Bora Canbula

This retrospective offline study examines how dense manual regional annotations can be organized for image-level grapevine disease recognition using pre-acquired HERMOS RGB images. After audit and filtering, 491 source images and 13,765 Pascal VOC regions were retained. A leakage-safe source-image split was used across full-image classification, regional crop classification, crop-to-image aggregation, annotation-guided multiple-instance learning (MIL), and YOLO11 localization. MIL bags in training, validation, and testing were constructed from crops derived from manual boxes; the results therefore describe annotation-assisted aggregation rather than raw-image deployment. The original ConvNeXt-Tiny MIL model achieved macro-F1 0.922 (95% source-image bootstrap CI 0.882–0.956), compared with 0.836 (0.754–0.895) for full-image ConvNeXt-Tiny and 0.869 (0.804–0.917) for the best fixed crop aggregation. Paired macro-F1 differences favored MIL over these baselines by 0.087 (95% CI 0.029–0.163; Holm-adjusted p = 0.013) and 0.053 (0.005–0.114; adjusted p = 0.036), respectively. Downy mildew precision and recall were 1.000 on nine positive test images, but the corresponding exact 95% binomial intervals were 0.664–1.000. Controlled ablations produced lower point estimates without learned attention or the auxiliary instance loss, whereas removing context expansion did not reduce performance; none of the three paired ablation tests was significant after Holm correction. Increasing detector capacity from YOLO11n to YOLO11s did not improve full-image mAP@0.50 (0.175 versus 0.170). These findings support annotation-guided local-evidence aggregation as an offline exploratory benchmark, while not establishing split-to-split stability, computational efficiency, or external-vineyard generalization.

This study investigates whether compact and physically interpretable time-domain descriptors can support accurate low-concentration gas classification without relying on full multichannel waveforms. Using the Gas Sensor Array Low-Concentration dataset, each sample was transformed from raw sensor signals into 120 descriptors that summarize baseline behavior, variability, transient dynamics, and response magnitude. Four classical learning pipelines were evaluated through repeated stratified 5-fold cross-validation, supported by principal component analysis, descriptor ranking, sensor correlation analysis, and sensor subset experiments. The best-performing model, a radial support vector machine trained on the compact descriptor set, achieved a mean accuracy of 0.9456 and a macro-averaged F1 score of 0.9437 across 100 test folds while using 75 times fewer inputs than the raw waveform representation. An important finding is that classification performance improved systematically with concentration: accuracy was 0.9067 at 50 ppb, increased to 0.9333 at 100 ppb, and reached 0.9967 at 200 ppb. Response mean and baseline mean emerged as the most informative descriptor families, while VOCS-P, 2M012, and MQ-137 were the most discriminative sensors. Overall, the results show that compact and interpretable descriptors provide an efficient, reproducible, and practically useful benchmark for low-concentration gas classification in resource-constrained electronic nose systems.

Reliable theoretical reaction cross sections require a physically consistent description of the nuclear level density, especially in the low and intermediate excitation regions where collective nuclear motion can modify the number of accessible compound nucleus states. In this work, the influence of collective effects was examined for four representative reactions: ⁴⁷Ti(d,2p)⁴⁷Sc, ⁷⁴Ge(γ,2n)⁷²Ge, ¹¹¹Cd(p,n)¹¹¹In, and ¹⁶⁰Gd(n,γ)¹⁶¹Gd. The cross section curves calculated with the Collective Semi-Classical Fermi Gas Model (CSCFGM) were compared with experimental data compiled in EXFOR. The calculation outputs show that the CSCFGM based curves reproduce the principal features of the experimental excitation functions, including the threshold rise, the peak region and the subsequent decrease or saturation where competing channels become important. The strongest peak agreement is obtained for the ⁷⁴Ge(γ,2n)⁷²Ge and ¹¹¹Cd(p,n)¹¹¹In reactions, whereas the ⁴⁷Ti(d,2p)⁴⁷Sc and ¹⁶⁰Gd(n,γ)¹⁶¹Gd cases reveal the expected sensitivity to direct reaction contributions, optical model parameters and resonance like capture behavior. An additional sensitivity analysis for the ⁷⁴Ge(γ,2n)⁷²Ge reaction indicates that the adopted gamma ray strength function also affects the photonuclear peak magnitude. Overall, the results confirm that collective nuclear effects are essential for improving the reliability of model based nuclear data evaluations.

The Low-Strain Pile Integrity Test (LSPIT), standardized in ASTM D5882, is widely used as a rapid and economical non-destructive technique for assessing pile continuity in deep foundation systems. However, interpretation of LSPIT reflectograms remains strongly dependent on expert judgment and is influenced by soil–pile interaction effects such as damping and radiation losses, which can alter waveform morphology and confound automated defect screening. This study proposes a soil-aware deep learning framework that combines image-based reflectogram features with categorical geotechnical context describing the dominant soil regime at the measurement site. Reflectogram images are processed with a pretrained ConvNeXt-Large backbone, while soil information derived from Unified Soil Classification System (USCS) logs is represented as a categorical auxiliary input and mapped to a learnable embedding. The resulting multimodal design conditions waveform interpretation based on site context rather than relying on signal morphology alone. The framework is examined on an assembled benchmark of 510 expert-labeled reflectograms (404 intact and 106 defective), including a nine-site subset of 182 field records with explicit soil annotations. On the assembled benchmark, the model yields 99.41% accuracy and a weighted F1-score of 0.9941; on the nine-site subset, the observed accuracy is 99.45% with zero missed defective cases. Balanced accuracy, specificity, missed-detection rate, false-alarm rate, and confidence intervals are additionally reported to better align the evaluation with engineering screening practice. The study also states the current limits of the evidence base, including partial soil annotation, dominant-soil simplification, restricted soil coverage, and the absence of leave-site-out and interpretability-focused validation. Overall, the results support soil-aware multimodal learning as a promising proof-of-concept direction for more context-aware automated LSPIT interpretation, while also identifying the validation steps still required for broad field deployment.

The accurate determination of nuclear level density (NLD) is essential for a wide range of applications in nuclear science, including reactor design, nuclear astrophysics, and nuclear data evaluation. Traditional phenomenological models often face challenges in capturing key physical effects, such as collective excitations and shell structure, particularly in heavy and transitional nuclei, where the level density grows exponentially. Machine learning (ML) approaches have shown promise in improving predictive accuracy but are often limited by their purely data-driven nature, leading to challenges in interpretability and performance in regions with sparse experimental data. In this study, we propose a Physics-Informed Neural Network (PINN) framework, enhanced through multi-task learning (MTL), to address these limitations. The proposed model simultaneously predicts cumulative levels and mean resonance spacings by integrating experimental data with theoretical constraints, ensuring consistency with nuclear structure theory and robustness in extrapolating beyond the training data. Validation against both cumulative and yrast levels highlights the model’s ability to accurately capture rotational and vibrational excitations across a wide range of isotopes. Comparative evaluations demonstrate that the PINN model significantly outperforms traditional phenomenological models and purely data-driven approaches, offering a comprehensive and interpretable framework for advancing nuclear level density predictions and supporting practical applications in nuclear energy and astrophysics.

Pile foundations are critical structural elements, and their integrity is essential for ensuring the stability and safety of construction projects. Low-strain pile integrity testing (LSPIT) is widely used for defect detection; however, conventional manual interpretation of reflectograms is both time-consuming and susceptible to human error. This study presents a deep learning-driven approach utilizing transfer learning with convolutional neural networks (CNNs) to automate pile defect detection. A dataset of 328 reflectograms collected from real construction sites, including 246 intact and 82 defective samples, was used to train and evaluate the model. To address class imbalance, oversampling techniques were applied. Several state-of-the-art pretrained CNN architectures were compared, with ConvNeXtLarge achieving the highest accuracy of 98.2%. The accuracy reported was achieved on a dedicated test set using real reflectogram data from actual construction sites, distinguishing this study from prior work relying primarily on synthetic data. The proposed novelty includes adapting pre-trained CNN architectures specifically for real-world pile integrity testing, addressing practical challenges such as data imbalance and limited dataset size through targeted oversampling techniques. The proposed approach demonstrates significant improvements in accuracy and efficiency compared to manual interpretation methods, making it a promising solution for practical applications in the construction industry. The proposed method demonstrates potential for generalization across varying pile lengths and geological conditions, though further validation with broader datasets is recommended.