diffusion models

Few-Shot Learning of Unconditional Latent Diffusion Models Based on Domain Adaptation and Domain-Independent Latent Space

Closed-Form Demagnetization-Safety Criteria for Dy-Reduced Nd–Fe–B Permanent Magnets Under Nonuniform Reverse Fields

This article presents a compact analytical framework for evaluating demagnetization safety in Dy-reduced Nd–Fe–B permanent magnets under a representative nonuniform reverse-field loading condition. The formulation combines temperature-dependent intrinsic-property degradation, a microstructure-sensitive coercivity model, and a normalized reverse-field map to derive a closed-form safety criterion for zero hazard. The hazard-area ratio is shown to be the upper-tail measure of the normalized reverse-field distribution, which leads directly to an explicit safe/unsafe boundary in the parameter space of Dy reduction and grain-boundary-diffusion (GBD) reinforcement. The closed-form expressions are further obtained for the critical reinforcement level required to eliminate hazardous regions and for the critical reverse-field suppression factor that preserves full safety. The 2-D finite-element simulations on a current-free magnetostatic prototype confirm that the proposed closed-form boundary accurately separates safe and unsafe regimes and captures the hazard-area evolution with respect to Dy reduction and reinforcement strength. The proposed framework provides a concise theoretical basis for demagnetization-aware design of low-Dy permanent magnets under high-field operation.

Brain Controllability and Control Energy in Gray–White Matter Fusion Network

These findings suggest that incorporating gray-matter morphological information into a DTI-supported network provides a complementary structural representation for studying brain network controllability and state transitions.

Diffusion-transformer hybrid model for robust Taijiquan motion assessment

Kinetic insights into propylene epoxidation with <scp> H ₂ </scp> and <scp> O ₂ </scp>

Direct propylene epoxidation with H 2 and O 2 over Au/TS‐1 is an attractive alternative for the production of propylene oxide (PO), and acquiring its reaction kinetics is essential for the industrialization of this process. Herein, its intrinsic kinetics are established by using Au/TS‐1‐B with template‐blocked micropores, which enables the elimination of catalyst deactivation and micropore diffusion effects. The established reaction network shows that propanal and acetone are formed via PO isomerization and propylene‐derived hydro‐oxidation, and CO 2 mainly originates from PO deep oxidation on Au sites. The co‐feeding experiments reveal that there exist strong and weak adsorption for C 3 H 6 or PO on Au sites, and competitive adsorption of C 3 H 6 , PO, and H 2 O on Ti sites. Based on the above understanding, a Langmuir–Hinshelwood intrinsic kinetic model is proposed and validated. This model exhibits excellent applicability to well predict the data obtained from our kinetic experiments and those reported in the literature.

Neural Network Imputation of the Pitch‐Angle‐Resolved Medium‐Energy Electron Flux Data in LEO

Pitch‐angle‐resolved electron flux data are crucial for investigating variations in electron flux within the magnetosphere and the underlying mechanisms, such as radial diffusion and wave‐particle interactions. The Medium‐Energy Electron Detector (MEED) onboard the Fengyun‐3E (FY‐3E) satellite in Low Earth Orbit (LEO) provides 18‐directional local pitch angle observations, enabling studies of medium‐energy electrons in LEO. However, the MEED cannot cover the full local pitch angle range continuously from 0° to 180° due to limitations of the satellite's three‐axis stabilized attitude control system. At mid‐latitudes, its pitch angle coverage spans approximately 100° (40°–140°), with improved coverage at low and high latitudes. To extend MEED's coverage toward full global pitch angle observations, we propose a data imputation method using the MEED data and machine learning technology. We have trained eight imputation models based on Multi‐Layer Perceptron (MLP) using electron flux data near 90° local pitch angle and satellite orbital data to generate missing data near 0° and 180° with 280–600 keV energies. For unseen data in test sets, the correlation coefficient r between the models' reconstructions and observations is at least 0.915, with a maximum Root Mean Square Error (RMSE) of 0.110 on the logarithmic scale, which indicates that the models can provide reasonably reliable estimates for missing values. Besides, these imputation models may support future near‐real‐time electron flux imputation when operational data streams become available. This work helps expand the mid‐latitude pitch angle coverage of the FY‐3E MEED, improving data usability, and aiding the development of electron flux prediction models in LEO.

Computational modeling of turbulent nanofluid heat transfer over a heated moving surface with local thermal non-equilibrium and machine-learned eddy viscosity

This study presents an adaptive modified Runge-Kutta compact scheme for the numerical simulation of unsteady k − ω turbulent nanofluid flow over a heated moving surface under local thermal non-equilibrium conditions. The surface-interfacial model incorporates mixed convection, viscous dissipation, turbulence transport, and separate energy equations for the base fluid and nanoparticle phases, with the effective thermal conductivity described by Xue’s formulation. The proposed time-integration method is explicit and combined with a compact finite-difference discretization that provides fourth-order spatial accuracy. The temporal coefficients are selected to achieve second-order accuracy, and the method is further enhanced through adaptive time stepping based on local error control. Stability analysis for the scalar convection–diffusion problem and conditional convergence analysis for the corresponding system formulation are also established. Numerical comparisons show that the proposed adaptive scheme yields lower error than existing adaptive Euler- and Runge-Kutta-based schemes. The computed results further demonstrate that thermal buoyancy increases the mean velocity, whereas larger Prandtl numbers reduce the thermal boundary-layer thickness of the fluid and nanoparticle phases. In addition, stronger interphase coupling modifies the two-temperature fields in a manner consistent with local thermal nonequilibrium. A machine-learning model is also employed to predict eddy viscosity, and its reliability is confirmed through profile comparisons, contour analyses, sensitivity assessments, and Taylor diagram evaluations. Overall, the proposed framework provides an accurate and efficient computational tool for surface-associated turbulent nanofluid transport with interfacial thermal nonequilibrium.

D-galactose enhances Maillard reaction and antioxidant activity of arginine-derived products: experimental and molecular insights

Arginine-derived Maillard reaction products (MRPs) have shown promising antioxidant activities. However, how structurally distinct hexose reducing sugars, such as D-fructose and D-galactose, affect the formation and functional properties of L-arginine-derived MRPs remains unclear. Arginine-fructose (Arg-Fru) and arginine-galactose (Arg-Gal) model MRPs were prepared by heating at 100 °C for 0.5, 1, and 2 h. The reaction rate was analyzed by monitoring the reaction process and using a kinetic model, while substrate consumption and apparent melanoidin production were measured. Early non-covalent interactions were further investigated using molecular docking and 200 ns molecular dynamics (MD) simulations. Finally, functional differences were evaluated by chemical and intracellular antioxidant activities. Compared with Arg-Fru, Arg-Gal showed a more great decrease in pH and higher A294 and A420 values under the same reaction conditions. Kinetic analysis showed that Arg-Gal had higher reaction rate constants. Meanwhile, Arg-Gal exhibited stronger UV-Vis absorption and fluorescence intensity, greater consumption of free amino groups and reducing sugars, and higher apparent melanoidin formation. Molecular docking results showed that D-galactose formed more hydrogen bonds with L-arginine than D-fructose; MD simulations further showed that the Arg-Gal system formed approximately 20 more hydrogen bonds on average, and the diffusion behavior of D-galactose was closer to that of L-arginine, suggesting stronger non-covalent association. Functionally, Arg-Gal showed stronger DPPH radical scavenging capacity, reducing power, and intracellular antioxidant activity than Arg-Fru. D-galactose may promote the Maillard reaction process by enhancing early non-covalent interactions with L-arginine and being associated with enhanced formation of intermediates and melanoidin-related substances, thereby enhancing the antioxidant activity of MRPs. These findings provide experimental and molecular insights into the role of sugar structure in modulating arginine-derived Maillard reaction systems.

Air-Dispersion-Model-Based Identification and Sparse Regression Inversion of Radon Sources in Uranium-Mine Roadways

Source identification in confined underground ventilation systems is essential for hazardous-gas monitoring, and uranium-mine radon provides a representative case in which release locations and strengths must be inferred from limited concentration measurements. This presents an underdetermined, ill-posed inverse problem whose solvability under different sparse-regression strategies and roadway configurations remains poorly understood. In this study, a computational fluid dynamics (CFD) forward model is coupled with sparse regression. The ventilation flow field and radon advection–diffusion process are solved in OpenFOAM to construct a source–sensor contribution matrix, and source recovery is formulated as a sparse linear inverse problem. Four methods—LASSO, LASSO with non-negative least-squares (NNLS) refitting, Elastic Net, and Elastic Net with NNLS refitting—are compared, and the contribution matrix is characterized by its mutual coherence, condition number, and singular-value spectrum. Numerical tests were conducted for single- and multiple-source scenarios in single-main and main–branch roadway models. The results indicate that inversion performance depends on the spatial information and local identifiability provided by the sensor configuration rather than on sensor number alone. LASSO and Elastic Net exhibited varying degrees of source-strength shrinkage or dispersion, whereas NNLS refitting reduced these effects when the first-stage support contained the dominant source candidates. In the prescribed three-source case, denser sensor coverage improved dominant-source localization and reduced the post hoc condition number of the prescribed-source submatrix, although the full-matrix condition number increased. This finding indicates improved local identifiability for the tested source combination rather than a general sensor-count effect. Because the synthetic observations and the inversion operator were derived from the same CFD response matrix, the results represent a controlled model-consistent proof of concept rather than an estimate of field-level performance.

Internal heating in rapidly rotating convection is not a shortcut to geostrophic turbulence

Convective turbulence in planets and stars is often driven by internal heating. This forcing mechanism has also been proposed as a means of accessing the diffusion-free scalings of the geostrophic turbulence (GT) regime at modest forcing. We test this with the asymptotically reduced quasi-geostrophic model, which is formally valid in the limit of vanishing Ekman number, $Ek \rightarrow 0$, and retains the Prandtl number as an independent parameter. We find no such shortcut: the Nusselt and Reynolds numbers are no closer to their diffusion-free predictions than in the boundary-heated case, and at $Pr = 7$ they are further from them; the predicted $Pr^{-1/2}$ collapse fails; and the prefactor depends on the form of the heating. Our no-slip/stress-free cases track the radiatively driven simulations of Hadjerci et al. (2024) case by case, yet a $25\%$ variation in $Nu$ persists between sweeps that share the same reduced Rayleigh and Prandtl numbers but differ in Ekman number. Those data span only the narrow range of forcing over which the compensated Nusselt number is stationary; across the wider range accessible to the reduced model it rises to a maximum and then falls. We argue that bulk transport scalings are incomplete diagnostics of GT, whereas the saturation of the interior mean temperature gradient and of the vertical velocity kurtosis remain reliable indicators.

Membership is Ownership: A Robust Ownership Verification Framework for Diffusion Models

This work investigates IP protection (i.e., model ownership verification) for diffusion models in a realistic commercial scenario with minimal model utility loss, and builds a framework for model ownership verification, termed ``{Membership is Ownership} (MiO)'', based on a population-level hypothesis test on a private member evidence dataset.

Deep generative modeling for AI-guided inverse design of perovskite photovoltaic devices

This work presents an end-to-end AI-guided inverse-design framework that learns the conditional distribution of device parameters given target photovoltaic figures of merit and offers a transferable, reproducible, and statistically rigorous methodology for accelerating design of next-generation perovskite PV devices.