spindle unit an inherent quality assurance system. This study proposes a deep learning-based approach to spindle health monitoring based on multi sensor vibration signal analysis. The two proposed AI methods is based on the analysis of acceleration and synchronous envelope vibrations by demodulating the signal based on the Hilbert transform to identify critical bearing defects and specific defects at high frequencies.

Abstract: This study investigates the geometric parameters of commercially available or recently published models of catalyst substrates for passenger vehicles and provides a numerical evaluation of their influence on thermal behavior for use in urban areas. To ensure the plausibility of the results, a range of scenarios have been meticulously designed to replicate real operational conditions. To ensure a reliable basis for comparison, all geometries were standardized to the same gas-solid exchange surface by adapt the total monolith length. The simulation experiments conducted revealed that the primary role in question is played by the mass of the monolith and its internal surface area, while the heat transfer coefficient plays a secondary role. The necessity of adapting the geometry to the deployment scenario was demonstrated by comparing the performance of different scenarios. The employment of contemporary technologies, such as start-stop systems and automatic catalytic converter preconditioning, has also been factored. The study's findings suggest that geometries characterized by high cell density and low wall thickness, in conjunction with modern materials and intelligent engine control, are ready to become the norm for future ICE applications. This claim is particularly salient in the context of the requirements of mobile exhaust gas purification systems.

Abstract: The present paper suggests a Federated Learning-based Distributed Solar Forecasting model based on GRU networks (FL-GRU) to smart buildings in Muscat, Oman. The growing adoption of rooftop photovoltaic (PV) systems in urban settings needs precise, privatizing, and scalable forecasting models able to manage geographically dispersed and statistically heterogeneous data. The suggested solution will include federated learning and gated recurrent unit (GRU) networks to train a global forecasting model across several smart buildings and avoid the exchange of raw energy data to overcome these challenges. The local GRU models are trained on local PV generation data and only parameters of the model are relayed to a central aggregation server. This provides privacy of data without compromising effectiveness of collaborative learning. The proposed framework is tested in a variety of realistic scenarios such as scalability analysis, non-identically distributed (non-IID) data, client dropout, communication constraints, seasonal variability, and privacy saving noise injection. Simulation outcomes show that the proposed FL-GRU model presents a final RMSE of 0.129, MAE of 0.100 and forecasting accuracy of 97%. When increasing the number of clients involved in the process, 2 to 10, RMSE decreases to 0.129, which supports the high scalability advantages. In non-IID scenarios, RMSE ranges between 0.129 and 0.167, and even with half of the clients dropping, the system is robust with RMSE of 0.172. The proposed FL-GRU is better than the benchmark models, Local GRU, centralized GRU, FL-LSTM, and FL-ANN with a maximum improvement of 22.29% in RMSE reduction. Also, the best predictive consistency is found with correlation analysis with R² = 0.957. On the whole, the suggested approach can offer an efficient, privacy-aware, and scalable solution to distributed solar energy prediction in smart cities.

Abstract: This study presents a thermodynamic analysis of an ideal integrated adiabatic compressed air energy storage (A-CAES) system with single-temperature thermal energy storage (TES) subjected to an external heat boost. The additional heat is assumed to be supplied by concentrated solar energy, although other heat sources, such as combustion, can also be used, at least in principle. We examine three principal performance metrics of the integrated A-CAES system: the round-trip coefficient of performance (CoP), the marginal thermal CoP of the solar booster, and the second-law efficiency of the overall integrated system. The principal thermodynamic constraint is formulated as a theorem for the upper bound of the integrated 2nd law efficiency. The analysis shows that the system can attain a high round-trip CoP while simultaneously utilising the external solar heat boost very efficiently. These results highlight the thermodynamic potential of integrating solar heat boosting with A-CAES to improve both storage performance and renewable heat utilisation.

Abstract: The growing adoption of distributed renewable energy systems (DRES) calls for ad-vanced planning methodologies capable of addressing their inherent complexity and multi-dimensional trade-offs. Multi-Criteria Decision-Making (MCDM) frameworks are widely used to balance diverse objectives, but their effectiveness depends heavily on the selection of criteria, weighting techniques, and integration methods. This paper undertakes a systematic review of existing literature to analyse how MCDM ap-proaches have been applied in the planning and optimization of DRES projects. The review focuses on the criteria considered in MCDM, the techniques used to assign their relative importance, and the methods employed to integrate these weights into mul-ti-objective evaluations. The analysis draws from a diverse set of peer-reviewed pa-pers, examining economic, technical, environmental, and social dimensions, as well as the relationships between project-specific features and the criteria selection process. Results show that social criteria remain underrepresented both in terms of frequency and of relative importance in the evaluation process, while economic criteria are the most used and influential, underlining the need for more balanced, context-sensitive, and socially inclusive MCDM frameworks. Among MCDM methods and weighting methods, TOPSIS and AHP are by far the most common approaches, respectively. This review provides a foundation for future research aimed at improving the adaptability and effectiveness of MCDM frameworks in DRES.

Abstract: Metal binder jetting (MBJ) is an additive manufacturing (AM) process that offers advantages such as high speed, low cost, and low residual stress, compared to the prevalent fusion-based metal AM methods. However, a major barrier to MBJ is the low density of manufactured parts, which restricts part quality and limits its applications. One key process parameter that affects part density is the packing density of the powder bed. In general, a higher packing density is preferable in MBJ. Although research has been conducted to enhance the packing density ex-situ, most proposed approaches lack robustness when applied to real-world printing, where environmental variations and stochastic powder behavior introduce inconsistencies. An in-situ sensing method for packing density can mitigate these issues in several ways. It enables the implementation of feedback control strategies to regulate packing density during printing, contributes to comprehensive in-situ process monitoring, and provides quantitative data to support post-processing analysis and optimization. However, effective in-situ methods for accurately sensing packing density remain limited. To fill this research gap, two methods, namely ultrasound (acoustic) and recoating-force sensing, are proposed as potential approaches for in-situ sensing of powder packing density. Using a dedicated test platform, their responses to different powder bed packing densities are measured and compared. The results show a strong correlation between packing density and the sensor measurements, with differing levels of estimation confidence, demonstrating promising potential for their implementation as in-situ packing density sensors. Furthermore, the concept of sensor fusion is tested by combining the force-sensing and acoustic-sensing data, leading to improvements in the estimation confidence.

Abstract: This study presents a hybrid additive manufacturing approach to fabricate bioinspired stainless steel 316L-copper (SS316L-Cu) multimaterial structures using laser powder bed fusion (LPBF). The present study incorporates honeycomb lattice structures with varying wall thicknesses (0.25 mm, 0.5 mm, 0.75 mm, and 1.0 mm) to investigate the effect of geometric parameters on mechanical performance. Mechanical testing was conducted according to ISO 6892 standards, and the results revealed a strong dependence of tensile strength and ductility on lattice thickness. Copper (Cu) infiltration into SS316L lattice structures improved ductility by 30% compared to the monolithic SS316L lattice, with minimal compromise in tensile strength. To complement experimental results, molecular dynamics (MD) simulations were performed to study atomic-scale deformation and validate the trend of strength enhancement with increasing wall thickness. The findings demonstrate the potential of combining LPBF and liquid Cu infiltration to develop multifunctional, mechanically robust, and thermally conductive metallic composites. This approach provides valuable insight into structure–property relationships and supports the design of next-generation multifunctional composites for structural and thermal applications.

Abstract: The foundation of nuclear power plants is special as large-scale earth filling is often required. The properties of the backfill soil differ significantly from naturally deposited soils with regard to deformation and bearing capacity. For pile foundations, a thick backfill layer near the top may change the bearing mode around the pile. In this paper, parallel multi-pile tests were conducted to thoroughly investigate the vertical and horizontal bearing characteristics of long rock-socketed piles at backfill nuclear site. The results show that longer piles tend to have higher vertical bearing capacity, however, they do not necessarily exhibit smaller displacements when subjected to vertical load. When the applied vertical load increases, the effect of end resistance of piles becomes more pronounced, meanwhile, deeper soil layers exert a more significant impact on the axial forces within the pile body. Short piles with more part in backfill layer may endure a hoop tightening effect at the upper part of the pile, resulting in very little frictional resistance being provided by the lower soil. At the vertical load level preceding failure, the distribution of axial force and shaft resistance along the pile length will change. The typical mechanism of transition from static to dynamic friction between soil and piles that lead to shaft resistance is more apparent for longer piles but inhomogeneous soil like backfill layer will make the transition complex. When subjected to lateral loading, piles with better integrity show more pronounced elastic features, smaller maximum horizontal displacement and less residual horizontal displacement. The selection of the proportional coefficient for determining piles’ horizontal bearing capacity should correspond to the specific load and displacement. The results and in-depth analysis of the piles’ bearing capacity will provide intuitive experience for the analysis of pile foundations, thus offer valuable references for the design and construction in similar engineering.

Abstract: The proliferation of Unmanned Aerial Vehicles (UAVs) in various applications has created a pressing need for robust and efficient communication systems. Fifth generation (5G) networks, with their high bandwidth and low latency, are poised to support the massive connectivity requirements of UAVs. However, the high mobility of drones presents significant challenges for handover management, leading to frequent service interruptions and degraded performance. This paper proposes a novel, first-of-its-kind framework that integrates multi-UAV trajectory prediction with proactive handover optimization in 5G networks. Our approach utilizes a Long Short-Term Memory (LSTM)-based Recurrent Neural Network (RNN) to predict the future flight path of each UAV. The predicted trajectories are then fed into a Deep Reinforcement Learning (DRL) agent, which makes optimal handover decisions to ensure seamless connectivity and high Quality of Service (QoS). Unlike existing solutions that primarily rely on simulated data, our framework is validated using a real-world drone trajectory dataset. The experimental results demonstrate that our proposed method significantly outperforms traditional and existing machine learning-based handover schemes in terms of handover success rate, average Signal-to-Interference-plus-Noise Ratio (SINR), and handover delay. The proposed framework paves the way for more reliable and efficient drone operations in 5G and beyond networks.

Abstract: Retrofitting existing residential buildings is a critical strategy for achieving urban decarbonization while addressing public health disparities, particularly in communities disproportionately affected by environmental and socioeconomic stressors. This study presents a scalable urban building energy modeling framework that integrates physics-based simulations with machine learning to evaluate and prioritize health-driven retrofit strategies across residential building stocks. Synthetic datasets were generated through parametric simulations of representative building archetypes and retrofit scenarios, capturing variations in envelope performance, HVAC systems, infiltration rates, and ventilation strategies. Machine learning models were trained as surrogate predictors of building energy performance, enabling rapid evaluation of retrofit impacts. A range of algorithms—including decision trees, random decision forests, gradient boosting machines, support vector machines, k-nearest neighbors, and artificial neural networks—were evaluated. An artificial neural network implemented as a multilayer perceptron was selected for further analysis due to its strong predictive performance (R² = 0.94) and ability to capture complex nonlinear relationships among retrofit variables. The final model used the Port optimization algorithm for stable convergence and improved generalization. The framework is applied to Seattle’s Duwamish Valley, a community experiencing disproportionate environmental and health burdens. The results highlight retrofit priorities—particularly infiltration reduction, HVAC upgrades, and improved envelope performance—that deliver co-benefits for energy efficiency, indoor environmental quality, and occupant health. The results demonstrate that machine learning–enhanced physics-based UBEM can significantly accelerate retrofit evaluation while preserving the interpretability of simulation-based approaches. The proposed framework provides a scalable approach for identifying health-informed retrofit pathways that support equitable urban decarbonization.

Abstract: A computational model of anisotropic biomass particle pyrolysis was used to study the influence of particle properties and process conditions. The model couples multicomponent CRECK kinetics with intraparticle heat and mass transport. Particle size and lignocellulosic composition significantly affect conversion time and product yields; aspect ratio was also found to be important for larger-diameter particles. Larger particles (8 mm diameter, 4:1 aspect ratio) showed conversion times more than twice those of 3 mm particles, and char yield increased from about 16% to 23% when comparing small and large particles. Lignin-rich materials (e.g., palm shell) produced higher char and lower volatile yields than cellulose-rich biomass (wood, sugarcane bagasse); for 3 mm particles, char changed from 16% (oak) to 23% (palm shell). Higher reactor temperatures and heating rates substantially shortened particle conversion time—by up to 75%—and noticeably affected product yields. Analysis of the Biot and Pyrolysis numbers indicates millimeter-scale particles operate in a transition regime where internal conduction, external convection, and chemical kinetics occur on comparable timescales, so models must include these phenomena to accurately predict conversion times and final yields for reactor design and optimization.

Abstract: We investigate the problem of stable signal estimation under irregular observation conditions characterized by missing samples, heteroscedastic noise, and reportability constraints. In such environments, the primary engineering challenge arises from the severe curvature distortion of the objective landscape, which renders traditional Euclidean gradient methods numerically unstable. To address this, we propose a layered information-geometric framework for stable estimation on a distorted statistical manifold.The architecture consists of three integrated components: (i) a deterministic basin-safe front-end that resolves global navigation on the non-convex information landscape; (ii) a diagonal-Fisher local refiner that performs local metric normalization to correct for geometric scaling under irregular weighting; and (iii) a PT-even gatekeeper that acts as an engineering feasibility constraint by restricting the optimization trajectory to the reportable subspace. As a concrete structured path within this layered estimator, the QFRT branch—a Quaternionic Fourier–Ramanujan representation used here as an arithmetic-periodicity anchor under shared whitening and weighting—serves as a structured source of robustness under gappy observations without altering the bounded scope of the present technical claim.Across a locked stress matrix, the resulting hybrid estimator exhibits a two-layer gain structure. First, a general refinement gain in RMSE_ω is achieved via curvature-aware updates in the classical-tangent regime. Second, a specific PT-sensitive gain emerges when nuisance-coupled sectors become observable, effectively suppressing non-reportable "ghost-mode" leakage (|z| ≈ 0) where unprojected baselines suffer from substantial parameter drift. Mechanism diagnostics support a seed-path shielding interpretation: the classical front-end resolves the ω-dominant basin selection problem, shielding the downstream PT-aware refinement from unfavorable seed geometry. The resulting contribution is a technical methods framework for auditable stable estimation under missing and heteroscedastic observations.