Search Results (1,353)

Uniform deposition of nanoparticle suspensions over large-area substrates remains challenging because coating formation is strongly influenced by suspension properties and spraying conditions. Therefore, this study employed computational fluid dynamics (CFD) to simulate the air-assisted spray deposition of nano-Al₂O₃ suspensions, with method validation. The effects of nanoparticle volume fraction, feed rate, and air pressure on deposition morphology and coating thickness distribution were investigated. Increasing the nanoparticle volume fraction from 1% to 4% improved thickness uniformity, reducing the normalized root-mean-square error and coefficient of variation (CV) to 5.52% and 38.90%, respectively, while a further increase to 5% provided no additional improvement. Increasing the feed rate from 1 g/s to 5 g/s increased the mean coating thickness from 3.09 μm to 10.97 μm and the effective coverage area from 2.5 × 10⁴ mm² to 4 × 10⁴ mm², but intensified local thickness accumulation. Increasing the air pressure from 0.1 MPa to 0.3 MPa enhanced droplet dispersion, increasing the effective coverage area from 2.45 × 10⁴ mm² to 3.57 × 10⁴ mm² while decreasing the mean coating thickness from 10.25 μm to 6.99 μm. These findings provide a reference for process selection in large-area nanoparticle suspension deposition. Full article

Deep learning has achieved high accuracy in machining condition monitoring, yet accuracy alone cannot reveal whether a model relies on physically meaningful evidence or on dataset shortcuts. The difficulty is that common tasks such as tool-wear and surface-roughness estimation lack a known ground-truth basis for discrimination, so attribution outputs cannot be validated. This study therefore constructs a benchmark task—three-class cutting-speed classification (Vc = 60, 80, 100 m/min)—in which, for a fixed tool diameter, the spindle-speed-dependent spectral content is an a priori known correct basis, enabling the reliability of explainable AI (XAI) methods to be examined under controlled conditions; the task is deliberately simple and serves as a validation framework for attribution methods rather than as a demanding monitoring problem. Cutting force signals were acquired from 3 × 3 full-factorial milling experiments on 17-4PH stainless steel using a Kistler 9170A rotating dynamometer; the torque Mz and resultant force Fa were segmented by sliding windows into 2970 single-channel samples (15 records × 99 windows × 2 channels), each represented by 50 time segments × 6 amplitude-independent statistics. A 1D-CNN, an LSTM, and a fully connected network (NN) were evaluated under five-fold StratifiedGroupKFold cross-validation. The 1D-CNN achieved 93.97% accuracy with roughly one-seventh of the parameters (NN 90.24%, LSTM 89.12%), the smallest fold-to-fold standard deviation, and 84.5% on the most difficult fold; as a single random seed was used, performance conclusions are descriptive only. Holding out whole cutting conditions reduced the mean accuracy to 83–86% (one feed level: 54–57%), so the accuracies describe generalization across records of the studied conditions, not to unseen conditions. Fold-wise application of LRP, LIME, and SHAP showed that the 1D-CNN and NN distributed importance across the entire signal (normalized entropy ≥ 0.97), whereas for the LSTM the gradient × input attribution used as the LRP implementation placed 9.4 times the uniform share on the final time segment, consistently across all five folds, a concentration that SHAP (expected gradients referenced to background training samples) and the perturbation-based LIME did not reproduce. Gradient-free occlusion and resampling reproduced it, and retraining showed that it follows the last input position while removing the final segments did not reduce accuracy; it therefore reflects positional sensitivity rather than reliance on the sequence end. In this run, cross-fold Spearman correlations of LIME rankings lay within −0.05 to 0.05, versus 0.20–0.70 for LRP and SHAP; because exploratory recalculations with alternative aggregation choices gave markedly different LIME correlations, this is reported as a run-specific observation rather than as a property of LIME. All three architectures relied predominantly on frequency-domain features; because the tooth-passing frequencies (159.2–265.3 Hz) exceed the 100 Hz Nyquist frequency of the 200 Hz sampling rate yet lie below the 300 Hz low-pass cutoff of the signal conditioner, they are aliased into the analyzed band, and the dominant-frequency feature reflects the spindle rotation frequency and aliased tooth-passing components, both proportional to spindle speed and hence consistent with the physical definition of cutting speed. The contributions are: a controlled validation framework that cross-validates attribution credibility through importance concentration, inter-method agreement, and cross-fold reproducibility; and the observation that the end-of-sequence concentration of the LSTM, revealed by gradient × input but not by SHAP or LIME, reflects positional sensitivity rather than the basis of classification, highlighting the need to verify attribution concentration with independent analyses rather than relying on a single attribution method. Full article

Kidney renal clear cell carcinoma (KIRC) is a major subtype of renal cell carcinoma with substantial molecular heterogeneity and variable clinical outcomes. Tetraspanins (TSPANs) are membrane proteins involved in cell adhesion, signaling, migration, tumor progression, and interactions with the tumor microenvironment (TME); however, their collective significance in KIRC remains insufficiently characterized. The expression profile of the TSPAN gene family was evaluated in the TCGA-KIRC cohort using GEPIA2. Differentially expressed TSPANs were further investigated through protein–protein interaction network (PPIN) and univariate survival analyses. Prognostically significant TSPANs were assessed for genomic alterations, promoter methylation, and associations with tumor-infiltrating immune cells. Feed-forward-loop (FFL) was constructed and visualized using Cytoscape. Thereafter, the findings were evaluated in an independent KIRC dataset. Several TSPANs displayed significant differential expression between KIRC and normal renal tissues. TSPAN1 and TSPAN6 were significantly associated with overall survival (OS), with higher expression corresponding to more favorable outcomes. Notably, higher TSPAN6 expression was associated with improved OS but a higher estimated risk of relapse indicating an endpoint-dependent and potentially complex prognostic relationship. Promoter methylation of TSPAN1 and TSPAN6 showed associations with selected clinicopathological characteristics, while both genes demonstrated significant relationships with specific immune-cell populations. Network analysis identified a ZNF260–miR-6794-5p–TSPAN6 FFL subnetwork motif, prioritizing TSPAN6 as the hub TSPAN. TSPAN6 was additionally detected among differentially expressed genes (DEGs) in an independent GEO cohort. This study identifies TSPAN6 as a promising candidate biomarker in KIRC and provides a multi-dimensional framework linking TSPAN dysregulation with prognosis, epigenetic regulation, immune infiltration, and regulatory networks. Experimental and clinical validation is warranted. Full article

Purpose: This study aimed to investigate the protective effects and potential mechanisms of exercise against long-term high-fat diet (HFD)-induced skeletal muscle dysfunction, with a focus on ADP-ribosylation factor 1 (ARF1)-related autophagy–lysosome markers and nuclear factor erythroid 2-related factor 2 (Nrf2) signaling. Methods: Forty male C57BL/6 mice were randomly divided into a normal diet group (DC) and an HFD group, and the latter was further subdivided into a model group (MG), exercise group (EG), Nrf2 inhibitor group (NG), and exercise combined with Nrf2 inhibitor group (ENG) after 12 weeks of HFD feeding. EG received 6 weeks of moderate-intensity treadmill exercise, NG was treated with Nrf2 inhibitor for 6 weeks, and ENG received both the 6-week treadmill exercise and inhibitor injections simultaneously. Serum lipid, blood glucose and insulin resistance indices were detected; skeletal muscle morphology was observed by H&E staining; mRNA and protein expression of ARF1, ULK1, ATP6V0A1 and other factors were determined by RT-qPCR, Western blotting and immunohistochemistry; the oxidative stress index was measured by an ELISA; and the colocalization of ARF1/ATP6V0A1 with the lysosomal marker Lamp1 was analyzed by immunofluorescence. Results: Long-term HFD significantly increased mouse body weight, serum TC, TG, LDL-C, FBG levels and the HOMA-IR index, induced morphological alterations in skeletal muscle (loose myofiber arrangement, abundant intracellular vacuoles), inhibited the mRNA expression of ARF1, ULK1 and ATP6V0A1, and elevated malondialdehyde (MDA) content (p < 0.05 or p < 0.01). Six weeks of moderate-intensity treadmill exercise effectively improved HFD-induced metabolic disorders and skeletal muscle morphological damage, upregulated the expression of ARF1, ULK1 and ATP6V0A1, and reduced oxidative stress (p < 0.05 or p < 0.01). Notably, luteolin-mediated inhibition of Nrf2 signaling may partially reverse exercise-induced alterations in ARF1-related autophagy–lysosome markers and skeletal muscle morphology. Conclusions: Long-term HFD may induce morphological alterations in skeletal muscle, accompanied by changes in ARF1-associated autophagy–lysosome markers and elevated oxidative stress. Conversely, moderate-intensity exercise may ameliorate HFD-related changes in skeletal muscle morphology, potentially via Nrf2 signaling, providing a potential molecular target and theoretical basis for the exercise intervention strategies against HFD-induced metabolic-related skeletal muscle dysfunction. Full article

Morphological and skeletal deformities compromise growth, swimming, feeding, welfare, and market value in farmed fish, but externally visible abnormalities may not reveal the underlying tissue lesions. This study characterized naturally occurring deformities in farmed rainbow trout (Oncorhynchus mykiss) using integrated macroscopic, radiographic, and histopathological evaluation. A total of 54 morphologically abnormal fish and one clinically normal fish were examined. Deformities were grouped by predominant external region, whole-body radiographs were evaluated for skeletal and structural abnormalities, and affected tissues were processed for routine histopathology. Caudal peduncle or tail-region deformities were most frequent (37.0%), followed by trunk/body deformities (25.9%), head-region deformities (24.1%), and specific or combined deformity patterns (13.0%). Vertebral compression was the dominant radiographic lesion (74.1%), followed by swim bladder anomalies (48.1%), contour irregularities (31.5%), combined deformity patterns (29.6%), and peduncle thickening (25.9%). Histopathology revealed ocular lesions compatible with cataract and microphthalmia, vertebral asymmetry and compression, cartilage and bone disorganization, intervertebral space loss, fibrosis, and skeletal muscle degeneration and atrophy. These findings show that visible deformities often reflect complex skeletal and soft-tissue pathology. Integrating radiography with histopathology provides a more informative framework for deformity assessment and welfare-oriented management in rainbow trout production. Full article

Semantic integration in Additive Manufacturing (AM) supervision demands unifying heterogeneous sensor streams and production context under strict latency and data-integrity constraints, yet conventional semantic stacks suffer from unpredictable query latencies, unguarded SPARQL (SPARQL Protocol and Resource Description Framework [RDF] Query Language) endpoints, and prohibitive reasoning overheads. This paper presents Semantic DREAM (Distributed Resilient and Emergent intelligence-based Additive Manufacturing), a security-aware architecture co-designed over a Web Ontology Language 2 Rule Language (OWL 2 RL) modular substrate with three mechanisms: (1) an abstract syntax tree (AST)-guarded read-only SPARQL engine that blocks destructive updates and caches template results; (2) a multi-protocol mashup layer aligning live Message Queuing Telemetry Transport (MQTT), Representational State Transfer (REST), Structured Query Language (SQL), and file feeds through Sensor, Observation, Sample, and Actuator (SOSA) observation minting and Quantities, Units, Dimensions, and Types (QUDT) unit normalization; and (3) an ingestion-time Shapes Constraint Language (SHACL) validation pipeline that gates observations before graph commit, decoupling structural checks from reasoning. Deductive materialization is performed offline at startup and verified with ROBOT as an intentional design choice to eliminate reasoning overhead from the streaming path, so online monitoring queries evaluate directly over the asserted graph. Single-host, lab-scale evaluations show that cached queries achieve sub-15 ms response times for recurring monitoring, while the worst-case event-driven decision cycle (incurring a post-invalidation cache miss) completes in 110 ms, well within the 200 ms supervisory deadline with 100% defect interception on injected anomalies. All measurements are single-host and lab-scale; factory-network transit and recorded telemetry are not evaluated. Full article

Current rations are formulated on the requirements of a representative cow plus a correction factor. This is not precision feeding because in practice cows are fed as a pen, and within a pen, dry matter intake (DMI) among cows is not uniform and has skew or kurtosis. Since ration formulation depends on DMI as a binding constraint, rations formulated to the pen DMI could be more precise if the distribution of DMI is better described. The objectives of this study were to: 1. Develop an algorithm, the distribution shape model (DSM), to predict pen DMI based on the distribution of cow measures which could easily be estimated on a commercial dairy farm; and 2. Compare predictions of pen DMI from the DSM to predictions of pen DMI based on a representative cow using the National Academies of Sciences, Engineering, and Medicine (NASEM) equation model (NSM). Individual cow DMI from previously published studies was used to train an Extreme Gradient Boosting (XGB) algorithm to predict the distribution shape of pen DMI (Beta or Generalized Normal) from parity, week in milk, and descriptive statistics of milk yield. The NSM predicted DMI for a representative cow from pen-average inputs, and this prediction was multiplied by the number of cows, simulating the scenario where DMI is assumed to be uniform across a pen. Both models were tested by comparing predicted with observed total pen DMI using an independent testing dataset of 229,909 pens generated from previously collected cow observations. For pens classified as Beta, the mean squared error of prediction (MSEP) of pen DMI from the DSM and NSM was 5249 and 9736 kg², respectively (root mean squared error of prediction, RMSEP, 72.5 and 98.7 kg/d), with 3.0% and 80.2% due to mean bias, 5.4% and 8.3% due to regression slope not equal to 1, and 91.6% and 11.5% due to random variation. For pens classified as Generalized Normal, MSEP was 2086 and 6431 kg² (RMSEP 45.7 and 80.2 kg/d), with 0.8% and 73.2% due to mean bias, 7.6% and 4.7% due to slope not equal to 1, and 91.6% and 22.1% due to random variation. The NSM consistently overpredicted pen DMI, contributing to its higher error. Including the distribution shape in the prediction of DMI improved precision and accuracy of pen DMI prediction and should be considered when estimating the total daily DMI fed to a pen. Full article

In large-scale pig farming systems, high visual similarity among pigs, complex posture variations, and frequent interactions between individuals pose significant challenges to vision-based multi-behavior detection. To overcome these limitations, this study proposes an improved YOLO11-based pig behavior detection model, termed DWEC-YOLO, for recognizing normal behaviors (standing, lying, sitting, and feeding) and aggressive behaviors (fighting, ear biting, and tail biting). First, the C3k2-DySnake module is embedded within the backbone architecture, enabling more effective extraction of fine-grained visual details. Second, an EUCB-SC module is employed during feature aggregation and works jointly with the CoordAtt mechanism to strengthen multi-scale feature representations, thereby enhancing the network’s sensitivity toward critical interaction areas. In addition, Wise-ShapeIoU is adopted as the bounding-box regression loss to refine localization optimization and produce more accurate object predictions. Comprehensive experiments show that DWEC-YOLO surpasses several representative lightweight detectors, including YOLOv5n, YOLOv7-Tiny, YOLOv8n, YOLOv10n, and YOLOv12n, in detection effectiveness. Relative to the original YOLO11 architecture, DWEC-YOLO improves precision, recall, and mAP@0.5 by 4.28, 4.46, and 2.65 percentage points, respectively, achieving 96.17%, 95.25%, and 97.57%. These outcomes demonstrate that DWEC-YOLO substantially enhances multi-behavior detection in group-housed environments, serving as a dependable tool for automated monitoring in smart livestock farming. Full article

The melon fly, Zeugodacus cucurbitae (Coquillett), is a globally invasive pest of cucurbit crops whose concealed larval feeding limits chemical control and has driven interest in reproduction-based management. The chorion gene s36 encodes an eggshell protein conserved among dipterans. However, whether s36 is required for female reproduction in Z. cucurbitae remains unknown. We generated a homozygous s36 mutant line by CRISPR/Cas9 and characterized its reproductive phenotype. Disruption of s36 strongly reduced transcript abundance and truncated the predicted protein. Mutant females showed elevated mortality after sexual maturation, fecundity was approximately halved, and none of the eggs laid hatched. Ovaries were markedly reduced in size and were scored as abnormal in every mutant female examined, and eggs were smaller, dull and shriveled rapidly in air, indicating compromised chorion integrity. Both vitellogenin genes were down-regulated in the ovary. Whole-body transcriptomics identified 179 differentially expressed genes enriched in lipid and hormone-related metabolism. Tissue-resolved RT-qPCR showed vitellogenin and yellow-g2 transcripts elevated in the fat body yet depleted in the ovary; whether this pattern reflects impaired delivery of yolk precursors to the oocyte or the secondary loss of the follicular tissue in which these genes are normally expressed remains to be determined. These findings identify s36 as required for normal ovarian maturation and egg viability in Z. cucurbitae, and as a candidate target for genetic control of this pest. Full article

Grasslands are among the most extensive and most contested biomes on Earth: published estimates of their global area range from 3.5% to 42% of the ice-free land surface, and a recent expert-validated reassessment places the figure at 22.8% (30.1 million km²). They sustain biodiversity, store roughly one-fifth of global soil carbon, and feed most of the world’s extensive livestock systems, yet degradation is widespread and assessment methods remain fragmented. This review examines the remote sensing indicators used to evaluate grassland conservation status, drawing on 179 studies retrieved from the Web of Science Core Collection and analysed under an adapted PRISMA 2020 workflow, with background and policy literature to early 2026. Indicators were grouped under three parameters adapted from the EU Habitats Directive reporting framework: Occupied Area, Structure and Function, and Future Prospects. Sentinel-2 is the most frequently used sensor in the recent literature, in 29 of the 73 studies published between 2019 and 2025 whose sensor could be determined (40%), owing to its 10–20 m pixels, red-edge bands and 5-day revisit; reported classification accuracies for grassland extent run from 65.0% to 98.9% across the 17 studies reporting overall accuracy, with a median of 91.5%. Standing aboveground biomass is retrieved with R² between 0.30 and 0.87 across 24 studies, median 0.67, when vegetation indices are combined with Random Forest, support vector or neural-network regression; fractional vegetation cover reaches 0.81 to 0.99 in three studies and leaf area index 0.67 in the one eligible study. The upper end of each range comes mostly from single-site studies with non-independent validation, and only two of the 24 biomass studies with eligible accuracy values were validated against an independent field campaign. Long-term Normalized Difference Vegetation Index (NDVI) trends support degradation detection over four decades, but the record that spans that period is the 30 m Landsat archive from 1984; the harmonised Landsat-Sentinel-2 product (HLS v2.0) is 30 m and begins in 2013, and the red-edge bands that most recent retrievals depend on exist only from Sentinel-2 in 2015. Species and plant-community composition is classified with overall accuracies of 75.3% to 98.9% across 11 studies, and canopy height is retrieved with R² of 0.51 and 0.58 in the two studies reporting it, one of which gives an RMSE of 3.94 cm; both remain constrained by acquisition cost and revisit frequency. Coverage is strongly uneven: Asia and Europe account for most of the corpus, while Africa, South America and the Southern Hemisphere generally are under-represented. We set out the physical basis of each index, the dependencies among indicators, a quality-tiered synthesis of reported accuracies, and a research agenda ranked by importance and feasibility. Full article

The balance between fed and extractive compartments is critical in recirculating multitrophic aquaponic systems (RMAS). This study evaluated the effects of Pacific oyster (Crassostrea gigas) stocking level on water quality, gilthead seabream (Sparus aurata) performance, oyster growth, glasswort (Salicornia europaea) production, and whole-system fresh-product output. Nine independent RMAS received no oysters (O-00), 30 oysters (O-30), or 60 oysters (O-60). Ninety juvenile seabream (79.7 ± 13.8 g) were reared for 134 days with salicornia, while oysters received no supplemental feed. Oyster inclusion was associated with lower nitrate and phosphate concentrations, whereas other water-quality variables showed no treatment-related differences. Fish growth, feed conversion ratio, survival, slaughter traits, fillet pH, and scale cortisol did not differ among treatments. Oyster survival was 100%, with modest positive biomass gain that did not increase proportionally at the higher stocking level. Salicornia yield was numerically highest in O-30, and feed-normalized apparent total and extractive fresh-product outputs were greater in O-30 than in O-60. Pacific oysters can therefore be integrated into saline RMAS without detectable adverse effects on fish performance, but increasing oyster abundance does not necessarily enhance system performance. Increasing oyster abundance did not provide proportional improvements in system performance under the conditions tested. Full article

Pulmonary sequestration is a rare congenital malformation supplied by an aberrant systemic artery, making parenchyma-sparing resection technically challenging because safe vascular control and accurate identification of the resection plane are both required. Indocyanine green (ICG) fluorescence has been used to delineate the perfusion boundary between sequestrated and normal lung, whereas preoperative embolization has traditionally been performed to reduce the risk of intraoperative bleeding. We describe a staged approach in which selective coil embolization of the aberrant feeding artery, performed one day before surgery, was followed by video-assisted fluorescence-guided wedge resection. Preoperative embolization established devascularization before surgery, while subsequent ICG fluorescence provided clear real-time delineation of the perfusion boundary, enabling tailored parenchyma-sparing resection. Although this observation derives from a single case and does not support routine preoperative embolization, it raises the hypothesis that the interval between embolization and surgery may influence the physiological conditions under which subsequent fluorescence-guided perfusion assessment is performed. This potential effect requires confirmation through quantitative fluorescence assessment and comparative studies. Full article

Necrotizing enterocolitis (NEC) is a devastating gastrointestinal emergency affecting up to 10% of infants < 1500 g. Although prior metabolomic studies have profiled metabolite derangements, few have focused on intestinal tissue, the primary site of injury. In this study, we performed metabolomic profiling of intestinal tissue from patients with NEC and surgical controls to identify metabolic alterations associated with NEC. Tissue was collected during surgery from twenty-one infants (8 controls, 13 NEC). Samples underwent liquid chromatography–tandem mass spectrometry. Raw data were extracted, peak-identified, and quality-control-processed, and compounds were identified by comparison with a purified standards library. Following normalization and log transformation, a mixed-effects model accounting for repeated samples from individual infants was used to identify metabolites that differed between NEC and control tissue. There were no significant differences in sex, gestational age, days of life, antibiotic use, or type of feeds between groups. Principal component analysis and hierarchical clustering demonstrated partial separation of NEC and control samples. Metabolomic profiling identified 1183 biochemical features, with 308 demonstrating nominal differential abundance between NEC and control tissue. Indole-related metabolites were increased in NEC tissue, including the nominally significant 3-indoxyl sulfate, a host–microbial co-metabolite associated with epithelial barrier dysfunction. Caffeine-related metabolites were also increased in NEC tissue, while docosahexaenoate demonstrated a trend toward increased relative abundance that did not reach statistical significance. Enrichment of retinol metabolism was observed in NEC tissue, suggesting alterations in retinoid-associated metabolic pathways. Hydroxybutyrate demonstrated a nonsignificant trend toward decreased abundance in NEC. Overall, NEC intestinal tissue demonstrated broad metabolic alterations compared with surgical controls, including differences in lipid, retinol, and host–microbial metabolic pathways. These findings characterize the intestinal metabolic environment associated with NEC and highlight the value of tissue-based metabolomics for understanding disease-associated intestinal alterations. Full article

Feed spacers can promote membrane-side transport in flat-sheet ultrafiltration channels, but the accompanying hydraulic resistance increases pumping demand. This study combined pure-water pressure-drop measurements with transient three-dimensional ANSYS Fluent simulations to compare a commercial 2/1 twill-woven spacer (T-Base) and 16 systematically varied candidate configurations. A geometry-resolved finite-volume model was evaluated through computational-domain, mesh, and time-step sensitivity assessments; the predicted pressure gradients agreed with the measurements with a mean relative deviation below 4%. The methodological novelty lies in integrating pressure-normalized near-wall efficacy, bilateral wall shear stress balance, and equal-area spatial sampling within a geometry-resolved, multi-indicator framework for comparing spacer configurations with different periodic lengths and mesh densities. At a 60° low attack angle and a superficial inlet velocity of 0.0551 m·s⁻¹, T3823 (filament spacing 0.38 mm, filament diameter 0.23 mm, spacer thickness 0.44 mm) provided the most favorable hydrodynamic trade-off among the 16 candidates. It produced a near-membrane velocity of 0.0220 m·s⁻¹ and a pressure gradient of 53.5 kPa·m⁻¹; its pressure-normalized near-wall efficacy was 1.23 relative to T-Base. T3823 also reduced bilateral membrane shear deviation from 24.91% to 8.30% and near-wall velocity dispersion by 50.7%. Overall, the results highlight the value of coordinated spacer-geometry optimization for balancing near-membrane flow enhancement and hydraulic resistance, with T3823 emerging as the most promising configuration among the candidates evaluated. Full article

Reverse osmosis (RO) digital twins require predictive models whose accuracy remains credible as plant operating conditions change. We evaluated the readiness of predictive models for a full-scale, two-stage RO plant using operational spreadsheets and temporally structured validation. An audit of 70 legacy worksheets identified heterogeneous layouts. A normalized source provided 476 train-day records from 17 operating months between June 2022 and November 2023, while 55 records from April and May 2024 formed an independent later-period holdout. Four supervised regression algorithms (Ridge Regression, Random Forest, Extra Trees, and Gradient Boosting) were evaluated for second-stage permeate flow and conductivity. Under a random split, Random Forest achieved R² values of 0.876 and 0.833, with root mean square errors of 8.48 m³/h and 63.26 µS/cm, respectively. Chronological validation reduced R² to 0.042 and 0.118. On the later holdout, R² fell to −1.173 and −0.918, with corresponding root mean square errors of 11.14 m³/h and 218.65 µS/cm. Rolling-origin tests revealed substantial variability among successive periods. Distributional diagnostics identified changes in feed pressure, first-stage permeate conductivity, second-stage flow, and second-stage conductivity; the available data could not establish their physical causes. These findings show that performance under random splitting is insufficient to establish readiness for future use. A membrane-process digital twin requires traceable data, chronological evaluation, operating-regime surveillance, controlled model updates, and human-supervised advisory operation. Full article