BACKGROUND Chemoresistance significantly limits the therapeutic efficacy of neoadjuvant chemotherapy (NACT) in advanced gastric cancer (AGC). There is an urgent need to identify robust biomarkers predictive of NACT response and to elucidate the molecular mechanisms that drive resistance. In this study, we systematically assess whether intercellular adhesion molecule 2 (ICAM2 ) predicts NACT response in patients with AGC and delineate its mechanistic role in chemoresistance. AIM To investigate the predictive significance and mechanistic role of ICAM2 in mediating 5-fluorouracil (5-FU) resistance in gastric cancer (GC). METHODS Real-time PCR, Western blotting, enzyme-linked immunosorbent assay, and immunohistochemistry were conducted to assess alterations in ICAM2 expression between 5-FU-sensitive and -resistant GC cells as well as in AGC patient samples. Cytotoxicity assays, colony formation, flow cytometry, analyses of apoptosis-related proteins, and xenograft experiments were employed to elucidate the role of ICAM2 in mediating chemoresistance. The mechanism underlying ICAM2 -mediated chemoresistance was further explored through RNA sequencing (RNA-seq), nuclear-cytosolic fractionation, co-immunoprecipitation, luciferase reporter, and chromatin immunoprecipitation assays. RESULTS Low ICAM2 expression correlated significantly with poor NACT response, advanced tumor stage, worse differentiation, and reduced overall survival and disease-free survival in AGC patients. Pre-NACT serum ICAM2 demonstrated high predictive accuracy (area under the curve = 0.876) in discriminating chemotherapy responders from non-responders. Mechanistically, ICAM2 knockdown conferred 5-FU resistance through two intertwined processes: Inhibition of caspase-dependent apoptosis and promotion of immunosuppressive M2 macrophage polarization within the tumor microenvironment. At the molecular level, loss of ICAM2 activated the TGF-β/Smad pathway, leading to transcription factor SP1-mediated pleiotrophin (PTN) upregulation. Elevated PTN further enhanced GC cell survival and may contribute to M2 macrophage polarization, thereby amplifying chemoresistance. Importantly, targeted inhibition of TGF-β signaling reversed ICAM2-associated chemoresistance in both cell culture and xenograft models. CONCLUSION Our study highlights the clinical impact of ICAM2 downregulation predicting poor outcome and NACT response in AGC patients, and reveals a novel ICAM2/TGF-β/Smad/SP1/PTN signaling mediating 5-FU resistance in GC.

Classification of secondary explosives with a 1D convolutional neural network technique using terahertz time-domain spectroscopy in reflection geometry.

Advancing quinoa(Chenopodium quinoa Willd.) quality assessment using hyperspectral imaging.

ABSTRACT Secondary tropical forests are essential for biodiversity conservation and carbon sequestration. However, estimating aboveground biomass (AGB) in these ecosystems remains challenging due to their structural complexity and the spectral interference caused by canopy shading. This study demonstrates the potential of hyperspectral data acquired by a remotely piloted aircraft, HS-RPA, for modelling AGB in forest remnants at different successional stages in the Atlantic Forest. Specifically, it investigates how variations in spatial resolution and canopy shading affect the accuracy of biomass estimates, providing insights for optimizing remote sensing approaches in structurally complex tropical forests. Thirty permanent plots (30 m × 30 m) were established in four forest remnants in southern Espírito Santo, Brazil. Hyperspectral images were acquired in the 397–1002 nm range, with native resolution of 0.11 m, later resampled to 0.5 m and 5 m. Four scenarios were analysed combining different spatial resolutions and the presence or absence of shadow masking. Regression models indicated that both spectral bands and vegetation indices, particularly Red-edge Vegetation Stress Index (RVSI) and Photochemical Reflectance Index (PRI), contributed to AGB estimation. The best performance was obtained in the scenario with 5 m resolution and without shadow masking (adjusted R2 = 0.78; RMSE = 50.50%). Notably, models without shadow masking outperformed those with shaded pixels removed, indicating that shadow exclusion did not improve predictive accuracy. Shaded areas may capture structural nuances of the canopy, providing complementary information on vegetation heterogeneity. In this study, the spatial resolution of 5 metres also promoted greater spectral stability by smoothing small-scale variations. Overall, airborne hyperspectral data for AGB estimation highlight important aspects related to spatial resolution and illumination conditions in tropical forest monitoring. The results obtained provide valuable support for the development of more accurate methodologies aimed at biomass estimation, contributing to the improvement of forest monitoring strategies.

Contemporary prostate cancer prognostic models do not include imaging and generally are based on pretreatment parameters. We sought to develop an externally validated model that used novel quantification of soft-tissue and bone disease, integrated with standard clinical and serum biomarkers, at baseline and up to 6 months of treatment. Two randomized phase 3 trials, Cougar COU-AA-302 (NCT00887198; for derivation) and Alliance A031201 (NCT01949337; for validation), were used to evaluate the added value of early on-treatment bone imaging and more than 1,000 radiomics features on CT, used in conjunction with clinical and serum biomarkers in first-line metastatic castration-resistant prostate cancer. Predictive accuracy measures were computed to determine whether these early on-treatment biomarkers could reliably sort patients into risk groups that inform overall survival (OS) and whether the patient-specific biomarker risk score could precisely predict their OS time. Imaging improved patient risk stratification but did not improve individual survival predictions. The strongest risk prediction model was developed for patients with bone-only metastases. This model was also the least complex, relying on just 16 risk factors, whereas all other models were high-dimensional, incorporating approximately 1,100 intercept and 1,100 slope features from the early on-treatment biomarker trajectories. Pretreatment and early on-treatment serum and automated quantitative imaging markers can well discriminate risk of death. Imaging improves this risk categorization relative to serum biomarkers alone. Such models can give early outcome predictions and can be used in future trials that involve imaging, even using traditional techniques such as bone scintigraphy.

Abstract. Merging physics-based with data-driven approaches in hybrid hydrological modeling offers new opportunities to enhance predictive accuracy while addressing challenges of model interpretability and fidelity. Traditional hydrological models, developed using physical principles, are easily interpretable but often limited by their rigidity and assumptions. In contrast, machine learning methods, such as Long Short-Term Memory (LSTM) networks, offer exceptional predictive performance but are often criticized for their black-box nature. Hybrid models aim to reconcile these approaches by imposing physics to constrain and understand what the ML part of the model does. This study introduces a quantitative metric based on Information Theory to evaluate the relative contributions of physics-based and data-driven components in hybrid models. Through synthetic examples and a large-sample case study, we examine the role of physics-based conceptual constraints: can we actually call the hybrid model “physics-constrained”, or does the data-driven component overwrite these constraints for the sake of performance? We test this on the arguably most constrained form of hybrid models, i.e., we prescribe structures of typical conceptual hydrological models and allow an LSTM to modify only its parameters over time, as learned during training against observed discharge data. Our findings indicate that performance predominantly relies on the data-driven component, with the physics-constraint often adding minimal value or even making the prediction problem harder. This observation challenges the assumption that integrating physics should enhance model performance by informing the LSTM. Even more alarming, the data-driven component is able to avoid (parts of) the conceptual constraint by driving certain parameters to insensitive constants or value sequences that effectively cancel out certain storage behavior. Our proposed approach helps to analyse such conditions in-depth, which provides valuable insights into model functioning, case study specifics, and the power or problems of prior knowledge prescribed in the form of conceptual constraints. Notably, our results also show that hybrid modeling may offer hints towards parsimonious model representations that capture dominant physical processes, but avoid illegitimate constraints. Overall, our framework can (1) uncover the true role of constraints in presumably “physics-constrained” machine learning, and (2) guide the development of more accurate representations of hydrological systems through careful evaluation of the utility of expert knowledge to tackle the prediction problem at hand.

The yttria-stabilized zirconia (YSZ)-based mixed potential gas sensor represents a promising platform for portable acetone detection systems, owing to their high sensitivity and selectivity. However, the practical deployment of these systems is hindered by sensor output drift, circuit noise, and signal fluctuations caused by variations in operating temperature and ambient humidity. To address these challenges, we developed GT-KANet, a novel hybrid deep-learning algorithm designed for robust temperature-humidity compensation and accurate acetone concentration prediction. The GT-KANet architecture integrates Gated Recurrent Unit (GRU) networks for initial temporal feature extraction, transformer layers enhanced with ContraNorm to mitigate oversmoothing during deep feature learning, and a Kolmogorov-Arnold Network (KAN) module for final concentration prediction. Trained on a comprehensive dataset acquired from our independently developed YSZ sensor under controlled conditions, the proposed GT-KANet achieved exceptional predictive accuracy (RMSE = 0.0920 ppm, MAE = 0.0542 ppm) across varying operating temperatures, ambient humidity, and gas concentrations, demonstrating excellent stability and adaptability. Furthermore, by leveraging knowledge distillation, we achieved a 55% reduction in the model's parameter size while maintaining exceptional prediction accuracy, significantly enhancing its feasibility for deployment on resource-constrained embedded platforms. A detailed comparative analysis during development systematically validated the efficacy of each module in boosting prediction accuracy. This work offers a competitive approach for temperature and humidity compensation using YSZ-based acetone sensors, demonstrating excellent application potential in rigorous detection scenarios.

We developed a Suitable Conditions Index (SCI) to predict dengue transmission in our prior work. However, the initial SCI was not refined with other important abiotic parameters. Therefore, in this study we refined the index by calculating three variants: temperature-based baseline daily average SCI (BDA-SCI), precipitation-weighted daily average SCI (PWDA-SCI), and waterbody-weighted daily average SCI (WWDA-SCI). We used the district-wise data for two South Asian dengue-endemic countries: Bangladesh and Sri Lanka. Temperature-suitable days specific to Aedes aegypti (17.05-34.61℃) and Aedes albopictus (15.84-31.51℃) were averaged (BDA-SCI) and weighted by district-level precipitation (PWDA-SCI) and waterbody data (WWDA-SCI). We assessed the association between dengue incidence and each SCI, along with other covariates using negative binomial regression models. Furthermore, a binomial logistic regression model (BLR) was used to measure the predictive accuracy of each SCI. The BDA-SCI for Ae. aegypti was highest in Sri Lanka at 0.96 (Standard deviation [SD] 0.04, range 0.85-1.00), compared to Bangladesh 0.68 (SD 0.06, range 0.61-0.87). For Ae. aegypti, WWDA-SCI (Relative risk [RR]aegypti = 1.06, p = 0.056, Akaike Information Criteria [AIC] 1218.6) and BDA-SCI (RRaegypti = 1.05, p = 0.008, AIC 1214.2) had a stronger association with dengue incidence in Bangladesh than PWDA-SCI (RRaegypti = 1.06, p = 0.056, AIC 1232.2), whereas in Sri Lanka, PWDA-SCI (RRaegypti = 1.06, p = 0.056, AIC 472.63) performed better (AICBDA-SCI: 481.36, AICWWDA-SCI: 475.89) in the multivariable model, similar to the findings for Ae. albopictus. The BLR model predicted districts with above-median dengue incidence, and model performance indicated that BDA-SCI achieved highest accuracy for Bangladesh, while WWDA-SCI performed best for Sri Lanka, based on higher sensitivity and the Area Under the Curve value. Overall, the SCI method demonstrated a practical approach for identifying dengue vector suitability and transmission risk. Refining this index with location-specific climatic and environmental variables may enhance the model accuracy and may be used for future predictions under climate change scenarios. Thus, our refined SCI will assist in creating a reliable early warning system and inform the policymakers to initiate vector control strategies, including monitoring and eliminating dengue breeding sites and implementing biocontrol strategies within hotspots.

Abstract Plastic pollution poses a significant environmental threat, with PET comprising 12% of global solid waste. Enzymatic degradation provides a sustainable recycling method, yet optimizing enzymes for industrial use remains difficult. This study utilizes protein and polymer language models to predict plastic degradation from protein sequences and polymer SMILES. The approach demonstrates high prediction accuracy with key structural features revealed by language models, enabling efficient in silico screening for protein design.

China continues to face a substantial burden of gastric cancer (GC), particularly with respect to metastasis-related mortality. However, population-based analyses of distant metastasis patterns in Chinese GC patients remain unavailable. Global Burden of Disease (GBD) data on GC from the 1990 to 2021 period was obtained through the Global Health Data Exchange (GHDx) query tool and integrative data of 18919 patients who underwent surgery were obtained from our hospital. Univariate and multivariate logistic regression identified independent risk factors for metastases, and survival analysis utilized univariate and multivariate Cox regression, Kaplan-Meier method, and log-rank test. Predictive nomograms were assessed using metrics such as the area under the curve (AUC), calibration curves, and decision curve analysis. According to the GBD database, GC demonstrates declining global trends in both incidence and mortality. Nevertheless, China continues to face a substantial GC burden, with progressive annual rises in distant metastasis prevalence and metastasis-related mortality. Clinical characteristics and temporal patterns vary significantly across metastatic types. Furthermore, metastatic profiles exhibit sex-, age-, and stage-specific variations. Univariate and multivariate regression analyses identified independent risk factors for overall GC metastasis and site-specific metastases. The resulting prediction models demonstrated excellent predictive accuracy for metastatic progression. The prognostic nomogram was developed to predict 1-, 5-, and 10-year overall survival (OS) in GC patients, with AUCs of 0.86 (0.84-0.88), 0.87 (0.85-0.89), and 0.80 (0.74-0.85) in the training set, respectively, which showed good discriminative ability. In this study, metastatic spectrums across diverse patient subgroups and temporal patterns of metastasis in GC were investigated. Furthermore, we developed clinical predictive nomograms for various metastatic patterns and OS in GC, which enhance the understanding of metastatic behavior and provide a robust tool for personalized risk assessment and prognosis prediction.

Predicting how residue variations affect protein stability is crucial for rational protein design and for assessing the impact of disease-related mutations. Recent advances in protein language models have revolutionized computational protein analysis, enabling more accurate predictions of mutational effects. However, balancing predictive accuracy with the fundamental laws of thermodynamics remains a challenge for sequence-based models. Here we show JanusDDG, a physics-informed neural network that leverages embeddings from protein language models and a bidirectional cross-attention transformer architecture to predict stability changes for both single and multiple residue mutations. By adopting a physics-informed paradigm, the model is explicitly constrained to satisfy fundamental thermodynamic principles, such as antisymmetry and transitivity, while maintaining high predictive performance. Instead of conventional self-attention, JanusDDG employs a cross-interleaved attention mechanism that computes the relationship between wild-type and mutant embeddings to capture mutation-induced perturbations while preserving essential contextual information. Our results demonstrate that JanusDDG achieves state-of-the-art performance in predicting stability changes from sequence alone, matching or exceeding the accuracy of structure-based methods for both single and multiple mutations.

Early action prediction aims to predict complete action intentions by observing the initial stages of action execution, which is crucial for safety-critical applications such as human-robot collaboration and intelligent monitoring. Existing methods primarily focus on improving prediction accuracy but neglect the inherent uncertainty in early observations, leading to systems that cannot distinguish between reliable predictions and uncertain guesses. This paper proposes the UncerTrans framework, which combines Temporal Transformer with Monte Carlo Dropout to achieve accurate and trustworthy early action prediction. Temporal Transformer extracts discriminative features from extremely short observation sequences through hierarchical temporal attention mechanisms and temporal decay positional encoding. Monte Carlo Dropout generates prediction distributions through multiple random forward propagations during inference to quantify the model's epistemic uncertainty. An adaptive sampling strategy is designed to dynamically adjust sampling frequency based on initial uncertainty, balancing prediction quality and computational efficiency. Experiments on the EPIC-KITCHENS-100 dataset demonstrate that UncerTrans achieves 65.5% accuracy with only 10% observation ratio and an Expected Calibration Error of merely 0.089, significantly outperforming baseline methods. Through selective rejection of high-uncertainty predictions, the system can improve the accuracy of remaining predictions to 84.2%. The research demonstrates that effective uncertainty quantification relies on high-quality feature extraction, and the combination of both components enables early prediction systems to adopt differentiated strategies based on confidence levels, providing a technical foundation for practical deployment.

As modernization and industrialization continue to accelerate, air pollution has become an increasingly pressing problem. Air quality prediction is considered an essential technical support for air pollution prevention and control. To achieve more accurate predictions for urban air pollution, we propose a hybrid model called CBLA, which consists of three parts: one-dimensional Convolutional Neural Networks (1D-CNNs), Bidirectional Long Short-Term Memory network (BiLSTM), and attention mechanism. Firstly, 1D-CNNs extract the deep features of the original data. Secondly, BiLSTM mines time-series features for initial prediction. Finally, the attention mechanism captures the effect of characteristic conditions on PM2.5 concentration at different times to further optimize the model. The eXtreme Gradient Boosting (XGBoosting) tree is used to integrate the preliminary prediction results and meteorological data to improve prediction accuracy further. We conducted extensive experimental evaluations using Beijing's air quality and meteorological datasets, which showed that the CBLA model has excellent performance and model expression power.

The development of reliable sensing platforms for synthetic food additives remains a critical challenge due to severe matrix interferences that limit selectivity and analytical accuracy. In this work, a multiphysics-guided framework is employed to design a ZIF-67/MWCNT-modified glassy carbon electrode (GCE) for the highly selective electrochemical detection of sunset yellow (SY) in complex food matrices. By integrating experimental electrochemical analysis with COMSOL-based modeling of mass transport, adsorption dynamics, charge transfer, and thermal effects, this study provides a mechanistic basis for material–analyte interactions that govern sensor performance. The ZIF-67/MWCNT hybrid exhibits synergistic surface chemistry, where π–π stacking between the azo-aromatic structure of SY and the graphitic domains of MWCNTs, together with electrostatic interactions with Co2+ centers in ZIF-67, yields a high adsorption constant (Kads = 5.41 × 104 m3 mol−1) and a dominant surface flux (3.47 × 10−7 mol m−2 s−1), surpassing those of common interferents. The optimized electrode delivers a steady-state current density of 5.22 µA m−2 at pH 7 and a 5 µm composite layer, while maintaining negligible faradaic contributions from ascorbic acid, citric acid, aspartame, and acesulfame potassium. Parametric simulations reveal robust performance under thermal variations (298–328 K), minimal sensitivity to electrolyte disturbances, and a direct correlation between surface heterogeneity and current attenuation. Model validation against experimental electrochemical impedance spectroscopy yields a low RMSE (0.0621), confirming predictive accuracy. These findings demonstrate how multiphysics analysis can rationally guide electrode engineering, offering a powerful design strategy for next-generation electrochemical sensors. The proposed platform provides a selective, sensitive, and scalable solution for trace-level SY detection, underscoring its relevance for food safety monitoring and real-sample analysis.

Preoperative differentiation of precursor glandular lesions (PGL), minimally invasive (MIA), and invasive adenocarcinoma (IAC) in stage IA lung adenocarcinoma (LUAD) is critical for surgical planning but remains challenging due to overlapping CT features and interobserver variability. While existing artificial intelligence (AI) models focus predominantly on binary classification with limited multicenter validation, this study developed and validated a ternary classification framework using pretrained TabPFN and traditional machine learning (ML) algorithms based on AI-derived histogram features, benchmarking against intraoperative frozen section analysis. This multicenter retrospective study utilized preoperative CT scans from three institutions between September 2014 and October 2023. Data were divided into training, internal validation, and external test sets. Histogram features (n =26) were automatically extracted using a commercial AI system (InferRead CT Lung). TabPFN and five ML algorithms were trained with selected clinical and histogram features. Performance was evaluated by accuracy, macro-AUC, sensitivity, specificity, and Cohen's Kappa. Statistical comparisons included DeLong tests for AUC and chi-square for categorical variables. The cohort comprised 584 stage IA LUAD patients (mean age 57.9±11.0years; 386 female), divided into training/validation sets (n =412, center 1) and external test sets (n =114, center 2; n =58, center 3). TabPFN achieved macro-AUC of 0.781-0.911 and accuracy of 67.2-78.9% across external test sets, outperforming other ML algorithms. Of note, TabPFN achieved an overall better prediction accuracy compared to frozen section analysis on all test sets (internal: 92.3% vs 84.6%, P =0.503; external 1: 87.5% vs 75%, P =1.000; external 2: 67.2% vs 43.1%, P <0.001). Subgroup analysis revealed superior performance for mGGN lesions (85%) on both external test sets. TabPFN enables robust, generalizable ternary classification of LUAD subtypes, surpassing conventional ML and frozen section analysis. Its integration with automated histogram analysis offers a scalable solution for preoperative stratification of early-stage lung cancer.

Accurate cardiovascular disease (CVD) risk prediction is crucial for personalized preventive medicine. To evaluate the effect of traditional risk factors on CVD mortality, and to validate the prognostic performance of the SCORE2 model for prediction of 18-year risk of CVD death. The cohort study included 6780 residents of Krakow (54% women), free of CVD and diabetes at baseline, recruited between 2002 and 2005. The mean baseline age of men and women was 57.2 (6.9) years and 56.6 (6.8) years respectively. In 50 246 and 62 906 person-years of follow-up in men and women, respectively, cumulative CVD mortality was 9.0% in men and 5.2% in women, with competing (non-CVD) mortality of 19.2% and 12.0%. Smoking and hypertension were strongly associated with CVD mortality, whereas associations with total cholesterol (TC) and HDL cholesterol (HDL-C) were weaker. A newly derived model including age, smoking, systolic blood pressure (SBP), TC and HDL-C achieved Harrell's C-Index of 0.693 in men and 0.757 in women. The model including SCORE2 as a continuous variable showed similar discrimination (C-index 0.718 in men and 0.754 in women), while SCORE2 categories demonstrated poorer predictive performance (C-index 0.589 and 0.676, respectively). Smoking and elevated blood pressure were confirmed as major long-term predictors of CVD mortality. The prognostic performance of SCORE2 as a continuous measure was good and comparable to the derived prediction model including age and traditional risk factors. However, the predictive accuracy of SCORE2 categories was lower, particularly in men.

Recurrent ischemic stroke (RIS) is a significant challenge in Malaysia, affecting approximately 33% of patients. However, studies using artificial intelligence (AI) to predict this event using real-world data remain very limited. This study aimed to develop and evaluate Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and RUSBoost models for predicting recurrent ischemic stroke using real-world data from the Malaysian National Neurology Registry. We established a retrospective study of 7,697 enrolled patients registered in the National Neurology Registry in Malaysia (2009-2016). We developed and evaluated several machine learning models, including SVM, KNN, and RUSBoost, to predict recurrent RIS. The Synthetic Minority Over-Sampling Technique (SMOTE) was applied to the training data to handle the imbalanced data. Ten-fold cross-validation was applied to assess the robustness and accuracy of the models, and performance was evaluated using criteria of accuracy, sensitivity, specificity, PPV, and area under the ROC curve (AUC). Among the evaluated machine learning models, RUSBoost demonstrated the strongest and most clinically relevant performance when assessed on validation (test) folds under stratified ten-fold cross-validation, achieving an AUROC of 0.943, sensitivity of 86.5%, and a favourable balance between sensitivity and PPV of 40.2% on the original imbalanced dataset. Although the application of SMOTE during training improved model discrimination for RUSBoost (training-fold AUROC = 0.986). The SHAP analysis showed that age, race, glucose level, hypertension, hyperlipidemia, and duration of diabetes were the most significant factors linked to an increased risk of recurrent ischemic stroke. This study demonstrates that applying machine learning models on real-world clinical data is a promising tool for predicting the risk of ischemic stroke recurrence. RUSBoost emerged as the most reliable and generalisable model for clinical risk prediction, proved effective in improving prediction accuracy and identifying patients at highest risk. While SMOTE enhanced model learning during training. The findings highlight the importance of integrating AI technologies into clinical practice to support early treatment decisions and enhance preventive interventions, opening new pathways for better patient care and reducing the health burden from recurrent stroke.

The project cycle of complex construction projects covers the whole process from project decision-making, design, bidding, construction, completion acceptance, and the initial stage of operation. Among them, the investment risk assessment of complex construction projects focuses on the early decision-making stage of the project, aiming to provide a basis for investment feasibility analysis. The investment risk of complex construction projects is highly nonlinear and uncertain, and the traditional risk assessment methods have limitations in model generalization ability and prediction accuracy. To improve the accuracy and reliability of quantitative risk assessment, this study proposed a novel investment risk assessment model based on the perspective of investors. Firstly, through literature research, a multi-dimensional comprehensive risk assessment index system covering policies and regulations, economic environment, technical management, construction safety, and financial cost was systematically identified and constructed. Subsequently, the Least Squares Support Vector Machine (LSSVM) was used to establish a nonlinear mapping relationship between risk indicators and final risk levels. Aiming at the problem that the parameter selection of the standard LSSVM model has a significant impact on the performance, this paper proposed an improved Firefly Algorithm (IFA) to automatically optimize the penalty factor and kernel function parameters of LSSVM, so as to overcome the blindness of artificial parameter selection and improve the convergence speed and generalization ability of the model. Compared with the classical Firefly Algorithm, IFA strengthens learning and adaptive strategies by adding depth. The conclusions are as follows. (1) Compared with the Backpropagation Neural Network (BPNN), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost), this model showed higher prediction accuracy on the test set, and its accuracy was reduced by about 3%. (2) Compared with FA, Genetic Algorithm (GA), and Particle Swarm Optimization (PSO), IFA had a stronger global retrieval ability. (3) The model could effectively fit the complex risk nonlinear relationship, and the risk assessment results were highly consistent with the actual situation. Therefore, the risk assessment model based on the improved LSSVM constructed in this study not only provides a more scientific and accurate quantitative tool for investment decision-making of construction projects, but also has important theoretical and practical significance for preventing and resolving significant investment risks.

In regard to shaping up a complex-shaped, largely strained industrial forming part, sheet hydroforming (SHF) is one of the primary processes utilized to address these challenges. Before actual tooling fabrication, the finite element (FE) simulation is, nowadays, commonly employed to assess the feasibility of the process and tooling design. Therein, material modelling, especially of the distinguishing deformation anisotropy unavoidable in cold rolled sheet metal, plays a vital role. This study, therefore, seeks to enhance the capability of the FE simulation on sheet hydroforming of an SPC270 mild steel sheet comparatively through the von Mises, Hill’48, and Yld2000-2d yield criteria. Additionally, the hybrid Swift-Voce (HSV) model is applied to refine and extend the experimentally determined uniaxial flow stress curve. The prediction accuracy is evaluated on the basis of two geometrical deviations such as the sheet thickness distribution and tank surface profile. The results show that the Yld2000-2d yield model obviously leads to the most accurate geometric estimation for both evaluation criteria.

A composite of polyvinyl alcohol (PVA) and beta-tricalcium phosphate (β-TCP) was synthesized as a biomaterial filament for 3D printers and its analytical and chemical evaluation was performed. PVA powder and β-TCP were mixed in the range of 0-20 wt% and hot-melt extruded at 200 °C using a single-screw extruder. Comprehensive material characterization of the synthesized filament was performed by powder X-ray diffraction (XRD), near-infrared spectroscopy (NIR), and scanning electron microscopy (SEM). XRD analysis confirmed that the amorphous nature of PVA and the crystalline nature of β-TCP coexisted and the physical mixture state was well maintained. In near-infrared spectroscopy, concentration-dependent spectral changes were observed by normalization, and principal component analysis showed that the first principal component explained 85.6% of the variance. In machine learning regression analysis, partial least squares regression (PLS), random forest (RF), and support vector machine (SVM) were compared, and SVM achieved the best prediction accuracy (R2 = 0.910). SEM observations confirmed streaky structures along the extrusion direction and uniform dispersion of β-TCP particles. This study demonstrated that a combined NIR spectroscopy and machine learning approach is effective as a non-destructive quality evaluation technique for composite filaments for 3D printing. This technique enables real-time composition monitoring and quality control of biomaterial filaments, and is expected to be applied to the manufacturing of patient-specific biomedical devices.