How microplastics affect nitrogen removal in nature-based stormwater infrastructures: A machine learning and meta-analysis study.

A low-pressure MEMS piezoresistive pressure sensor using multilayer perceptron neural network for structure optimization

Pathological microenvironments linked to aging, trauma, malignancies, and metabolic disorders significantly hinder bone fractures and frequently result in fracture nonunion, posing substantial worldwide clinical difficulties. Widely prevalent therapies encounter difficulties in addressing diverse anatomical defects and variable illness conditions due to their inflexible designs and limitations in empirical optimization. Efficient strategies are critical to restore mechanics, improve pathological microenvironments, enhance neovascularization, and adapt to anatomical defects and clinical conditions. Deep learning networks (DLN) excel at analyzing extensive nonlinear relationships, enabling predictions of biomaterial‑biological interactions, hence accelerating biomaterial development. This study presents a synergistic DLN and 4D printing approach to fabricate a microenvironment-adaptive bioactive scaffold (MABS) for enhanced osteogenesis and angiogenesis. The scaffold integrates bioactive glass and a shape-memory PgP matrix, with a multilayer perceptron (MLP) neural network optimizing its design via nonlinear parameter-performance analysis. In vivo investigations revealed that the DLN-optimized scaffold enhanced shape-morphing adaptability and promoted the formation of dense bone tissue and vascular networks. This paradigm shift-employing DLN to integrate 4D printing dynamics, degradation kinetics, and multi-scale biological responses-transforms bone implants from static entities to dynamically adaptive systems, offering a scalable, intelligent framework for precise bone repair that rectifies the deficiencies of current strategies.

ABSTRACT Driven by significant concerns about the risk of ship collisions within the maritime industry, this paper aims to propose an enhanced and intelligent method for assessing collision risks utilizing deep learning techniques. A novel research framework is introduced for assessing potential collision risk of ships entering monitored waterways by leveraging the Automatic Identification System (AIS) data, deep learning methods, and expert knowledge. Specifically, this framework includes 1) the selection of multifaceted ship collision assessment indicators. 2) The development of an intelligent model for risk assessment under limited data. A deep learning model, based on variational autoencoder (VAE) and incorporating multilayer perceptron (MLP) and convolutional neural network (CNN), is proposed. The MLP is employed to upsample features and address issues related to small sample sizes and variable correlations, while the CNN serves as the feature extractor. To improve iterative effectiveness, a novel loss function is introduced. Comparative experiments have shown that the proposed model outperforms existing baselines in predictive performance. By applying the proposed method in the Yangtze Estuary Deepwater Channel, the study reveals several findings and provides recommendations. Results indicate that the most critical risk indicators are the degree of ship course deviation, ship size, and ship speed.

Accurate pre-harvest yield estimation of underground bulb crops such as onion and garlic is important for precision agriculture, harvest planning, and food-security-oriented decision-making. However, their harvestable organs develop below ground and cannot be directly observed using conventional remote sensing methods. This study aimed to develop a non-destructive yield estimation framework by integrating UAV-based hyperspectral imaging with hybrid machine learning models. Field experiments were conducted in Muan-gun, Korea, using onion and garlic as representative underground bulb crops. UAV-based hyperspectral images, crop growth traits, and destructive live bulb weight measurements were collected during the growing period. Hyperspectral images were processed through geometric correction, radiometric correction, and Savitzky-Golay spectral smoothing. Three dimensionality reduction methods, including genetic algorithm (GA), principal component analysis (PCA), and clustering, were used to reduce spectral redundancy. Five prediction models, including random forest (RF), XGBoost, partial least squares regression (PLSR), multilayer perceptron (MLP), and residual network (ResNet), were then evaluated for live bulb weight prediction. Significant spectral differences were observed in the 550-680nm and 730-800nm bands, which were closely associated with crop yield and below-ground bulb development. GA was the most effective feature selection method for extracting yield-related spectral bands. For onion yield prediction, the GA + RF model achieved the highest predictive accuracy, with an R2 of 0.9656 and an NRMSE of 18.55%. For garlic yield prediction, PLSR showed the best performance, with an R2 of 0.9260 and an NRMSE of 27.20%. The proposed UAV-based hyperspectral framework enables accurate, real-time, and non-destructive yield estimation for underground bulb crops. This approach reduces reliance on labor-intensive destructive sampling and provides a practical tool for precision crop monitoring and data-driven agricultural management.

Steel is an ideal recycling material as it can be recycled almost indefinitely, and steel recycling is a lot more energy-efficient than iron ore steel production. Steel making with usage of steel scrap in electric arc furnaces is heavily influenced by contaminants in the scrap, including non-ferrous metals, stones, or plastic. To produce high-quality steel, it is important to know what contaminants are part of the scrap to adapt the recycling process accordingly. This work presents a three-part processing pipeline to determine the scrap composition and optimise the recycling process parameters. The first part is hyperspectral imaging, recording images with 437 spectral bands in the short-wave infrared range. Next, deep learning-based image recognition, with multilayer perceptron, 2D and 3D convolutional neural networks (CNNs) have been compared. The 3D-CNN showed the best performance in detecting the 14 material classes in the scrap samples. Finally, a mixed integer optimisation is used to select the best scrap mix for the steel classes that are to be produced. The evaluation shows an accuracy of about 76% in detecting the 14 material classes correctly, with a higher accuracy for the steel class. The detected non-ferrous materials determine the scrap class which is input for the optimisation with its constraints that each scrap class has a minimum consumption to prevent storage overflows, and as little energy and additives as possible should be used.

Accurate modeling of laryngeal motor control is key to understanding typical and disordered voice production. However, traditional biomechanical plant models based on ordinary differential equations (ODEs) often involve high computational costs and numerical instabilities, limiting their use in real-time closed-loop control frameworks. This study evaluates feature-driven machine learning (ML) regressors, specifically Random Forest (RF), Multilayer Perceptron Neural Networks (NN), and Polynomial Regression (PR), as surrogate forward models mapping laryngeal motor inputs to fundamental frequency and sound pressure level. Training data were generated with two biomechanical vocal fold models: the extended body-cover and the triangular body-cover. Results demonstrate that ML surrogates reduce execution times from seconds to milliseconds (e.g., 2 ms for PR), enabling stable real-time tracking via inverse Jacobian control. While RF provides the highest accuracy, NN and PR offer smoother control signals and smaller memory footprints. A practical performance threshold was identified near N = 1,000 training samples, below which accuracy degraded substantially when models were trained from scratch. These findings support ML surrogates as efficient and adaptable alternatives to direct numerical simulation, providing a foundation for future subject-specific modeling through transfer learning in data-limited clinical scenarios.

The essential role of electricity supply for public and private services highlights the need to monitor the stability of power transmission networks during, or immediately after, hazardous events. In the aftermath of calamities, traditional field inspections may be impractical or unsafe, leaving operators without timely information on the condition of critical assets. In this paper, we present and discuss the performance of two automatic Artificial Intelligence (AI)-based models (Multi-Layer Perceptron (MLP) neural network architectures and Support Vector Machine (SVM) model) designed to automatically assess the status of high-voltage transmission towers and power lines through multi-temporal spaceborne Synthetic Aperture Radar (SAR) image analysis. Model development and testing rely on real COSMO-SkyMed Stripmap observations of damaged towers and power lines affected by documented hazardous events across Italy, complemented by simulated tower data generated with a physics-guided, signature-based SAR simulator designed to preserve the observed target-to-background contrast and spatial footprint patterns of real SAR tower signatures. Results indicate that the MLP, trained on either real or simulated data, achieved 100% Overall Accuracy (OA) with no observed false positives or false negatives within the considered visibility-screened real test set, while providing inference times on the order of tenths of milliseconds per target… Computational performance characteristics, operational advantages, and the potential pathway toward satellite on-board porting are discussed to enhance situational awareness and support the prioritisation of interventions during critical events.

Recently, there has been an increase in the grid integration of electric vehicles (EVs) and solar photovoltaic (PV) systems, primarily driven by two goals: lowering energy costs and decreasing emissions. Numerous research studies have concentrated on the separate effects of integrating PVs and EVs into the grid. Nevertheless, it is important to recognize that as the adoption of PVs and EVs continues to grow, the supply grid will face the cumulative effects of PV and EV integration on power quality (PQ) challenges. To provide a comprehensive understanding, this study examines the joint impact of PVs and EVs on PQ aspects in detail. This study has indicated that EVs and PVs alone can adversely impact grid reliability and PQ because of the variable character of PV source and the unpredictability of EV demand. But multiple research efforts have shown that coordination between PVs and EVs can help to alleviate certain problems that arise from their individual integration. This study demonstrates PQ enhancement in a grid system integrated with PV and EV using a multilayer perceptron neural network (MLPNN) approach. In the system with PV integration, the GWO-ANFIS, MPPT technique is employed for optimizing power extraction. Under balanced non-linear loading conditions, simulation results show that the THD is initially 25.97% without compensation, then decreases to 12.57% with a shunt passive filter (SPF), 3.37% with the application of recursive least squares (RLS), and 1.37% with MLPNN. With much lower THD and quicker convergence, the suggested MLPNN-based controller exhibits improved harmonic mitigation. A comparison between the proposed and existing methods are drawn using the MATLAB/Simulink platform.

To enhance the control performance of permanent magnet synchronous motors (PMSM) under complex operating conditions such as sudden load changes and parameter perturbations caused by strong external disturbances, this paper proposes an incremental model predictive control (IMPC) strategy based on a disturbance observer (DOB), referred to as DOB-IMPC. Firstly, an incremental predictive model of PMSM is constructed, taking advantage of its inherent integral characteristic to improve the system's robustness. Secondly, a DOB is introduced to estimate the aggregated disturbances, including load disturbances and parameter uncertainties in real time, and the estimated values are fed forward to compensate for the predictive model. Subsequently, an internal optimization process of IMPC is designed based on this model, and the optimal control input sequence is obtained by solving a constrained quadratic programming problem. For the parameter setting of the IMPC itself and the additional tuning burden introduced by the DOB, conventional approaches relying on empirical expertise and repeated trial-and-error are inefficient and cannot readily guarantee optimal control performance. This paper develops a parameter-tuning mechanism based on a multi-layer perceptron (MLP) neural network, which integrates the real-time state variables of the PMSM with the overall structural information of the DOB-IMPC framework to achieve online tuning of key control parameters. Simulation and experimental results demonstrate that, compared with conventionally manually tuned MPC, the MLP-based tuning strategy reduces the speed root mean square error (RMSE) by 46.2% and the maximum speed fluctuation by 29.0% while maintaining zero overshoot. With the further incorporation of the DOB, the maximum speed fluctuation under sudden load disturbance is further reduced by 31.2% compared with MPC, and the speed RMSE under parameter perturbation conditions is reduced by 75.8%. Collectively, these results confirm that the proposed strategy delivers measurable gains in dynamic response, disturbance rejection, and parametric resilience-without compromising stability or implementation feasibility.

The classification of plant varieties is a key task in plant breeding and variety registration. Red fescue (Festuca rubra L.), a widely cultivated grass species, includes numerous closely related varieties, making automated classification a challenging multi-class problem. This study aimed to develop and evaluate a multilayer perceptron (MLP) neural network combined with a subset-based decision framework for accurate classification of red fescue varieties and recognition of previously unseen varieties. The study analyzed 76 varieties described by seven morphological features. To address the complexity of the multi-class problem, the dataset was divided into multiple subsets and the effectiveness of different partitioning strategies was evaluated. A confidence-based and majority-based decision rule (majority ratio ≥ 0.9 and mean Softmax confidence ≥ 0.8) was introduced to improve the reliability of final predictions and enable open set recognition. The model was evaluated using accuracy, precision, F1 score, and recall. The most optimal solution was to divide the dataset into 15 subsets, with the first subset containing six varieties and the remaining subsets containing five varieties each. This approach provided the best balance between predictive performance and decision consistency, enabling correct classification of known varieties and stable detection of unknown samples. Combining MLP neural networks with strategic subset division and confidence-driven decision rules offers a robust solution to high-dimensional, multi-class classification challenges in plant variety recognition. The model's ability to recognize new varieties is crucial for its practical application, ensuring the algorithm's flexibility. This is particularly useful in agriculture and horticulture, where new varieties are bred over the years.

Autism Spectrum Disorder (ASD) affects 1 in 100 children globally, yet early detection remains challenging due to reliance on subjective behavioral assessments and limited specialist availability, particularly in resource-constrained settings. Eye-tracking biomarkers offer objective, quantitative measurements of visual attention patterns, but existing Neural Network approaches lack systematic evaluation across diverse neural network architectures and multi-site data. This study provides a comprehensive comparison of Graph Neural Networks, Long Short-Term Memory networks, and Multilayer Perceptrons for ASD detection using multimodal eye-tracking features. We combined two independent datasets: CILIA (57 participants from France, structured viewing paradigm) and Saliency4ASD (27 participants from Italy, free-viewing paradigm), creating a diverse multi-site European cohort of 84 participants (40 ASD, 44 typically developing). Four neural network architectures were evaluated: Graph Neural Networks (GNNs) with attention mechanisms, GNN without attention, bidirectional Long Short-Term Memory (LSTM), and Multi-Layer Perceptron (MLP) networks. Eye-tracking features included fixation patterns, saccade dynamics, pupil diameter, and spatial attention distributions. Leave-one-subject-out cross-validation ensured robust generalization. Statistical comparisons employed DeLong tests for area under the ROC curve differences, McNemar tests for accuracy, and bootstrap confidence intervals (10,000 iterations). Per-dataset performance analysis assessed generalization across acquisition protocols. LSTM achieved the highest AUC of 0.859 (95% CI: 0.770-0.934), followed by GNN with attention (AUC = 0.855, 95% CI: 0.763-0.931). Contrary to expectations, attention mechanisms did not significantly improve GNN performance (GNN without attention: AUC = 0.823, p = 0.102). LSTM significantly outperformed MLP (AUC = 0.811, p = 0.002) but showed no significant difference versus GNN with attention (p = 0.749). GNN with attention demonstrated the highest specificity (93.2%) and positive predictive value (90.6%). Performance remained consistent across both datasets (LSTM: AUC = 0.840 on CILIA versus 0.956 on Saliency4ASD), supporting generalization across datasets despite different viewing paradigms, equipment, and cultural populations. All models exhibited large effect sizes (Cohen's d exceeding 1.0) for discriminating ASD from typically developing predictions. This systematic evaluation demonstrates that LSTM and graph-based approaches achieve equivalent performance for ASD detection from eye-tracking biomarkers, with attention mechanisms providing no significant benefit. The high specificity (93.2%), real-time capability, and cross-dataset generalization make these models viable for scalable, objective ASD screening in clinical settings, potentially enabling earlier intervention in resource-constrained environments.

Machine learning-based models for estimating the rich amine acid gas loading based on industrial data

Measurements of the mean free path of ionizing photons from high-redshift quasar spectra at z ∼ 5-6 constrain the reionization history, but interpreting them requires modeling the kiloparsec-scale clumping that large-volume reionization simulations cannot resolve. We present a deep learning emulator for the mean free path (MFP) trained on high-resolution cosmological radiative transfer simulations of ionization fronts sweeping through small 2 comoving Mpc/h volumes. Using a residual multi-layer perceptron neural network, we predict the MFP at a given redshift as a function of the reionization redshift, photoionization rate, wavelength, and box-scale density, achieving a median relative error of 1.3% across nearly four orders of magnitude in MFP. Integrating its predictions over box-scale overdensity and an extended reionization history allows the emulator to predict the global MFP. We apply the emulator to extended reionization histories constrained by observed photoionization rates, finding that models prefer late reionization with substantial neutral fractions persisting at z ≲ 6. Fitting a parametric ionization history yields a midpoint of reionization of z re = 6.58 ± 1.2 for reionization durations consistent with Planck and kinetic Sunyaev-Zeldovich constraints, and the universe being 10% neutral still at z < 5.8 (6.3) at 1 (2)σ. Global ionizing emissivity inferences using measurements of the photoionization rate and MFP plus our emulator, which avoids common power-law assumptions, suggest a factor of 2-3 decline between z = 6 and 4.8, in agreement with previous studies. Our method provides an efficient (and more converged) alternative to large-volume radiative-hydrodynamic simulations of reionization for interpreting MFP measurements, and can also serve as a subgrid prescription for the ionizing opacity within such simulations.

Predicting soymilk odors using a multilayer perceptron neural network model.

Active disturbance rejection control with neural network-based ESO for gas turbine control system by fractional fuzzy-PSO optimization.

Predicting the impact sensitivity of energetic materials is crucial for safe handling yet remains challenging due to complex molecular interactions. In this study, quantitative structure–property relationship (QSPR) models were developed to predict the impact sensitivity (logH50) of 404 nitro compounds using multilayer perceptron (MLP) and one-dimensional convolutional neural network (1D-CNN). The MLP model based on 24 Dragon descriptors, achieved a coefficient of determination R2 of 0.853, root mean square (rms) error of 0.151 and mean absolute error (MAE) of 0.122 for the test set. The CNN model, employing 17 Dragon descriptors, attained test R2 = 0.843, rms = 0.155, and MAE = 0.121, with 99.0% of predictions having standardised residuals within ±3, indicating reliable performance for structurally diverse compounds. Both models satisfy all QSPR acceptance criteria, ranking among the best-performing models for this endpoint. Mechanistic interpretation revealed that impact sensitivity is governed by electronic distribution, molecular bulk, polarity, specific functional groups, and hydrogen bonding interactions. This work provides accurate predictive tools and physicochemical insights for designing safer nitro compounds.

Extensive research has been conducted on vehicle queuing and dissipation near signalized intersections. However, existing prediction methods for vehicle departure time primarily rely on assumptions of steady-state homogeneous traffic flow, utilizing shockwave theory and vehicle kinematic modeling. These methods encounter challenges in addressing traffic uncertainties during queue formation and dissipation, particularly in scenarios involving multiple lanes. This paper introduces a novel approach by leveraging unmanned aerial vehicle (UAV) trajectory data to construct fleet state features and proposes a prediction method for tail vehicle departure time based on Gaussian process regression. The objective of this method is to optimize the green light crossing time window and eco-driving trajectory for connected vehicles at signalized intersections. The findings reveal that the departure time of the tail vehicle within a specified distance adheres to a Gaussian process, demonstrating the applicability of Gaussian process regression for departure time prediction modeling. The effectiveness of the proposed method was validated using a field-measured dataset collected from three typical multi-lane signalized intersections. Notably, compared to four benchmark models (linear regression, decision trees, multilayer perceptron neural networks, and eXtreme Gradient Boosting—XGBoost), the mean absolute percentage error (MAPE) was reduced by an average of 5.146% on the test set under a random 70/30 split. Additionally, a robustness assessment demonstrates that the proposed model performs well, albeit slightly less effectively than the XGBoost model. We emphasize that the conclusions are drawn for the studied intersections; generalization to unseen intersections requires further validation with cross-site data.

AimsEmergency department overcrowding, especially in cardiac units, delays care and raises mortality. Conventional triage is error-prone. We developed an AI-based model integrating routine data and automated ECGs to improve early risk classification.Methods and resultsThis retrospective cross-sectional study involved 600 medical records of patients presenting with suspected cardiac symptoms. Model development was conducted in three phases: Designing a triage model using routine triage data, designing a triage model based on ECG images, and combining the ECG-based model and triage data. Model performance was evaluated regarding standard clinical outcomes within the first 24 h and compared against the Emergency Severity Index. The best-performing model based on triage data alone (i.e. multilayer perceptron neural network) yielded an accuracy of 89.42%, F-score of 84.51, and area under the curve between 0.815 and 0.858. The best-performing model based on ECG interpretation alone (i.e. convolutional neural network) yielded an accuracy of 93.83%, F-score of 91.08, and area under the curve ranging from 0.852 to 0.914. The fusion model demonstrated superior performance, with an accuracy of 97.22%, F-score of 94.60, and area under the curve between 0.881 and 0.938—significantly outperforming the conventional Emergency Severity Index. In the fusion model, the key predictive variables included ECG interpretation, heart rate, and mode of entry to emergency department.ConclusionGiven its advantages over models using only routine data, ECG, or conventional triage, the fused AI-based triage model may effectively prioritize and predict cardiac emergency outcomes, providing a foundation for developing reliable, intelligent support systems in acute care.

Mitochondrial metabolism (MM) abnormalities have been implicated in multiple diseases, but its contribution to sepsis-associated encephalopathy (SAE) remains insufficiently understood. The purpose of this research was to explore the candidate biomarkers associated with MM in SAE and the underlying mechanisms. The relevant transcriptome data were acquired from the public databases. MM-related genes (MM-RGs) were scoured from published articles. Candidate biomarker identification integrated differential expression profiling, PPI network construction, machine learning algorithms, ROC curve evaluation, and expression quantification. Immune infiltration, multilayer perceptron (MLP) network, Gene Set Enrichment Analysis (GSEA), molecular regulatory network, drug prediction, and molecular docking were applied to probe the mechanisms of candidate biomarkers in SAE. The expression levels of candidate biomarkers were further validated in brain tissues of the LPS induced SAE mouse model. Four MM-related genes-INSIG1, SREBF1, CIDEC, and PNPLA3-were identified as candidate biomarkers, and their expression patterns in the rodent model were consistent with bioinformatics predictions. Within the GSE135838 dataset, the MLP model demonstrated good performance in distinguishing SAE from control samples. Differential immune cells included resting mast cells, naïve B cells, and activated natural killer (NK) cells. Gene set enrichment analysis (GSEA) indicated significant enrichment in the allograft rejection pathway. The regulatory network involved transcription factors (e.g., THRB), miRNAs (e.g., hsa-mir-29c-3p), and lncRNAs (e.g., KCNQ1OT1). Furthermore, experimental validation confirmed that these candidate biomarkers were significantly downregulated in SAE samples. This study identified and experimentally validated four MM-related candidate biomarkers in SAE, revealing their potential roles in immune dysregulation. These candidate biomarkers require further validation in independent cohorts and in a more diverse range of animal models before clinical translation.