Yu Wang

Rapid prediction of cavity sound pressure spectra is essential for the acoustic design of high-speed vehicles. A spectral database is built from 20 flow conditions (4 Mach numbers × 5 angles of attack) and eight cavity measurement points, giving 160 condition–point combinations, each densely sampled at 1147 amplitude–frequency points over 6–7000 Hz. Four models—RF, XGBoost, LightGBM and SVM—are compared on two prediction tasks, flow-condition and spatial-position prediction, before and after Bayesian optimization. Optimization markedly improves in-distribution accuracy and reduces dominant-frequency amplitude errors under unseen conditions, but reverses the ranking: RF and LightGBM dominate at baseline, whereas XGBoost ranks first after tuning, attaining 0.76 dB in the flow-condition task and staying on a par with RF (0.09 dB apart) in the spatial-position task. Under unseen conditions the SVM degenerates to a constant output of 117.11 dB, essentially the training-target mean, owing to the locality of the RBF kernel—a structural limitation that tuning cannot remedy. The direction of the peak-amplitude bias is not uniform: systematic underestimation occurs in the spatial-position task, whereas after tuning XGBoost overestimates the peak at Ma = 1.5, the condition with the highest measured amplitude. No single model is optimal across all scenarios.

Three-dimensional graph neural networks predict molecular and atomistic properties from interatomic geometry. Existing geometric backbones are compared mainly by replacing one architecture with another, so any accuracy gain is attributed to the new design as a whole. GeoRC augments an existing graph-level predictor with a geometry-aware residual branch while preserving its base prediction path. Radial distance features and element-wise edge gating build directed local interactions, gated recurrent unit updates refine the node states, and graph-level residual fusion combines the resulting candidate correction with the base prediction through a bounded, magnitude-conditioned screening map and a learnable global scale. The complete GeoRC-SphereNet model lowers MD17 force mean absolute error (MAE) for all eight evaluated molecules by 3.14% on average. In the repeated component-deletion study, removing the radial basis features increases mean force MAE for all eight molecules, which identifies the geometric encoding as the component with the broadest effect, and the OC20 macro energy-MAE reduction is 0.0639% across three paired runs. The results support GeoRC as an effective residual design for molecular and atomistic property prediction, and the proposed method provides a useful reference for future work in this area.

Grid operators scheduling wind power act on the width of the predicted distribution, and diffusion-based forecasters have reported strong accuracy on wind benchmarks. That accuracy can come from a sampling chain of a hundred or more evaluations per member, with reported implementations incurring multi-second latency per farm against a dispatch cycle of minutes, while a one-step student distilled from such a chain can concentrate its members into a spread narrower than its own error. This paper introduces Consist-Wind, a one-step probabilistic wind power forecaster. A frozen point backbone supplies the conditioning context; a teacher diffusion model on numerical weather prediction (NWP) residuals is distilled into a one-step consistency student, one function evaluation per member, and a stabilized shifted-logit warp maps generated members to the physical power range. A fourth phase calibrates the student by training its ensemble against the continuous ranked probability score (CRPS), which recovers the spread and raises the accuracy of the one-step forecast. Evaluated at every update, its pairwise form governs the training footprint, so the sorted-rank form of the CRPS is used, quasilinear in time and linear in memory. The cost of calibration is decoupled from the ensemble width by a scalable ensemble calibration principle, with the objective on order statistics and query-chunked attention (QCA) inside the kernel launch limit, so the whole pipeline trains and runs on a single device with limited compute. In our experiments on GEFCom2014-W, SDWPF, and GEFCom2012-W, Consist-Wind achieves the lowest CRPS, 3.1% below the strongest reproduced baseline, reduces the calibration error of the distilled student by 45%, issues a forecast in 0.03 ms for a speedup of about 66× relative to the reproduced multi-step reference, and reduces the memory required to evaluate the pairwise penalty term by 98%, admitting 500-member calibration ensembles on the evaluation device. Distributional forecasts of this accuracy, trained within a limited-compute budget, bring uncertainty-aware scheduling within reach of operators on commodity hardware.

Unsupervised industrial anomaly detection learns from normal images and identifies defects unseen during training. In reconstruction-based detectors, the features used to represent normality and the residuals used to localize defects vary across images, yet these signals are commonly selected with fixed category-level rules. We propose MSPNet, a Mirrored Selection Policy Network that learns these choices from validation performance. Its feature selection policy chooses channels that characterize normal appearance, and its residual selection policy chooses residuals for localization. The policies share a reward based on image-level and pixel-level AUROC while acting on complementary stages of the reconstruction pipeline, so the pair is symmetric in structure while the stages it acts on are not. Their scores are combined by SA-Fuzzy, a compact Takagi–Sugeno layer whose inputs are symmetric under exchange of the two branches, and whose membership functions are symmetric in branch agreement but asymmetric in anomaly level, so that increasing and decreasing anomaly signals receive different responses. On MVTec-AD, VisA, MPDD and BTAD, MSPNet obtains Image-AUROC/Pixel-AUROC values of 99.6/99.1, 98.2/99.0, 97.1/98.6 and 96.6/98.1, respectively. The learned policies are most useful for images whose selections depart from category averages, and SA-Fuzzy improves over a raw-score combiner on the images where the two branches disagree. On the full set, the two combiners are indistinguishable. Most of that margin comes from the exchange-invariant parameterization of the layer inputs. The present resolution does not separate fuzzy inference itself from a parameter-matched alternative. The learnable asymmetric memberships do separate from their symmetric counterpart on the same subset. These results show that input-adaptive selection can improve reconstruction-based anomaly detection across changing image appearances.