Signed RC–Schmitt oscillator arrays can realize bipartite (anti-phase two-cluster) timing patterns useful for differential signaling and quadrature clock generation, but existing event-triggered and reinforcement-learning designs for such arrays either ignore non-memoryless operating-mode changes or leave data-driven threshold tuning unconstrained, risking loss of lock. This paper closes both gaps with a single framework: a phase-reduction and structural-balance gauge argument reduces the signed, mode-switching circuit array to an unsigned Kuramoto–Sakaguchi network amenable to Lyapunov analysis; hidden semi-Markov switching, with general holding-time distributions and imperfect mode detection, models load, supply, and comparator-delay variation; and a memory-aware event-triggered rule is combined with Q-learning threshold adaptation that is provably confined to a certified admissible interval derived from the analytical locking margin. Three theorems establish practical mean frequency locking under arbitrary hidden semi-Markov switching, asymptotic bipartite phase locking for identical oscillators, and preservation of the certified locking margin throughout all Q-learning iterations, with a pinned bipartite tracking corollary for leader-following arrays. Six-node MATLAB simulations confirm chatter-free circuit operation, bipartite locking under three hidden semi-Markov modes, and a 60.7% reduction in triggering events relative to periodic coupling without violating the certified safety margin.
The Fourth Industrial Revolution has increased the demand for intelligent wireless systems with real-time sensing and high-speed communication. This growth necessitates compact frequency-selective surfaces (FSSs) for microwave sensing and high-frequency communication applications, yet current designs often lack robustness and material adaptability. This work introduces compact, symmetric, and polarization-independent FSS for communication and sensing across Ku-, K-, and Ka bands. The design exhibits single-negative (SNG) properties with an electrical size of 0.236λ₀ × 0.236λ₀. It blocks transmission below −10 dB across three bands at 14.13, 23.15, and 36.76 GHz, demonstrating polarization independence with angular stability up to 60°. The FSS exhibits high Q-factors of 235.5, 211.32, and 369.16, with good sensitivity of 0.2-0.4 GHz/ε for both solid materials and bioplastics. The proposed FSS exhibits a stable linear frequency shift (R2) when sensing various bioplastics and solid materials. Measurements and experiments validate its effectiveness for next-generation microwave sensing and communication systems.
Research on high-dimensional fractional-order multi-wing hyperchaotic systems remains limited. This paper proposes a novel 6D fractional-order hyperchaotic system that incorporates a memristor and nonlinearities and possesses an infinite number of equilibria. First, the system is solved numerically using the Adomian decomposition method. Next, phase portraits, Lyapunov exponents, bifurcation diagrams, and basins of attraction are presented. The results demonstrate that the system exhibits chaotic behavior over a wide parameter range. Varying the control parameters alters the attractor topology, thereby controlling the number of wings and driving morphological evolution. Moreover, different initial conditions yield attractors with distinct shapes. Compared with previously reported systems, the proposed system achieves a lower fractional order q and larger Lyapunov exponents. Spectral entropy analysis reveals high dynamical complexity. Finally, NIST statistical tests and DSP implementation validate the excellent pseudo-randomness of the system, confirming the hardware feasibility.
Discovering governing equations from data is essential for interpretable modeling of physical systems, but the required derivative estimation remains highly sensitive to observational noise. To address this issue, we propose a robust symbolic PDE discovery framework based on a context-aware attention neural field and uncertainty-weighted symbolic regression. The context-aware attention neural field constructs a differentiable surrogate of the underlying solution by adaptively integrating neighboring spatiotemporal observations instead of relying on fixed numerical differentiation stencils. Temporal and spatial derivatives are then obtained by automatic differentiation, while an independently trained ensemble provides pointwise derivative uncertainty estimates. These uncertainty estimates are incorporated into the symbolic regression objective as reliability weights, suppressing the contribution of derivative samples with high estimated uncertainty during the equation search. Experiments on six PDE benchmarks under Gaussian noise and data subsampling demonstrate robust equation and coefficient recovery with reduced computational cost compared with representative baselines.
We construct an integrable semidiscretization of the principal chiral field equation with a self-consistent potential, replacing space by a one-dimensional lattice while keeping time continuous. A Lax pair is proposed whose compatibility yields the semidiscrete system and an exact discrete conservation law. Using the associated Darboux transformation, we derive one-fold, two-fold and N-fold transformations in closed determinant form and obtain explicit multi-soliton solutions. The spin field remains on the unit sphere, while the potential forms localized travelling wells. A coalescence limit of spectral parameters produces higher-order smooth positons; the second-order case exhibits double-loop spin motion, curved space-time trajectories, and logarithmically slow separation of its components. Closed expressions are given for the world lines and the velocities of the two constituents. The solutions preserve the spin and potential constraints and reduce, in the continuum limit, to those of the continuous principal chiral field model. This provides a structure-preserving discrete framework for degenerate nonlinear excitations in integrable spin systems.
Multichannel metasurface holography requires accurate control of wavelength- and polarization-dependent responses, but deep-learning design is hindered by inter-channel domain shifts and repeated database construction. Here, we propose a transfer-learning-assisted framework for rapid, data-efficient design of multi-wavelength and multi-polarization metasurface holograms. A fully connected forward neural network is first trained on a source channel to learn the mapping between anisotropic TiO₂ meta-atom geometries and complex transmission responses. A freeze-and-fine-tune strategy then transfers the learned geometry-dependent features to target polarization and wavelength channels, enabling accurate cross-channel response prediction with limited target-domain data. An adaptive weighted Gerchberg-Saxton algorithm is further introduced to jointly optimize target-channel phase matching and non-target-channel response suppression. The framework is validated by six-channel polarization-multiplexed holography at 540 nm and twelve-channel wavelength-polarization multiplexed holography at 450, 540, and 630 nm. Numerical results show clear reconstructions, reduced crosstalk, while the proposed transfer-learning workflow reduces the cumulative workflow time by approximately 82.6% relative to direct full-wave database construction for five target channels.
