Search Results (2,140)

Aiming at the problems of the severe active–reactive power coupling, insufficient adaptive inertia–damping regulation, and degraded dynamic performance of virtual synchronous generators (VSGs) under the operating conditions of a weak grid, high resistance-to-reactance ratio, and large power angle, this paper proposes a cooperative control strategy that combines reactive power feedforward decoupling with adaptive damping–inertia regulation. First, a small-signal power model of the VSG is established, and a dynamic relative gain array is employed to quantitatively analyze the effects of the resistance-to-reactance ratio and power angle on power coupling characteristics, revealing that large power angles and high resistance-to-reactance ratios significantly aggravate active–reactive power coupling. Based on this analysis, a reactive-power-oriented feedforward decoupling strategy is designed to suppress the cross-coupling between reactive power and power angle while preserving the intrinsic inertia support characteristics of the active power loop. Eigenvalue migration analysis further demonstrates that the proposed reactive-power-oriented decoupling provides higher damping ratios and larger stability margins than conventional full active–reactive power decoupling. Furthermore, a deep deterministic policy gradient-based adaptive damping–inertia control method is developed by incorporating frequency deviation, power fluctuation, voltage deviation, and coupling degree into the state space, enabling the online coordinated optimization of virtual inertia and damping coefficients. The hardware-in-the-loop experimental results verify that the proposed strategy effectively suppresses active–reactive power coupling, reduces power overshoot and oscillation, enhances frequency support capability and dynamic response speed, and maintains superior stability under weak grid conditions. Sensitivity analysis under grid impedance estimation errors further confirms its strong robustness against parameter uncertainty, while tests under composite disturbance scenarios demonstrate excellent transient performance. The proposed strategy provides an effective solution for improving the grid-connected operation performance and adaptability of VSGs in low-inertia power systems. Full article

Streamflow variations in the Middle Minjiang River Basin (MMR) are vital for the flood mitigation and water resources management of the Chengdu metropolitan area which is important for the development of Southwest China. However, how climate change, Chengdu metropolitan area and Zipingpu Reservoir influence streamflow in the MMR remains unclear. Hence, we coupled the Geomorphology-Based Ecohydrological Model (GBEHM), the Physic-aware Hybrid Learning (PaHL) model and the Extreme Gradient Boosting (XGBoost) model to reproduce streamflow variations at Pengshan station—the outlet cross section of MMR—from 1980 to 2019, subsequently performing attribution analysis. Annual streamflow at Pengshan station exhibits a decreasing trend from 1980 to 2019. Coupled simulations effectively reproduce daily streamflow at Pengshan station during 35 years, with values of NSE, R² and KGE exceeding 0.96. The dominant influence of anthropogenic disturbance on daily streamflow decrease is generally steady at Pengshan station, explaining 62.3% and 430.8% of it before and after the impoundment of Zipingpu Reservoir (in 2006), respectively. Majority of the climate change’s influence is notably concentrated from June to September, suggesting a potential temporal imbalance in water resources and a threat of extreme hydrological events. Our study contributes to flood mitigation and water resources management in the MMR. Full article

Background/Objectives: Heart failure (HF) classification based on left ventricular ejection fraction (LVEF) provides an incomplete representation of disease complexity, as it does not fully integrate functional impairment, structural remodeling, and clinical severity within a unified framework. Although N-terminal pro-B-type natriuretic peptide (NT-proBNP) is widely used for diagnosis and risk stratification, prior studies have primarily evaluated its role in isolation or within individual HF phenotypes, leaving its phenotype-specific distribution and integrative capacity across the HF spectrum insufficiently defined. This study aimed to address this gap by systematically evaluating NT-proBNP across HF phenotypes and assessing its potential as an integrative biomarker linking ventricular dysfunction, structural remodeling, and clinical severity. Methods: A cross-sectional study was conducted including 125 participants, comprising 65 clinically stable HF patients and 60 age- and sex-matched controls. HF patients were stratified according to LVEF into HF with reduced EF (HFrEF) (n = 28), (HFmrEF) (n = 20), and HF with preserved EF (HFpEF) (n = 17). Serum NT-proBNP concentrations were measured using a standardized electrochemiluminescence immunoassay. Clinical and echocardiographic parameters, including LVEF, left ventricular end diastolic diameter (LVEDD), left atrial diameter (LAD), and New York Heart Association (NYHA) functional class, were recorded and analyzed. Results: NT-proBNP levels were significantly higher in HF patients compared with controls (1845 ± 620 vs. 95.7 ± 40.5 pg/mL; p < 0.001) and demonstrated a clear stepwise increase across phenotypes (HFrEF: 2850.6 ± 710.4; HFmrEF: 1620.8 ± 480.2; HFpEF: 920.9 ± 310.3 pg/mL; p < 0.001). NT-proBNP showed a strong inverse correlation with LVEF (r = −0.68, p < 0.001) and significant positive correlations with LVEDD (r = 0.61, p < 0.001) and LAD (r = 0.57, p < 0.001). Higher levels were associated with more advanced NYHA functional class (III–IV vs. II: 2510 ± 680 vs. 980 ± 340 pg/mL; p < 0.001). ROC analysis demonstrated robust discriminatory performance across HF phenotypes, with the highest accuracy observed in HFrEF. Conclusions: NT-proBNP exhibits a phenotype-dependent gradient and consistently reflects ventricular dysfunction, adverse structural remodeling, and clinical severity across the HF spectrum. These findings support its role as an integrative biomarker that captures the multidimensional nature of HF, with potential implications for phenotype-based risk stratification and more precise clinical decision making. Full article

Background: Cervicogenic headache (CEH) is a secondary headache disorder originating from dysfunction in the cervical spine. In addition to pain, individuals with CEH frequently experience disturbances in postural control and sensorimotor integration, which may compromise functional capacity and quality of life. Conventional clinical assessments typically focus on pain intensity and cervical range of motion; however, these measures often fail to capture the multifactorial mechanisms underlying balance impairments in this population. Machine learning (ML) methods offer the ability to integrate multidimensional clinical data and may provide a more comprehensive approach for identifying patterns of postural stability and the factors influencing balance regulation in CEH. Methods: A secondary analysis was conducted using baseline data pooled from three registered randomized controlled trials, comprising 68 independent participants diagnosed by a neurologist according to the International Classification of Headache Disorders, 3rd edition (ICHD-3). Postural Stability Class served as the primary outcome and was derived from quantitative stability scores categorized as High, Moderate, or Low. Predictor variables included demographic characteristics (age, gender), clinical measures (pain intensity, headache frequency, symptom duration, cervical range of motion), and sensorimotor parameters (center-of-pressure sway and gaze accuracy). Five machine learning algorithms—Random Forest, XGBoost, Support Vector Machine, Logistic Regression, and Gradient Boosting—were trained and evaluated using 10-fold cross-validation with procedures implemented to reduce overfitting. Results: The Gradient Boosting classifier demonstrated the best performance, achieving an accuracy of 0.857 and an F1 score of 0.857, with a cross-validated accuracy of 0.802 ± 0.063. Random Forest and XGBoost achieved accuracies of 0.786. Feature importance analysis identified center-of-pressure sway and pain intensity as the most influential predictors of stability classification, followed by cervical flexion range of motion and gaze accuracy. Demographic variables showed minimal contribution to model performance. Conclusions: Machine learning models were able to distinguish different levels of postural stability in individuals with CEH. The findings highlight the central role of pain and sensorimotor control in balance regulation and suggest that predictive analytics may support precision physiotherapy by enabling rehabilitation strategies tailored to individual sensorimotor profiles. Full article

Post-harvest losses of mango (Mangifera indica L.) in developing economies are estimated at 5% to 30%, largely due to manual management practices that depend on subjective visual assessments. This paper proposes a lightweight deep learning architecture, termed SE-DBIRNet, for real-time classification of mangoes into five shelf-life stages: unripe, semi-ripe, fully ripe, overripe, and perished. The model incorporates three key design strategies: (i) depthwise separable convolutions, achieving an 88.5% reduction in parameters when integrated into the ResNet50 backbone; (ii) a double-branch inverted residual (DBIR) module designed to enhance feature diversity and richness; and (iii) a squeeze-and-excitation (SE) attention mechanism for adaptive channel-wise recalibration. Using a public benchmark dataset of 4428 RGB images (Mendeley Data) under 10-fold cross-validation, SE-DBIRNet achieved 98.24% accuracy. Among lightweight CNN architectures (EfficientNetB0, MobileNetV2, ResNet50), SE-DBIRNet outperformed the best lightweight baseline (EfficientNetB0: 96.57%) by 1.67 percentage points. While dedicated attention-based DenseNet variants (e.g., DSA-DenseNet: 99.20%) achieved higher accuracy, SE-DBIRNet offers a superior trade-off among accuracy, inference speed (56.9 ± 1.8 FPS), and memory efficiency (8871 ± 45 MB CPU memory). EigenCAM activation visualizations revealed that the model focuses on biologically relevant and stage-discriminative features, including surface color gradients, texture uniformity, lenticel patterns, and decay boundaries. Overall, SE-DBIRNet achieves a Pareto-optimal balance among accuracy, speed, and memory efficiency, making it a strong candidate for real-time, edge-deployable post-harvest mango quality-monitoring systems, particularly when computational resources are limited. Full article

This study examines the retail food environment surrounding public schools in Michigan using a multiscale, multidimensional framework. A cross-sectional spatial analysis integrates relative healthfulness (modified Retail Food Environment Index, mRFEI), availability (outlet counts), and accessibility (network-based walking time) across school districts, census tracts, block groups, and school-centered buffers. The analysis includes 3530 public schools, 7680 fast food restaurants, and 2065 convenience stores. Results show pronounced spatial heterogeneity and clustering of unhealthful outlets (Nearest Neighbor Index = 0.284, p < 0.001), with many located near schools. Approximately 34% of schools are within a 10 min walk of a fast food restaurant, increasing to 65% within a 20 min walk. Urban schools face significantly greater exposure—2.27–2.80 times more fast food outlets and shorter walking times than rural schools (p ≤ 0.002)—with consistent gradients across city, suburban, town, and rural contexts. Overall, school neighborhood food environments are highly structured, obesogenic, and inequitable. By integrating multiple spatial scales and complementary measures of food environments, this study advances food environment research and provides policy-relevant evidence for targeted, place-based interventions to improve access to healthier food around schools. Full article

Cost overrun remains a pervasive issue in building construction projects, yet most predictive studies operate within a single data domain, ignoring the systemic interactions across project, schedule, resource, quality, and safety dimensions. This study quantifies the incremental predictive value of integrating these five construction data domains and identifies the cross-domain interaction patterns that explain prediction accuracy. As a simulation-based methodological study, an integrated dataset of 100,000 records was synthesised with theory-grounded causal structures derived from the construction management literature; no real project data were used. Gradient Boosting (GB), Random Forest (RF), and Linear Regression were evaluated on an 80/20 hold-out test split, with robustness verified through alternative domain orderings and hyperparameter sensitivity. SHAP analysis, including exact interaction values, was used to interpret feature importance and cross-domain synergies. The full five-domain GB model achieved R² ≈ 0.97 and MAPE ≈ 6%, a 220% relative R² improvement over the Project-domain baseline (R² rising from 0.305 to 0.975), robust across three ordering schemes. Schedule and Quality contributed the largest marginal gains (ΔR² = +0.312 and +0.255), whereas Resource integration yielded approximately one-thirty-first of Schedule’s return. Because the dataset is synthetic, the results are interpreted as a methodological demonstration rather than empirical evidence from real projects; they provide a reusable framework for prioritising data-integration investment and show that, within the simulated causal structure, cross-domain interactions—particularly Schedule × Risk and Project Type × Change Cost—carry predictive information that single-domain analyses cannot recover. Validation on real, partially integrated datasets is identified as essential future work. Full article

Background/Objectives: Adult epilepsy is clinically heterogeneous, and individual clinical predictors may not fully capture the multidimensional burden associated with drug-resistant epilepsy (DRE). This study aimed to identify latent clinical phenotypes in adults with epilepsy and examine their cross-sectional associations with DRE and broader disease burden. Methods: This regional observational cohort study used a source database of 1100 patients with epilepsy. After excluding two patients aged <18 years, the adult analytic cohort included 1098 patients. Complete-case latent class analysis (LCA) was performed in 1054 patients using age at onset, disease duration, seizure type, seizure frequency, serial seizures/status, postictal confusion, neurological status, neuroimaging category, and number of antiseizure medications. Model selection was based on statistical fit, class size, and clinical interpretability. Internal clinical validation outcomes included DRE, quality of life, cognitive screening, and stigma scores. Post hoc characterization described the classes by epilepsy etiology, derived epilepsy type, and seizure categories aligned with current terminology. Results: A three-class solution was selected, with class sizes of 314, 465, and 275. DRE prevalence increased stepwise across classes: 5.7%, 14.2%, and 33.1%, respectively (p < 0.001). In adjusted analysis, Class 2 had higher odds of DRE than Class 1 (odds ratio 2.70, 95% confidence interval 1.56–4.67), while Class 3 showed the strongest association (odds ratio 8.19, 95% confidence interval 4.15–16.16; both p < 0.001). Higher-burden classes showed lower quality-of-life and cognitive scores and higher stigma scores. Conclusions: LCA identified three clinically interpretable, burden-enriched phenotypic profiles associated with a stepwise gradient in DRE and broader multidimensional disease burden. These cross-sectional profiles may provide a useful framework for describing clinical heterogeneity in adult epilepsy and generating hypotheses for future validation studies. Full article

Ceramic additive manufacturing offers strong potential for fabricating geometrically complex and application-specific components, yet achieving reliable dimensional fidelity remains challenging because dimensional deviation is governed by highly coupled material, process, thermal, and environmental factors. To address this problem, this study proposes an uncertainty-aware computational framework for dimensional error prediction in ceramic 3D printing under variable material and process conditions. The contribution is positioned as a system-level integration of established learning, uncertainty estimation, calibration, and reliability-interpretation components within a ceramic additive manufacturing dimensional-error prediction workflow, rather than as a fundamental methodological breakthrough. The validation is conducted using the publicly available Ceramic 3D Printing Process Control Dataset, a 1000-sample tabular dataset, and the resulting findings are therefore interpreted as dataset-specific computational evidence rather than direct proof of industrial deployment readiness. The methodology begins with a structured data-driven preprocessing pipeline that transforms the Ceramic 3D Printing Process Control Dataset into a multi-condition feature space through data cleaning, one-hot material encoding, min–max normalization, and engineered descriptors capturing extrusion–speed balance, thermal gradients, cooling intensity, deposition density, and material-conditioned interactions. A multi-branch deep computational architecture is then developed to encode material, process, thermal-environmental, and engineered-feature streams separately, followed by adaptive cross-condition fusion to learn nonlinear dependencies across ceramic printing regimes. To improve reliability beyond deterministic regression, the framework jointly models aleatoric and epistemic uncertainty and incorporates calibration refinement to align predictive confidence with observed error behavior, thereby enabling preliminary reliability-oriented interpretation of stable and high-risk operating conditions. Experimental results demonstrate that the full model achieves the best overall within-dataset performance, with a test MAE of 0.0118, RMSE of 0.0172, $R^2=0.999$, MAPE of 1.74%, calibration error of 0.003, PICP of 0.996, reliability score of 0.992, and a stable prediction rate of 98.7%. Although these values indicate strong predictive behavior under the current structured dataset, the exceptionally high $R^2$ should be interpreted cautiously because external experimental validation, larger measured datasets, and cross-machine ceramic printing trials are still required. These findings show that the proposed framework provides an effective system-level computational strategy for dataset-specific reliability-aware dimensional quality prediction in ceramic additive manufacturing and offers a preliminary data-driven foundation for uncertainty-aware intelligent process optimization. Full article

Background: Accurate preoperative prediction of length of hospital stay (LOS) after surgery for colorectal liver metastases (CRLMs) could improve patient counselling and resource planning, yet reliable risk tools are lacking. We aimed to develop an interpretable machine learning model predicting LOS following first-time liver-directed surgery for CRLMs. Methods: In this multicenter cohort study, we included patients who underwent first-time liver resection, ablation, or a combination for CRLMs at three Danish hepatobiliary centers between 2016 and 2023. Preoperative features from two national registries were used to train Elastic Net, Random Forest, HistGradientBoosting, and Explainable Boosting Machine (EBM) algorithms. Hyperparameters were optimized using five-fold cross-validation. Performance was evaluated on a 20% hold-out test sample using mean absolute error (MAE) with bootstrapped 95% confidence intervals (CIs). Results: Among 915 patients, median LOS was 4.0 days (interquartile range (IQR) 3.0–6.0). All four algorithms achieved comparable prediction error (MAE 3.0–3.1 days). The EBM (MAE 3.1 days, 95% CI 2.6–4.3) algorithm was selected for its inherent interpretability. Surgical approach was the strongest predictor, where percutaneous and laparoscopic approaches were associated with reductions of 1.9 and 1.2 days, respectively. Tumor burden, including number of lesions and largest lesion diameter, showed progressive non-linear associations with longer stays. Nonetheless, overall explained variance was low (R² ≤ 0.10), and calibration showed systematic underestimation of stays beyond five days. Conclusions: An inherently interpretable machine learning model matched the predictive performance of opaque algorithms for LOS after CRLM surgery, although overall predictive accuracy was modest and longer stays were underestimated. Explainability analysis identified surgical approach and tumor burden as the most influential predictors. External validation in healthcare systems with different discharge practices is warranted. Full article

High-resolution CTD profiles, with SVP cross-validation of the sound speed field, were recorded at a single station in the outer roadstead of the Port of Constanța (northwest Black Sea; 44°07′41″ N, 28°53′15″ E; depth ≈ 25 m; June 2024), revealing a strongly stratified, five-layer water column driven by three combined forcing mechanisms: seasonal thermal stratification with an abnormally shallow Cold Intermediate Water layer (7.3–15.6 m), Danube-sourced freshwater input, and anthropogenic disturbances consistent with port and anchorage activity. A contextual hypothesis is proposed that conflict-related marine traffic intensification may contribute to observed signals, but physical measurements cannot establish causation. At the main pycnocline (7.31–15.62 m), a density difference of Δρ = 4.02 kg m⁻³ yields a maximum Brunt–Väisälä frequency of N² = 2.37 × 10⁻³ s⁻², reducing vertical eddy diffusivity by two orders of magnitude (Kz ≈ 10⁻⁶ m² s⁻¹). Physical conditions—a shallow mixed layer (~0.7–1.2 m) and strong pycnocline—support the theoretical expectation of surface-layer contaminant accumulation; however, no chemical measurements were carried out to confirm contaminant presence. All contamination inferences rely exclusively on physical proxies (turbidity, dissolved oxygen, and density gradients), and contaminant retention remains untested for lack of direct chemical evidence. A dimensional Stratification-Controlled Retention Index (SCRI = N²/Kz; units: m⁻² s⁻¹) is introduced, and its consistency with the observed hydrographic structure is demonstrated. Full article

End-to-end motion control of 7-degree-of-freedom (DOF) redundant manipulators under sparse reward signals presents a fundamental challenge in deep reinforcement learning (DRL) for robotics: the vast configuration space and absence of dense gradient information combine to produce severe cold-start failures and high cross-seed training variance. This paper proposes Curriculum-SAC-HER, a novel fusion framework integrating Soft Actor–Critic (SAC), Hindsight Experience Replay (HER), and a performance-driven three-stage Automatic Curriculum Learning (ACL) scheduler, designed to resolve the cold-start exploration bottleneck within a training budget of 300,000 environment interaction steps. The core methodology progressively expands the spatial target distribution across three stages of increasing difficulty, conditioning each stage transition on an 80% rolling success threshold to guarantee kinematic prior consolidation before advancing. A rigorous evaluation across 15 independent training runs (five seeds per group, all retained without filtering) demonstrates that the proposed framework achieves a final mean success rate of 84.8% (std: 11.0%), substantially surpassing the SAC + HER ablation (70.3%, Mann–Whitney U test, p = 0.028) and the DDPG baseline (22.3%, p = 0.008), while compressing cross-seed variance by 67% relative to the ablation. Zero-shot robustness evaluations under simulated domain perturbations further reveal that the learned policy maintains above 92% success across extreme friction variations and sustains 71.8% success under a 1.5× payload increase, demonstrating that the ACL module fosters generalized kinematic representations rather than over-fitting to specific contact mechanics. Full article

Rolling bearings, as core components of rotating machinery, are prone to failure under harsh working conditions, and their fault diagnosis is crucial for the safe operation of industrial systems. Aiming at resolving the problems of weak fault feature representation, poor model generalization ability and high dependence on manual preprocessing in traditional bearing fault diagnosis methods, an end-to-end fault diagnosis model named KanMSConv is proposed for one-dimensional raw vibration signals. The model abandons complex time–frequency transformation and manual feature engineering, and constructs a multi-scale feature extraction module based on depthwise separable convolution to capture local impulsive components and global modulation characteristics of fault signals simultaneously. The SE channel attention mechanism is integrated to adaptively enhance fault-related critical features and reduce redundant channel responses. Residual connection is introduced to alleviate the gradient degradation problem of deep networks and improve feature reuse capability. On this basis, the Kolmogorov–Arnold Network (KAN) is used to replace the traditional fully connected layer, which enhances the model’s ability to fit complex nonlinear mapping relationships and distinguish fault classification boundaries. Experimental verification is carried out on three representative rolling bearing datasets (CWRU, PU, SDUST) under multi-load, multi-class and cross-platform conditions. The results show that the KanMSConv model achieves 100% accuracy on the CWRU dataset, 99.93% on the PU dataset and 99.80% on the SDUST dataset, which is significantly superior to the existing mainstream fault diagnosis models in terms of Accuracy, Precision, Recall and F1-Score. And the ablation and computational cost analyses further support this conclusion. Full article

We develop SAFIRE (Self-Attention Fuzzy-Inspired Rule Estimator), a differentiable fuzzy-inspired rule-scoring surrogate for binary medical tabular classification coupling multi-head self-attention, Gaussian membership functions, and Hard Concrete gates for continuous rule scoring. We position SAFIRE as a smooth surrogate of the discrete $L_0$-regularised rule-selection problem and establish five mathematical results and one complexity remark: (1) the relaxed objective is differentiable almost everywhere under positive Gaussian widths (enforced by a Softplus reparameterisation) and fixed batch-normalisation statistics; (2) the deterministic-inference active threshold is strictly stricter than the expected-nonzero training threshold, identifying Hard Concrete gates as continuous rule-scoring devices rather than automatic pruning mechanisms; (3) per-sample forward complexity identifies attention and rule layers as the dominant terms; (4) the Softplus–BatchNorm–linear rule operator violates all four triangular-norm axioms—with necessary and sufficient conditions per axiom and a no-finite-parameterisation impossibility result—while a Softplus reparameterisation restores coordinate-wise monotonicity; (5) a margin-based upper bound characterises disagreement between the full classifier and a top-k rule-only surrogate; and (6) the Softplus-reparameterised constrained variant is provably coordinate-wise monotone with explicit asymptotic regimes. Evaluated on four University of California, Irvine (UCI), medical binary tabular benchmarks under repeated stratified cross-validation, SAFIRE-Prog is statistically competitive with strong interpretable, modern, and gradient-boosting baselines, with one Bonferroni-significant gain over RuleFit on the Diabetic Retinopathy Debrecen corpus. The 48-configuration Hard Concrete sweep, constrained-variant comparison, and a top-k fidelity analysis (per-fold range 0.73–0.95) provide quantitative companion measurements for the mathematical framework. A supplementary large-scale hospital electronic health record (EHR) benchmark (Diabetes 130-US Hospitals, $n=101,766$) shows the rule-scoring mechanism scales to ∼${10}^5$ records and, under severe class imbalance, statistically matches gradient boosting on accuracy while significantly exceeding it on macro-F1. The results offer a mathematically auditable pathway towards interpretable, auditable rule scoring for medical tabular classification, with rule signatures defined in a projected latent space rather than over raw clinical variables. Full article

Flow-cell ultrasonication of gelatinized rice flour slurries alters cultivar-dependent water solubility, viscosity, and retrogradation of pregelatinized rice flour, properties important for plant-based beverages and convenience foods. We tested whether cultivar-level composition descriptors, amylose, protein, and fiber, can represent cultivar-associated variation in ultrasonication responses while separating process-only prediction, within-domain cultivar representation, and unseen-cultivar transfer. Six rice cultivars were processed across nine amplitude-time combinations and two slurry concentrations. Water solubility index, apparent viscosity at a shear rate of 50 s⁻¹, and setback viscosity were modeled using ElasticNet, partial least squares regression, support vector regression, random forest, and extreme gradient boosting. Three input formulations were compared: process variables alone, process variables plus composition descriptors, and process variables plus cultivar identity. Repeated nested group cross-validation showed insufficient process-only prediction and substantial improvement from composition descriptors. Within-domain validation showed comparable composition-descriptor and cultivar-identity performance under nonlinear algorithms. However, because cultivar identity is undefined for absent cultivars, leave-one-cultivar-out transfer of the composition-descriptor model remained uncertain. Cross-fitted Shapley additive explanations showed predictions used process and composition variables. For the validated cultivar-process domain, this approach can screen cultivar-process combinations for beverage and convenience-food applications, but replacing categorical source identifiers with continuous descriptors requires explicit transfer validation. Full article