Efficient Model Research Articles

Tracer-Separator: A Deep Learning Model for Brain PET Dual-Tracer (18F-FDG and Amyloid) Separation.

Multiplexed PET imaging revolutionized clinical decision-making by simultaneously capturing various radiotracer data in a single scan, enhancing diagnostic accuracy and patient comfort. Through a transformer-based deep learning, this study underscores the potential of advanced imaging techniques to streamline diagnosis and improve patient outcomes. The research cohort consisted of 120 patients spanning from cognitively unimpaired individuals to those with mild cognitive impairment, dementia, and other mental disorders. Patients underwent various imaging assessments, including 3D T1-weighted MRI, amyloid PET scans using either 18F-florbetapir (FBP) or 18F-flutemetamol (FMM), and 18F-FDG PET. Summed images of FMM/FBP and FDG were used as proxy for simultaneous scanning of 2 different tracers. A SwinUNETR model, a convolution-free transformer architecture, was trained for image translation. The model was trained using mean square error loss function and 5-fold cross-validation. Visual evaluation involved assessing image similarity and amyloid status, comparing synthesized images with actual ones. Statistical analysis was conducted to determine the significance of differences. Visual inspection of synthesized images revealed remarkable similarity to reference images across various clinical statuses. The mean centiloid bias for dementia, mild cognitive impairment, and healthy control subjects and for FBP tracers is 15.70 ± 29.78, 0.35 ± 33.68, and 6.52 ± 25.19, respectively, whereas for FMM, it is -6.85 ± 25.02, 4.23 ± 23.78, and 5.71 ± 21.72, respectively. Clinical evaluation by 2 readers further confirmed the model's efficiency, with 97 FBP/FMM and 63 FDG synthesized images (from 120 subjects) found similar to ground truth diagnoses (rank 3), whereas 3 FBP/FMM and 15 FDG synthesized images were considered nonsimilar (rank 1). Promising sensitivity, specificity, and accuracy were achieved in amyloid status assessment based on synthesized images, with an average sensitivity of 95 ± 2.5, specificity of 72.5 ± 12.5, and accuracy of 87.5 ± 2.5. Error distribution analyses provided valuable insights into error levels across brain regions, with most falling between -0.1 and +0.2 SUV ratio. Correlation analyses demonstrated strong associations between actual and synthesized images, particularly for FMM images (FBP: Y = 0.72X + 20.95, R2 = 0.54; FMM: Y = 0.65X + 22.77, R2 = 0.77). This study demonstrated the potential of a novel convolution-free transformer architecture, SwinUNETR, for synthesizing realistic FDG and FBP/FMM images from summation scans mimicking simultaneous dual-tracer imaging.

A novel tetra hybrid bio-nanofluid model with stenosed artery

Abstract For treating and diagnosing cardiovascular diseases, the field of biomedical engineering is significant because it develops new ways and techniques. Stenosis is the narrowing of an artery, and it leads reduction in the flow rate of blood. This study investigates the blood flow mechanism in an artery using a mathematical model of Carreau nanofluid with four distinct nanoparticles. Tetra nanofluid model produces significant advancement in the simulation of blood flow through the stenosed arteries. The model is capable of predicting the pressure drop and velocity distribution for diagnosing and treating stenosis. The spectral relaxation approach is used to present the model's efficiency and effectiveness, which makes it a suitable method for solving the governing equations of this study. The findings of this study have important implications for the development of new treatments and diagnostic techniques for stenosis and other cardiovascular diseases.

Sequential lexicon enhanced bidirectional encoder representations from transformers: Chinese named entity recognition using sequential lexicon enhanced BERT

Lexicon Enhanced Bidirectional Encoder Representations from Transformers (LEBERT) has achieved great success in Chinese Named Entity Recognition (NER). LEBERT performs lexical enhancement with a Lexicon Adapter layer, which facilitates deep lexicon knowledge fusion at the lower layers of BERT. However, this method is likely to introduce noise words and does not consider the possible conflicts between words when fusing lexicon information. To address this issue, we advocate for a novel lexical enhancement method, Sequential Lexicon Enhanced BERT (SLEBERT) for the Chinese NER, which builds sequential lexicon to reduce noise words and resolve the problem of lexical conflict. Compared with LEBERT, it leverages the position encoding of sequential lexicon and adaptive attention mechanism of sequential lexicon to enhance the lexicon feature. Experiments on the four available datasets identified that SLEBERT outperforms other lexical enhancement models in performance and efficiency.

Predictive modeling of membrane reactor efficiency using advanced artificial neural networks for green hydrogen production

The imperative to decarbonize the energy sector has prompted substantial advancements in clean electricity generation, with hydrogen emerging as a promising low-carbon energy carrier. While hydrogen synthesis from renewable sources is crucial, challenges persist, necessitating innovative approaches for efficient and sustainable production. This study leverages diverse artificial neural network (ANN) models to assess and predict system efficiency based on key operational variables in membrane reactor systems. The multilayered perceptron (MLP) and radial basis function (RBF) methodologies are employed, with the MLP models optimized across twelve training algorithms and eight activation functions, exploring up to three hidden layers with variable neuron counts. The MLP model, utilizing the Levenberg-Marquard training algorithm and Tangent-Sigmoid activation function, achieved a high correlation coefficient (R2) of 0.9975 for training and 0.9962 for testing, and a mean squared error (MSE) of 0.00425 for training and 0.23951 for testing, indicating precise and accurate efficiency predictions. The Log-Sigmoid activation function also performed well, with R² values of 0.9971 (training) and 0.9961 (testing), and MSE values of 0.004086 (training) and 0.17694 (testing). Optimization of the RBF network identified the best performance with a spread parameter of 1 and 35 neurons, although the MLP model demonstrated superior accuracy and reduced computational time. Statistical analysis, encompassing correlation coefficient, mean squared error, Root Mean Squared error, absolute average deviation, absolute average relative deviation, and runtime, confirms the network’s consistent and accurate estimation of system efficiency across various input variables. The study highlights that applying tansig and logsig activation functions, configured with neuron counts of 20, 17, 6 and 23, 20, 2 at the first, second and third hidden layers, respectively, offers enhanced accuracy and reliability. The MLP model’s high performance underscores its potential to identify optimal conditions for H2 generation based on system efficiency, thereby advancing membrane reactor technology for hydrogen production.

Modeling and Optimization of Maize Yield and Water Use Efficiency under Biochar, Inorganic Fertilizer and Irrigation Using Principal Component Analysis

This study was conducted to predict the grain yield of a maize crop from easy-to-measure growth parameters and select the best treatment combinations of biochar, inorganic fertilizer, and irrigation for the maize grain yield and water use efficiency (WUE) using the Principal Component Analysis (PCA) technique. Two rates of biochar (0 and 20 t ha−1) and fertilizer (0 and 300 kg ha−1) were applied to the soil, with maize crop planted, and subjected to deficit irrigation at 60, 80, and 100% of full irrigation amounts (FIA). Maize growth parameters (number of leaves—NL, leaf area—LA, leaf area index—LAI, and plant height—PH) were measured weekly. The results showed that the developed principal component regression (PCR) from the easy-to-measure growth parameters were strong and moderate in predicting the maize yield and WUE, with coefficient of determination; r2 values of 0.92 and 0.56, respectively. Using the PCA technique, the integration of irrigation with the least amount of water (60% FAI) with biochar (20 t ha−1) and fertilizer (300 kg ha−1) produced the highest ranking on grain yield and water use efficiency. This optimization technique showed that with the adoption of the integrative approach, 40% of irrigation water could be saved for other agricultural purposes

Advancing IoMT security: A two-factor authentication model employing PUF and Fuzzy logic techniques

The rapid integration of Internet of Things technologies in healthcare has catalyzed the development of the Internet of Medical Things, markedly enhanced patient care while posing significant security risks. This paper introduces a comprehensive computational framework to safeguard Internet of Medical Things devices and healthcare providers through a sophisticated registration and authentication process. Our model incorporates cryptographic technologies such as Physical Unclonable Functions, fuzzy extractors, and hash functions to bolster the security during the registration and authentication processes for Internet of Medical Things devices and healthcare providers. The Physical Unclonable Function module enhances device security by producing unique, non-replicable responses for device authentication, significantly reinforcing the system's defense against physical and cloning attacks. Furthermore, the model leverages fuzzy logic for the real-time classification of patient health states, enhancing the decision-making accuracy. A comparative analysis confirms that our model exceeds existing models in communication cost, computational efficiency and security. The proposed scheme has been rigorously tested against various attacks using the Scyther tool. By employing a unique identifier generation method through Physical Unclonable Function and utilizing fuzzy logic for secure data transmission and patient classification, our framework addresses vulnerabilities such as man-in-the-middle, denial of service, impersonation, identity guessing, password guessing and replay attacks, which are prevalent in current Internet of Medical Things frameworks.

Efficient Optimized YOLOv8 Model with Extended Vision.

In the field of object detection, enhancing algorithm performance in complex scenarios represents a fundamental technological challenge. To address this issue, this paper presents an efficient optimized YOLOv8 model with extended vision (YOLO-EV), which optimizes the performance of the YOLOv8 model through a series of innovative improvement measures and strategies. First, we propose a multi-branch group-enhanced fusion attention (MGEFA) module and integrate it into YOLO-EV, which significantly boosts the model's feature extraction capabilities. Second, we enhance the existing spatial pyramid pooling fast (SPPF) layer by integrating large scale kernel attention (LSKA), improving the model's efficiency in processing spatial information. Additionally, we replace the traditional IOU loss function with the Wise-IOU loss function, thereby enhancing localization accuracy across various target sizes. We also introduce a P6 layer to augment the model's detection capabilities for multi-scale targets. Through network structure optimization, we achieve higher computational efficiency, ensuring that YOLO-EV consumes fewer computational resources than YOLOv8s. In the validation section, preliminary tests on the VOC12 dataset demonstrate YOLO-EV's effectiveness in standard object detection tasks. Moreover, YOLO-EV has been applied to the CottonWeedDet12 and CropWeed datasets, which are characterized by complex scenes, diverse weed morphologies, significant occlusions, and numerous small targets. Experimental results indicate that YOLO-EV exhibits superior detection accuracy in these complex agricultural environments compared to the original YOLOv8s and other state-of-the-art models, effectively identifying and locating various types of weeds, thus demonstrating its significant practical application potential.

Integrated energy management for enhanced grid flexibility: Optimizing renewable resources and energy storage systems across transmission and distribution networks

This study explores the enhancement of electric grid flexibility and the realization of smart grid objectives through the integration of renewable energy (RE) resources and energy storage systems (ESS). While prior research has mainly concentrated on optimizing ESS operations either within the transmission network as a versatile power source or within the distribution network using small-scale batteries, our approach offers a more holistic perspective. We introduce a bi-level stochastic model for integrated energy management that encompasses renewable energy, demand side management (DSM), transmission, and distribution networks as interconnected entities. To tackle the binary variables inherent in the model, we employ a precise method based on reformulation and decomposition techniques to ensure globally optimal solutions. We evaluate the efficacy of our proposed model and the influence of ESS on the networks using various integrated transmission and distribution network systems. Our findings demonstrate the model's efficiency and underscore the cost-saving benefits of integrating energy storage systems. Specifically, incorporating ESS into the distribution grid results in a 13 % reduction in distribution network costs, while deploying large batteries in the transmission grid leads to an impressive 83 % cost reduction.

Predictive residual stress analysis in gear units: an artificial intelligence approach based on finite element data

Determining residual stresses is essential for ensuring the safety and longevity of gear components. Traditional methods, such as analytical, numerical, and experimental approaches, are often costly, time-consuming, and sometimes destructive. This study proposes a new method to predict residual stresses in gear units using artificial intelligence based on data from finite element analysis. To achieve this, a commercial finite element tool models the heat treatment and carburization processes, grounded in thermomechanical metallurgical physics. The residual stresses obtained from finite element analysis are then compared to the ones from experiment using deep hole drilling approach. This is followed by a sensitivity analysis. Next, a specific class of gearbox component geometries and the computed stress fields are utilized to train artificial neural networks. The performance of the artificial neural networks approach is compared to that of two different machine learning regression methods. It has been concluded that this technique successfully anticipates residual stresses in components with scaled geometries and similar heat treatment steps, outperforming the two regression models in accuracy and efficiency. Importantly, it achieves this with incomparable, significantly less computation time compared to standard finite element analysis.

Large Eddy Simulation of an impinging Jet for Cooling of Blades

Abstract The impinging jet shows extremely high performance in local blade cooling and has been widely applied. In this study, high precision Large Eddy Simulation (LES) was employed to simulate compressible fully turbulent impinging jet under specific conditions, using three subgrid-scale (SGS) models - the Smagorinsky model (SM), the Coherent-structure Smagorinsky Model (CSM), and the Selective Mixed Scale Model (SMSM). The predictive accuracy and computational efficiency of the three SGS models were evaluated. Compared with existing DNS and experimental results, the results indicate that there is no significant difference among the three models in computational efficiency. However, subgrid models have a crucial impact on heat transfer. The SMSM model, which utilizes small-scale turbulence information, can better simulate the heat transfer in the impingement region compared with the SM model and the CSM model.

Comparative Study of LSTM and ANN Models for Power Consumption Prediction of Variable Refrigerant Flow (VRF) Systems in Buildings

The optimized control of variable refrigerant flow (VRF) system requires an accurate time series forecast model for power consumption. Currently, physics-based and black-box models widely used for forecasting power consumption may not be able to capture the dynamic and non-linear behavior of such complex systems. This study presents a long-short-term memory (LSTM), deep learning-based model to accurately predict the power consumption of a VRF system with heat recovery units. The model training used one year of VRF system field test data. The feature selection through the Pearson correlation coefficient was implemented to improve the model's accuracy and computational efficiency. The sensitivity analysis of feature selection was performed by preparing three feature sets, including different levels of relationship with the predicted target. Additionally, the hyperparameters of the models were optimized by Bayesian optimization with the Tree-structured Parzen Estimator algorithm. The deep learning model, LSTM model, was compared to the baseline machine learning model, Artificial Neural Network (ANN) and decision tree. The results show that LSTM-30feat with input time step 4 has the best testing performance of Coefficient of the Variation of the Root Mean Square Error (CvRMSE) 23.3%. The best ANN model is ANN-10feat with input time step 8, which has a CvRMSE of 27.8% in testing and 13,569 trainable parameters. However, LSTM-10feat with input time step 4 has the CvRMSE of 24.8% in testing, and the trainable parameters are 1,809. A higher number of trainable parameters in models might result in increased memory usage on the computer and be computationally expensive.

Solar energy prediction with synergistic adversarial energy forecasting system (Solar-SAFS): Harnessing advanced hybrid techniques

An effective solar forecasting is an important part of managing and enhancing the effectiveness of solar power systems. The purpose of this paper is to develop a distinct and smart energy prediction system for solar PV systems. An efficient Solar Synergistic Adversarial Energy Forecasting System (Solar-SAFS) has been developed, which greatly enhances the precision and reliability of solar power predictions. It combines various deep learning methods including Graph Convolutional Networks (GCNs), Variational Autoencoders (VAEs), Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to effectively understand spatial and time-related patterns in solar data. The strength and accuracy of this model are increased by using adversarial training, which makes sure it gives reliable predictions in different situations. For making the hyper parameters better and to increase model's effectiveness and efficiency, we are using Grey-Oystercatcher Hybrid Optimization (GOHO) technique. The Solar-SAFS is validated and tested with different data sets including Fingrid and DSK solar. It shows that this method gives better results compared to current deep learning models. Our findings suggest that Solar-SAFS improves the accuracy and stability of solar energy prediction by a large margin, offering an advanced solution for managing renewable energy resources. Moreover, it significantly helps to obtain better predictions for solar energy production, which is a crucial need in this sector. It also encourages more efficient and dependable forecasting methods for renewable energy sources like sunlight. Compared to existing deep learning models, the Solar-SAFS demonstrates significantly lower error rates across multiple performance metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Bias Error (MBE).

A Hybrid Approach for Efficient DDoS Detection in Network Traffic Using CBLOF-Based Feature Engineering and XGBoost

This is one of the greatest challenges in computer network security and cannot be dealt with without a set of most recent detection techniques. This paper lays down a new hybrid technique that combines Clustering-Based Local Outlier Factor (CBLOF) and Extreme Gradient Boosting (XGBoost) to enhance accuracy while detecting Distributed Denial of Service (DDoS) from network traffic. The proposed hybrid model utilizes a CBLOF for outlier detection as feature engineering. Over the detected anomalies, classification is to be done using XGBoost classification to attain the objective. The proposed hybrid model was tested extensively on CICIDS 2017 and CICIDS 2018 datasets Compared with traditional ones, the proposed model outperformed the traditional ones with an accuracy rate of 99.99%, precision of 100%, and F1 score reflecting perfection. These results confirm this model's efficiency in terms of known and novel attack patterns and introduce a further reliable framework for the timely detection of DDoS attacks. Even if it is computation-heavy, optimization could be made towards real-time large-scale data.

Predictive Artificial Intelligence Models for Energy Efficiency in Hybrid and Electric Vehicles: Analysis for Enna, Sicily

Developments in artificial intelligence techniques allow for an improvement in sustainable mobility strategies with particular reference to energy consumption estimates of electric vehicles (EVs). This research proposes a vehicle energy model developed on the basis of deep neural network (DNN) technology. This study also explores the potential application of the model developed for the movement data of new vehicles in the province of Enna, Sicily, Italy, which are characterized by numerous attractors and the increasing number of hybrid and electric cars circulating. The energy model for electric vehicles shows high accuracy and versatility, requiring vehicle velocity and acceleration as input data to predict energy consumption. This research article also provides recommendations for the energy modeling of electric vehicles and outlines additional steps for model development. The implemented methodological approach and its results can be used by transport decision-makers to plan new transport policies in Italian cities aimed at optimizing vehicle charging infrastructure. They can also help vehicle users accurately estimate energy consumption, generate maps, and identify locations with the highest energy consumption.

End-to-end deep learning on classification of neutral-to-ground voltage in secondary distribution systems

Neutral-to-ground voltage (NTGV) abnormalities in secondary distribution systems (SDS) pose significant power quality (PQ) challenges, including safety hazards, power losses, and equipment damage. Despite their importance, these abnormalities remain relatively understudied. Accurate classification of NTGV events is crucial for effective mitigation strategies. Existing research primarily relies on machine learning (ML) models trained on manually extracted features from simulated or real-world signals. This paper introduces a novel end-to-end deep learning approach that leverages Gate Recurrent Units (GRU) to bypass manual feature extraction, directly utilizing real-world signals from three NTGV event categories: ground fault, lightning strike, and normal conditions. This is first time that GRU has been used for NTGV classification using raw data. The model's generalizability is assessed through 5-fold cross-validation. A comparative analysis with baseline models and traditional ML techniques demonstrates the proposed model's superior performance and computational efficiency due to its ability to directly process raw data.

Detecting emerald ash borer boring vibrations using an encoder-decoder and improved DenseNet model.

Forest ecosystems are under constant threat from wood-boring pests such as the Emerald ash borer (EAB), which remain elusive owing to their hidden life cycles within tree trunks. Early detection is vital to mitigate economic and ecological damage. The main current monitoring method is manual detection which is ineffective at early stages of infestation. This study introduces VibroEABNet, a deep learning-based joint recognition network designed to enhance the detection of EAB boring vibration signals, with a novel approach integrating denoising and recognition modules. The proposed VibroEABNet model demonstrated exceptional performance, achieving an average accuracy of 98.98% across multiple signal-to-noise ratios (SNRs) in test datasets and a remarkable 97.5% accuracy in real forest datasets, surpassing traditional models and other deep learning networks evaluated in this study. These findings were supported by rigorous noise resistance analysis and real dataset evaluation, indicating the model's robustness and reliability in practical applications. Furthermore, the model's efficiency was highlighted by its inference time of 26 ms and a compact model size of 8.43 MB, underscoring its suitability for deployment in resource-limited environments. The development of VibroEABNet marks a significant advancement in pest detection methodologies, offering a scalable, accurate and efficient solution for early monitoring of wood-boring pests. The integration of a denoising module within the network structure addresses the challenge of environmental noise, one of the primary limitations in acoustic monitoring of pests. Currently, this research is limited to a specific pest. Future work will focus on the applicability of this network to other wood-boring pests. © 2024 Society of Chemical Industry.

MRI-Based Brain Tumor Classification using ResNet-50 and Optimized Softmax Regression

Accurate classification of brain tumors is crucial for effective treatment planning and patient management. This study presents a new hybrid deep learning classification method based on transfer learning by feature extraction to automate the categorization of MRI brain image datasets into four classes: meningioma, glioma, pituitary tumor, and no tumor. The proposed method combines a finely-tuned ResNet-50 model, a state-of-the-art convolutional neural network architecture, with optimized Softmax Regression (SR) for classification. The study explores the use of data augmentation techniques and evaluates the model's performance on both augmented and unaugmented images. The results demonstrate that the proposed method achieves an impressive accuracy of 98.4%, outperforming existing methods for automatic brain tumor detection. Furthermore, a detailed comparative analysis is presented to evaluate the proposed model's accuracy and efficiency against previous state-of-the-art hybrid models for brain tumor classification. The study suggests that the proposed methodology could be employed as a diagnostic tool to aid radiologists in identifying questionable brain regions, potentially improving the accuracy and efficiency of brain tumor diagnosis.

A constraint programming approach for resource-constrained flexible assembly flow shop scheduling problem with batch direct delivery

Production and distribution are both crucial components of supply chains. Integrated production and distribution scheduling (IPDS) in the context of flexible assembly flow shop scheduling and batch delivery problems is often overlooked. A realistic problem inspired by the production and distribution processes of a dishwasher factory can be modeled as a resource-constrained flexible assembly flow shop scheduling problem with batch direct delivery (RCFAFSP-BDD). In this problem, order requirements are decomposed into several production tasks processed at different stages of the workshop, and are then delivered in batches via a third-party logistics provider to regional distributors at various locations. Auxiliary resource restrictions, hierarchical coupling constraints, machine eligibility restrictions and sequence-dependent setup times, are incorporated into the problem as operational constraints. To the best of our knowledge, this is the first attempt to solve this problem. This work formulates a mixed-integer linear programming (MIP) model to minimize total costs, including tardiness, inventory, and delivery costs. Given the problem’s strong NP-hard nature, the focus is on developing an efficient solution approach using constraint programming (CP). A CP model is proposed and enhanced with multiple redundant constraints. To reduce runtime, two branching strategies are designed for the CP model. Numerical experiments with varying instance scales reveal that the proposed CP model outperforms the MIP model in accuracy and efficiency within the given time limit. The redundant constraints and search strategy can reduce CP model runtime by up to 263.83%. Compared to manual scheduling at the studied factory, the CP model can cut costs by up to 26.59% for real data, offering viable alternatives for factory planners.

IGATTLDA: Integrative graph attention and transformer-based model for predicting lncRNA-Disease associations.

Long non-coding RNAs (lncRNAs) have emerged as significant contributors to the regulation of various biological processes, and their dysregulation has been linked to a variety of human disorders. Accurate prediction of potential correlations between lncRNAs and diseases is crucial for advancing disease diagnostics and treatment procedures. The authors introduced a novel computational method, iGATTLDA, for the prediction of lncRNA-disease associations. The model utilised lncRNA and disease similarity matrices, with known associations represented in an adjacency matrix. A heterogeneous network was constructed, dissecting lncRNAs and diseases as nodes and their associations as edges. The Graph Attention Network (GAT) is employed to process initial features and corresponding adjacency information. GAT identified significant neighbouring nodes in the network, capturing intricate relationships between lncRNAs and diseases, and generating new feature representations. Subsequently, the transformer captures global dependencies and interactions across the entire sequence of features produced by the GAT. Consequently, iGATTLDA successfully captures complex relationships and interactions that conventional approaches may overlook. In evaluating iGATTLDA, it attained an area under the receiver operating characteristic (ROC) curve (AUC) of 0.95 and an area under the precision recall curve (AUPRC) of 0.96 with a two-layer multilayer perceptron (MLP) classifier. These results were notably higher compared to the majority of previously proposed models, further substantiating the model's efficiency in predicting potential lncRNA-disease associations by incorporating both local and global interactions. The implementation details can be obtained from https://github.com/momanyibiffon/iGATTLDA.

MH-DCNet: An improved flow field prediction framework coupling neural network with physics solver

With the development of intelligent computing technology, deep learning methods have provided an efficient solution for rapid flow field prediction in computational fluid dynamics (CFD) problems. However, existing methods have limitations in handling interference among physical variables due to different data distributions, leading to a decline in prediction performance. In this paper, we propose MH-DCNet, an improved flow field prediction framework that couples a neural network with a physics solver. Specifically, to address the data distribution problem, we design a multi-head deep convolutional neural network that decouples the prediction of physical variables through multiple encoders and decoders. We also develop a hybrid loss function by introducing the mean structural similarity to better capture the complex spatial structures and distribution features of flow fields. We evaluate MH-DCNet with unseen geometries and various flow conditions. Experimental results show that MH-DCNet outperforms other advanced models in efficiency and generalization capability. It accelerates the prediction process by 2.35 times compared to the CFD method while meeting the convergence constraints.