Photovoltaic–thermal (PVT) systems suffer from reduced electrical efficiency due to heat accumulation on PV surfaces. This study numerically investigates anovelstaircase-configured PVT air system with passive, perforated V-shaped fins—a geometric combination not previously reported in the PVT literature. This design disrupts thermal boundary layers and induces micro‑turbulence, enhancing convective cooling. Using validated 3D CFD simulations, the proposed system achieves a peak thermal efficiency of 31.0% (vs. 29.4% for a conventional design) and an electrical efficiency of 14.8% (vs. 14.6%), while reducing PV surface temperature by 9 K. Peak outputs reach approximately 400 W (thermal) and 180 W (electrical) per panel. Sensitivity analysis shows that a 7 mm perforation diameter and 0.04 kg/s airflow optimize performance, yielding 32.5% thermal efficiency and a 10 K temperature drop. An integrated energy‑economic‑environmental (3E) assessment estimates a 5.8‑year payback period (clear‑sky), 80 tons CO2 mitigation over 25 years, and an energy payback time of 0.86 years. The passive, modular design is scalable from residential to commercial applications, offering a cost‑effective and sustainable solution for arid, high‑irradiance regions.
Cellular morphology and tissue-level architecture must be captured for automated breast cancer grade classification from histopathological images, many current single-resolution CNN–Transformer models fall short in this regard. A multi-magnification pyramid encoder with cross-scale flexible attention, a graph neural network branch modeling inter-nuclear interactions, stain-aware contrastive pretraining, and an uncertainty-gated adaptive fusion mechanism are all integrated into our proposed hybrid framework, HM-GraphSwin. The model combines multi-scale and graph-based features for better biological interpretability by utilising morphologically conditioned cross-attention. The proposed model is evaluated on the BreakHis benchmark dataset, demonstrating superior performance over existing state-of-the-art methods. The performance accuracies of 99.32 ± 0.28% (40×), 98.41 ± 0.37% (100×), 97.56 ± 0.52% (200×), and 96.71 ± 0.64% (400×) for cancer classification are observed. These results demonstrate superior performance over existing methods while providing reliable uncertainty estimates and interpretable visualizations suitable for clinical deployment.
Quantum processors encode computational information in relative phases and are therefore sensitive to spatially and temporally correlated dephasing. We formulate gravitational redshift as a physically constrained template for such noise. A fixed, known redshift produces a deterministic phase, whereas dephasing arises only when unresolved displacement, timing, or frequency fluctuations are transduced into phase uncertainty; for vertical motion, the conversion factor is βg=ω0g/c02. The gravity-specific content is thus the mapping from geometry fluctuations to phase covariance and frequency-noise spectra, while the resulting Gaussian dephasing dynamics are generic. We test spatial-kernel dependence using exponential, Gaussian, Matérn, and power-law correlations. Circuit-level effects are evaluated with an explicit GHZ-preparation circuit, showing that a full-GHZ-per-layer exponential approximation overestimates loss. For temporal correlations, we derive Sδω=βg2Sh and use a mechanically motivated drift-plus-vibration model, which attributes memory to the mechanical process rather than gravity. Platform estimates indicate that the gravity-transduced contribution is negligible compared with present technical dephasing. The framework therefore provides a falsifiable geometry-to-noise model while keeping mitigation and compilation claims quantitatively cautious.
Traffic flow forecasting requires stable and accurate multi-step predictions under non-stationary spatiotemporal coupling and noise perturbations. However, existing methods often lack explicit modeling of latent temporal state evolution, which can lead to error accumulation over long prediction horizons. Moreover, spatial dependencies are commonly generated dynamically for individual samples, making it difficult to retain reusable prototypical interaction patterns and thereby limiting model generalization and interpretability. To address these issues, a GRU-SSM and Pattern Bank Enhanced Linear Attention (GSPB-LA) model is proposed for traffic flow forecasting. First, GRU-SSM recurrence is residually coupled with linear attention along the temporal dimension to strengthen the representation of long-range dynamic dependencies. Second, a global pattern bank is constructed to perform node-level pattern retrieval in the spatial dimension, thereby improving the modeling of complex spatial dependencies. Finally, a gating mechanism is introduced into the temporal embedding layer to integrate time-of-day (TOD) information with Fourier-based periodic priors, yielding a more robust representation of temporal periodicity. The proposed GSPB-LA model is evaluated on multiple real-world traffic flow datasets. Experimental results show that GSPB-LA outperforms recent representative baseline models and exhibits stronger stability and generalization across multiple forecasting horizons.
Dual-band millimeter-wave (mmWave) heterogeneous networks (HetNets) operating at 28 and 73 GHz offer high capacity, but interference-coupled SINR, subchannel reuse, and per-base-station power constraints make transmit-power allocation nonconvex. This paper proposes Weight-Driven Particle Swarm Optimization (WD-PSO), which uses a recursively decaying weight to jointly regulate particle momentum and attraction strength. WD-PSO is evaluated against five PSO variants, Equal Power Allocation, Random Feasible Search, and Differential Evolution (DE) over 30 paired Monte Carlo realizations. Results show that WD-PSO achieves final throughput comparable to the strongest PSO variants while consistently outperforming the additional baselines. Across K=5 and K=10, WD-PSO improves throughput over DE by 4.48–6.10%. Its mean throughput varies by less than 0.8% across a 5000-fold range of the initialization coefficient α. Scalability analysis further shows near-linear runtime growth with population size and iteration count, supporting a practical balance between solution quality, robustness, and computational cost.
