Complex-valued neural networks(CVNNs)have shown their excellent efficiency compared to their real counterparts in speech enhancement,image and signal processing.Researchers throughout the years have made many efforts ...Complex-valued neural networks(CVNNs)have shown their excellent efficiency compared to their real counterparts in speech enhancement,image and signal processing.Researchers throughout the years have made many efforts to improve the learning algorithms and activation functions of CVNNs.Since CVNNs have proven to have better performance in handling the naturally complex-valued data and signals,this area of study will grow and expect the arrival of some effective improvements in the future.Therefore,there exists an obvious reason to provide a comprehensive survey paper that systematically collects and categorizes the advancement of CVNNs.In this paper,we discuss and summarize the recent advances based on their learning algorithms,activation functions,which is the most challenging part of building a CVNN,and applications.Besides,we outline the structure and applications of complex-valued convolutional,residual and recurrent neural networks.Finally,we also present some challenges and future research directions to facilitate the exploration of the ability of CVNNs.展开更多
In this paper, a novel design procedure is proposed for synthesizing high-capacity auto-associative memories based on complex-valued neural networks with real-imaginary-type activation functions and constant delays. S...In this paper, a novel design procedure is proposed for synthesizing high-capacity auto-associative memories based on complex-valued neural networks with real-imaginary-type activation functions and constant delays. Stability criteria dependent on external inputs of neural networks are derived. The designed networks can retrieve the stored patterns by external inputs rather than initial conditions. The derivation can memorize the desired patterns with lower-dimensional neural networks than real-valued neural networks, and eliminate spurious equilibria of complex-valued neural networks. One numerical example is provided to show the effectiveness and superiority of the presented results.展开更多
In this paper, the singularity and its effect on learning dynamics in the complex-valued neural network are elucidated. It has learned that the linear combination structure in the updating rule of the complex-valued n...In this paper, the singularity and its effect on learning dynamics in the complex-valued neural network are elucidated. It has learned that the linear combination structure in the updating rule of the complex-valued neural network increases the speed of moving away from the singular points, and the complex-valued neural network cannot be easily influenced by the singular points, whereas the learning of the usual real-valued neural network can be attracted in the neighborhood of singular points, which causes a standstill in learning. Simulation results on the learning dynamics of the three-layered real-valued and complex-valued neural networks in the neighborhood of singularities support the analytical results.展开更多
Recently,deep learning has been used to establish the nonlinear and nonintuitive mapping between physical structures and electromagnetic responses of meta-atoms for higher computational efficiency.However,to obtain su...Recently,deep learning has been used to establish the nonlinear and nonintuitive mapping between physical structures and electromagnetic responses of meta-atoms for higher computational efficiency.However,to obtain sufficiently accurate predictions,the conventional deep-learning-based method consumes excessive time to collect the data set,thus hindering its wide application in this interdisciplinary field.We introduce a spectral transfer-learning-based metasurface design method to achieve excellent performance on a small data set with only 1000 samples in the target waveband by utilizing open-source data from another spectral range.We demonstrate three transfer strategies and experimentally quantify their performance,among which the“frozen-none”robustly improves the prediction accuracy by∼26%compared to direct learning.We propose to use a complex-valued deep neural network during the training process to further improve the spectral predicting precision by∼30%compared to its real-valued counterparts.We design several typical teraherz metadevices by employing a hybrid inverse model consolidating this trained target network and a global optimization algorithm.The simulated results successfully validate the capability of our approach.Our work provides a universal methodology for efficient and accurate metasurface design in arbitrary wavebands,which will pave the way toward the automated and mass production of metasurfaces.展开更多
Deep neural networks(DNNs)are effective in solving both forward and inverse problems for nonlinear partial differential equations(PDEs).However,conventional DNNs are not effective in handling problems such as delay di...Deep neural networks(DNNs)are effective in solving both forward and inverse problems for nonlinear partial differential equations(PDEs).However,conventional DNNs are not effective in handling problems such as delay differential equations(DDEs)and delay integrodifferential equations(DIDEs)with constant delays,primarily due to their low regularity at delayinduced breaking points.In this paper,a DNN method that combines multi-task learning(MTL)which is proposed to solve both the forward and inverse problems of DIDEs.The core idea of this approach is to divide the original equation into multiple tasks based on the delay,using auxiliary outputs to represent the integral terms,followed by the use of MTL to seamlessly incorporate the properties at the breaking points into the loss function.Furthermore,given the increased training dificulty associated with multiple tasks and outputs,we employ a sequential training scheme to reduce training complexity and provide reference solutions for subsequent tasks.This approach significantly enhances the approximation accuracy of solving DIDEs with DNNs,as demonstrated by comparisons with traditional DNN methods.We validate the effectiveness of this method through several numerical experiments,test various parameter sharing structures in MTL and compare the testing results of these structures.Finally,this method is implemented to solve the inverse problem of nonlinear DIDE and the results show that the unknown parameters of DIDE can be discovered with sparse or noisy data.展开更多
For the diagnostics and health management of lithium-ion batteries,numerous models have been developed to understand their degradation characteristics.These models typically fall into two categories:data-driven models...For the diagnostics and health management of lithium-ion batteries,numerous models have been developed to understand their degradation characteristics.These models typically fall into two categories:data-driven models and physical models,each offering unique advantages but also facing limitations.Physics-informed neural networks(PINNs)provide a robust framework to integrate data-driven models with physical principles,ensuring consistency with underlying physics while enabling generalization across diverse operational conditions.This study introduces a PINN-based approach to reconstruct open circuit voltage(OCV)curves and estimate key ageing parameters at both the cell and electrode levels.These parameters include available capacity,electrode capacities,and lithium inventory capacity.The proposed method integrates OCV reconstruction models as functional components into convolutional neural networks(CNNs)and is validated using a public dataset.The results reveal that the estimated ageing parameters closely align with those obtained through offline OCV tests,with errors in reconstructed OCV curves remaining within 15 mV.This demonstrates the ability of the method to deliver fast and accurate degradation diagnostics at the electrode level,advancing the potential for precise and efficient battery health management.展开更多
The ability to accurately predict urban traffic flows is crucial for optimising city operations.Consequently,various methods for forecasting urban traffic have been developed,focusing on analysing historical data to u...The ability to accurately predict urban traffic flows is crucial for optimising city operations.Consequently,various methods for forecasting urban traffic have been developed,focusing on analysing historical data to understand complex mobility patterns.Deep learning techniques,such as graph neural networks(GNNs),are popular for their ability to capture spatio-temporal dependencies.However,these models often become overly complex due to the large number of hyper-parameters involved.In this study,we introduce Dynamic Multi-Graph Spatial-Temporal Graph Neural Ordinary Differential Equation Networks(DMST-GNODE),a framework based on ordinary differential equations(ODEs)that autonomously discovers effective spatial-temporal graph neural network(STGNN)architectures for traffic prediction tasks.The comparative analysis of DMST-GNODE and baseline models indicates that DMST-GNODE model demonstrates superior performance across multiple datasets,consistently achieving the lowest Root Mean Square Error(RMSE)and Mean Absolute Error(MAE)values,alongside the highest accuracy.On the BKK(Bangkok)dataset,it outperformed other models with an RMSE of 3.3165 and an accuracy of 0.9367 for a 20-min interval,maintaining this trend across 40 and 60 min.Similarly,on the PeMS08 dataset,DMST-GNODE achieved the best performance with an RMSE of 19.4863 and an accuracy of 0.9377 at 20 min,demonstrating its effectiveness over longer periods.The Los_Loop dataset results further emphasise this model’s advantage,with an RMSE of 3.3422 and an accuracy of 0.7643 at 20 min,consistently maintaining superiority across all time intervals.These numerical highlights indicate that DMST-GNODE not only outperforms baseline models but also achieves higher accuracy and lower errors across different time intervals and datasets.展开更多
The increasing popularity of the Internet and the widespread use of information technology have led to a rise in the number and sophistication of network attacks and security threats.Intrusion detection systems are cr...The increasing popularity of the Internet and the widespread use of information technology have led to a rise in the number and sophistication of network attacks and security threats.Intrusion detection systems are crucial to network security,playing a pivotal role in safeguarding networks from potential threats.However,in the context of an evolving landscape of sophisticated and elusive attacks,existing intrusion detection methodologies often overlook critical aspects such as changes in network topology over time and interactions between hosts.To address these issues,this paper proposes a real-time network intrusion detection method based on graph neural networks.The proposedmethod leverages the advantages of graph neural networks and employs a straightforward graph construction method to represent network traffic as dynamic graph-structured data.Additionally,a graph convolution operation with a multi-head attention mechanism is utilized to enhance the model’s ability to capture the intricate relationships within the graph structure comprehensively.Furthermore,it uses an integrated graph neural network to address dynamic graphs’structural and topological changes at different time points and the challenges of edge embedding in intrusion detection data.The edge classification problem is effectively transformed into node classification by employing a line graph data representation,which facilitates fine-grained intrusion detection tasks on dynamic graph node feature representations.The efficacy of the proposed method is evaluated using two commonly used intrusion detection datasets,UNSW-NB15 and NF-ToN-IoT-v2,and results are compared with previous studies in this field.The experimental results demonstrate that our proposed method achieves 99.3%and 99.96%accuracy on the two datasets,respectively,and outperforms the benchmark model in several evaluation metrics.展开更多
This study investigates photonuclear reaction(γ,n)cross-sections using Bayesian neural network(BNN)analysis.After determining the optimal network architecture,which features two hidden layers,each with 50 hidden node...This study investigates photonuclear reaction(γ,n)cross-sections using Bayesian neural network(BNN)analysis.After determining the optimal network architecture,which features two hidden layers,each with 50 hidden nodes,training was conducted for 30,000 iterations to ensure comprehensive data capture.By analyzing the distribution of absolute errors positively correlated with the cross-section for the isotope 159Tb,as well as the relative errors unrelated to the cross-section,we confirmed that the network effectively captured the data features without overfitting.Comparison with the TENDL-2021 Database demonstrated the BNN's reliability in fitting photonuclear cross-sections with lower average errors.The predictions for nuclei with single and double giant dipole resonance peak cross-sections,the accurate determination of the photoneutron reaction threshold in the low-energy region,and the precise description of trends in the high-energy cross-sections further demonstrate the network's generalization ability on the validation set.This can be attributed to the consistency of the training data.By using consistent training sets from different laboratories,Bayesian neural networks can predict nearby unknown cross-sections based on existing laboratory data,thereby estimating the potential differences between other laboratories'existing data and their own measurement results.Experimental measurements of photonuclear reactions on the newly constructed SLEGS beamline will contribute to clarifying the differences in cross-sections within the existing data.展开更多
Spectrum-based fault localization (SBFL) generates a ranked list of suspicious elements by using the program execution spectrum, but the excessive number of elements ranked in parallel results in low localization accu...Spectrum-based fault localization (SBFL) generates a ranked list of suspicious elements by using the program execution spectrum, but the excessive number of elements ranked in parallel results in low localization accuracy. Most researchers consider intra-class dependencies to improve localization accuracy. However, some studies show that inter-class method call type faults account for more than 20%, which means such methods still have certain limitations. To solve the above problems, this paper proposes a two-phase software fault localization based on relational graph convolutional neural networks (Two-RGCNFL). Firstly, in Phase 1, the method call dependence graph (MCDG) of the program is constructed, the intra-class and inter-class dependencies in MCDG are extracted by using the relational graph convolutional neural network, and the classifier is used to identify the faulty methods. Then, the GraphSMOTE algorithm is improved to alleviate the impact of class imbalance on classification accuracy. Aiming at the problem of parallel ranking of element suspicious values in traditional SBFL technology, in Phase 2, Doc2Vec is used to learn static features, while spectrum information serves as dynamic features. A RankNet model based on siamese multi-layer perceptron is constructed to score and rank statements in the faulty method. This work conducts experiments on 5 real projects of Defects4J benchmark. Experimental results show that, compared with the traditional SBFL technique and two baseline methods, our approach improves the Top-1 accuracy by 262.86%, 29.59% and 53.01%, respectively, which verifies the effectiveness of Two-RGCNFL. Furthermore, this work verifies the importance of inter-class dependencies through ablation experiments.展开更多
Identifying cyberattacks that attempt to compromise digital systems is a critical function of intrusion detection systems(IDS).Data labeling difficulties,incorrect conclusions,and vulnerability to malicious data injec...Identifying cyberattacks that attempt to compromise digital systems is a critical function of intrusion detection systems(IDS).Data labeling difficulties,incorrect conclusions,and vulnerability to malicious data injections are only a few drawbacks of using machine learning algorithms for cybersecurity.To overcome these obstacles,researchers have created several network IDS models,such as the Hidden Naive Bayes Multiclass Classifier and supervised/unsupervised machine learning techniques.This study provides an updated learning strategy for artificial neural network(ANN)to address data categorization problems caused by unbalanced data.Compared to traditional approaches,the augmented ANN’s 92%accuracy is a significant improvement owing to the network’s increased resilience to disturbances and computational complexity,brought about by the addition of a random weight and standard scaler.Considering the ever-evolving nature of cybersecurity threats,this study introduces a revolutionary intrusion detection method.展开更多
Metabolomics covers a wide range of applications in life sciences,biomedicine,and phytology.Data acquisition(to achieve high coverage and efficiency)and analysis(to pursue good classification)are two key segments invo...Metabolomics covers a wide range of applications in life sciences,biomedicine,and phytology.Data acquisition(to achieve high coverage and efficiency)and analysis(to pursue good classification)are two key segments involved in metabolomics workflows.Various chemometric approaches utilizing either pattern recognition or machine learning have been employed to separate different groups.However,insufficient feature extraction,inappropriate feature selection,overfitting,or underfitting lead to an insufficient capacity to discriminate plants that are often easily confused.Using two ginseng varieties,namely Panax japonicus(PJ)and Panax japonicus var.major(PJvm),containing the similar ginsenosides,we integrated pseudo-targeted metabolomics and deep neural network(DNN)modeling to achieve accurate species differentiation.A pseudo-targeted metabolomics approach was optimized through data acquisition mode,ion pairs generation,comparison between multiple reaction monitoring(MRM)and scheduled MRM(sMRM),and chromatographic elution gradient.In total,1980 ion pairs were monitored within 23 min,allowing for the most comprehensive ginseng metabolome analysis.The established DNN model demonstrated excellent classification performance(in terms of accuracy,precision,recall,F1 score,area under the curve,and receiver operating characteristic(ROC))using the entire metabolome data and feature-selection dataset,exhibiting superior advantages over random forest(RF),support vector machine(SVM),extreme gradient boosting(XGBoost),and multilayer perceptron(MLP).Moreover,DNNs were advantageous for automated feature learning,nonlinear modeling,adaptability,and generalization.This study confirmed practicality of the established strategy for efficient metabolomics data analysis and reliable classification performance even when using small-volume samples.This established approach holds promise for plant metabolomics and is not limited to ginseng.展开更多
Severe ground-level ozone(O_(3))pollution over major Chinese cities has become one of the most challenging problems,which have deleterious effects on human health and the sustainability of society.This study explored ...Severe ground-level ozone(O_(3))pollution over major Chinese cities has become one of the most challenging problems,which have deleterious effects on human health and the sustainability of society.This study explored the spatiotemporal distribution characteristics of ground-level O_(3) and its precursors based on conventional pollutant and meteorological monitoring data in Zhejiang Province from 2016 to 2021.Then,a high-performance convolutional neural network(CNN)model was established by expanding the moment and the concentration variations to general factors.Finally,the response mechanism of O_(3) to the variation with crucial influencing factors is explored by controlling variables and interpolating target variables.The results indicated that the annual average MDA8-90th concentrations in Zhejiang Province are higher in the northern and lower in the southern.When the wind direction(WD)ranges from east to southwest and the wind speed(WS)ranges between 2 and 3 m/sec,higher O_(3) concentration prone to occur.At different temperatures(T),the O_(3) concentration showed a trend of first increasing and subsequently decreasing with increasing NO_(2) concentration,peaks at the NO_(2) concentration around 0.02mg/m^(3).The sensitivity of NO_(2) to O_(3) formation is not easily affected by temperature,barometric pressure and dew point temperature.Additionally,there is a minimum IRNO_(2) at each temperature when the NO_(2) concentration is 0.03 mg/m^(3),and this minimum IRNO_(2) decreases with increasing temperature.The study explores the response mechanism of O_(3) with the change of driving variables,which can provide a scientific foundation and methodological support for the targeted management of O_(3) pollution.展开更多
This paper addresses the consensus problem of nonlinear multi-agent systems subject to external disturbances and uncertainties under denial-ofservice(DoS)attacks.Firstly,an observer-based state feedback control method...This paper addresses the consensus problem of nonlinear multi-agent systems subject to external disturbances and uncertainties under denial-ofservice(DoS)attacks.Firstly,an observer-based state feedback control method is employed to achieve secure control by estimating the system's state in real time.Secondly,by combining a memory-based adaptive eventtriggered mechanism with neural networks,the paper aims to approximate the nonlinear terms in the networked system and efficiently conserve system resources.Finally,based on a two-degree-of-freedom model of a vehicle affected by crosswinds,this paper constructs a multi-unmanned ground vehicle(Multi-UGV)system to validate the effectiveness of the proposed method.Simulation results show that the proposed control strategy can effectively handle external disturbances such as crosswinds in practical applications,ensuring the stability and reliable operation of the Multi-UGV system.展开更多
Real-time prediction and precise control of sinter quality are pivotal for energy saving,cost reduction,quality improvement and efficiency enhancement in the ironmaking process.To advance,the accuracy and comprehensiv...Real-time prediction and precise control of sinter quality are pivotal for energy saving,cost reduction,quality improvement and efficiency enhancement in the ironmaking process.To advance,the accuracy and comprehensiveness of sinter quality prediction,an intelligent flare monitoring system for sintering machine tails that combines hybrid neural networks integrating convolutional neural network with long short-term memory(CNN-LSTM)networks was proposed.The system utilized a high-temperature thermal imager for image acquisition at the sintering machine tail and employed a zone-triggered method to accurately capture dynamic feature images under challenging conditions of high-temperature,high dust,and occlusion.The feature images were then segmented through a triple-iteration multi-thresholding approach based on the maximum between-class variance method to minimize detail loss during the segmentation process.Leveraging the advantages of CNN and LSTM networks in capturing temporal and spatial information,a comprehensive model for sinter quality prediction was constructed,with inputs including the proportion of combustion layer,porosity rate,temperature distribution,and image features obtained from the convolutional neural network,and outputs comprising quality indicators such as underburning index,uniformity index,and FeO content of the sinter.The accuracy is notably increased,achieving a 95.8%hit rate within an error margin of±1.0.After the system is applied,the average qualified rate of FeO content increases from 87.24%to 89.99%,representing an improvement of 2.75%.The average monthly solid fuel consumption is reduced from 49.75 to 46.44 kg/t,leading to a 6.65%reduction and underscoring significant energy saving and cost reduction effects.展开更多
The isolated fracture-vug systems controlled by small-scale strike-slip faults within ultra-deep carbonate rocks of the Tarim Basin exhibit significant exploration potential.The study employs a novel training set inco...The isolated fracture-vug systems controlled by small-scale strike-slip faults within ultra-deep carbonate rocks of the Tarim Basin exhibit significant exploration potential.The study employs a novel training set incorporating innovative fault labels to train a U-Net-structured CNN model,enabling effective identification of small-scale strike-slip faults through seismic data interpretation.Based on the CNN faults,we analyze the distribution patterns of small-scale strike-slip faults.The small-scale strike-slip faults can be categorized into NNW-trending and NE-trending groups with strike lengths ranging 200–5000 m.The development intensity of small-scale strike-slip faults in the Lower Yingshan Member notably exceeds that in the Upper Member.The Lower and Upper Yingshan members are two distinct mechanical layers with contrasting brittleness characteristics,separated by a low-brittleness layer.The superior brittleness of the Lower Yingshan Member enhances the development intensity of small-scale strike-slip faults compared to the upper member,while the low-brittleness layer exerts restrictive effects on vertical fault propagation.Fracture-vug systems formed by interactions of two or more small-scale strike-slip faults demonstrate larger sizes than those controlled by individual faults.All fracture-vug system sizes show positive correlations with the vertical extents of associated small-scale strike-slip faults,particularly intersection and approaching fracture-vug systems exhibit accelerated size increases proportional to the vertical extents.展开更多
Severe acute respiratory syndrome coronavirus 2(SARS-CoV-2)mutations are influenced by random and uncontrollable factors,and the risk of the next widespread epidemic remains.Dual-target drugs that synergistically act ...Severe acute respiratory syndrome coronavirus 2(SARS-CoV-2)mutations are influenced by random and uncontrollable factors,and the risk of the next widespread epidemic remains.Dual-target drugs that synergistically act on two targets exhibit strong therapeutic effects and advantages against mutations.In this study,a novel computational workflow was developed to design dual-target SARS-CoV-2 candidate inhibitors with the Envelope protein and Main protease selected as the two target proteins.The drug-like molecules of our self-constructed 3D scaffold database were used as high-throughput molecular docking probes for feature extraction of two target protein pockets.A multi-layer perceptron(MLP)was employed to embed the binding affinities into a latent space as conditional vectors to control conditional distribution.Utilizing a conditional generative neural network,cG-SchNet,with 3D Euclidean group(E3)symmetries,the conditional probability distributions of molecular 3D structures were acquired and a set of novel SARS-CoV-2 dual-target candidate inhibitors were generated.The 1D probability,2D joint probability,and 2D cumulative probability distribution results indicate that the generated sets are significantly enhanced compared to the training set in the high binding affinity area.Among the 201 generated molecules,42 molecules exhibited a sum binding affinity exceeding 17.0 kcal/mol while 9 of them having a sum binding affinity exceeding 19.0 kcal/mol,demonstrating structure diversity along with strong dual-target affinities,good absorption,distribution,metabolism,excretion,and toxicity(ADMET)properties,and ease of synthesis.Dual-target drugs are rare and difficult to find,and our“high-throughput docking-multi-conditional generation”workflow offers a wide range of options for designing or optimizing potent dual-target SARS-CoV-2 inhibitors.展开更多
Timely and accurate forecasting of storm surges can effectively prevent typhoon storm surges from causing large economic losses and casualties in coastal areas.At present,numerical model forecasting consumes too many ...Timely and accurate forecasting of storm surges can effectively prevent typhoon storm surges from causing large economic losses and casualties in coastal areas.At present,numerical model forecasting consumes too many resources and takes too long to compute,while neural network forecasting lacks regional data to train regional forecasting models.In this study,we used the DUAL wind model to build typhoon wind fields,and constructed a typhoon database of 75 processes in the northern South China Sea using the coupled Advanced Circulation-Simulating Waves Nearshore(ADCIRC-SWAN)model.Then,a neural network with a Res-U-Net structure was trained using the typhoon database to forecast the typhoon processes in the validation dataset,and an excellent storm surge forecasting effect was achieved in the Pearl River Estuary region.The storm surge forecasting effect of stronger typhoons was improved by adding a branch structure and transfer learning.展开更多
Accurate estimation of mineralogy from geophysical well logs is crucial for characterizing geological formations,particularly in hydrocarbon exploration,CO_(2) sequestration,and geothermal energy development.Current t...Accurate estimation of mineralogy from geophysical well logs is crucial for characterizing geological formations,particularly in hydrocarbon exploration,CO_(2) sequestration,and geothermal energy development.Current techniques,such as multimineral petrophysical analysis,offer details into mineralogical distribution.However,it is inherently time-intensive and demands substantial geological expertise for accurate model evaluation.Furthermore,traditional machine learning techniques often struggle to predict mineralogy accurately and sometimes produce estimations that violate fundamental physical principles.To address this,we present a new approach using Physics-Integrated Neural Networks(PINNs),that combines data-driven learning with domain-specific physical constraints,embedding petrophysical relationships directly into the neural network architecture.This approach enforces that predictions adhere to physical laws.The methodology is applied to the Broom Creek Deep Saline aquifer,a CO_(2) sequestration site in the Williston Basin,to predict the volumes of key mineral constituents—quartz,dolomite,feldspar,anhydrite,illite—along with porosity.Compared to traditional artificial neural networks (ANN),the PINN approach demonstrates higher accuracy and better generalizability,significantly enhancing predictive performance on unseen well datasets.The average mean error across the three blind wells is 0.123 for ANN and 0.042 for PINN,highlighting the superior accuracy of the PINN approach.This method reduces uncertainties in reservoir characterization by improving the reliability of mineralogy and porosity predictions,providing a more robust tool for decision-making in various subsurface geoscience applications.展开更多
Accurate prediction of molecular properties is crucial for selecting compounds with ideal properties and reducing the costs and risks of trials.Traditional methods based on manually crafted features and graph-based me...Accurate prediction of molecular properties is crucial for selecting compounds with ideal properties and reducing the costs and risks of trials.Traditional methods based on manually crafted features and graph-based methods have shown promising results in molecular property prediction.However,traditional methods rely on expert knowledge and often fail to capture the complex structures and interactions within molecules.Similarly,graph-based methods typically overlook the chemical structure and function hidden in molecular motifs and struggle to effectively integrate global and local molecular information.To address these limitations,we propose a novel fingerprint-enhanced hierarchical graph neural network(FH-GNN)for molecular property prediction that simultaneously learns information from hierarchical molecular graphs and fingerprints.The FH-GNN captures diverse hierarchical chemical information by applying directed message-passing neural networks(D-MPNN)on a hierarchical molecular graph that integrates atomic-level,motif-level,and graph-level information along with their relationships.Addi-tionally,we used an adaptive attention mechanism to balance the importance of hierarchical graphs and fingerprint features,creating a comprehensive molecular embedding that integrated hierarchical mo-lecular structures with domain knowledge.Experiments on eight benchmark datasets from MoleculeNet showed that FH-GNN outperformed the baseline models in both classification and regression tasks for molecular property prediction,validating its capability to comprehensively capture molecular informa-tion.By integrating molecular structure and chemical knowledge,FH-GNN provides a powerful tool for the accurate prediction of molecular properties and aids in the discovery of potential drug candidates.展开更多
基金partially supported by the JSPS KAKENHI(JP22H03643,JP19K22891)。
文摘Complex-valued neural networks(CVNNs)have shown their excellent efficiency compared to their real counterparts in speech enhancement,image and signal processing.Researchers throughout the years have made many efforts to improve the learning algorithms and activation functions of CVNNs.Since CVNNs have proven to have better performance in handling the naturally complex-valued data and signals,this area of study will grow and expect the arrival of some effective improvements in the future.Therefore,there exists an obvious reason to provide a comprehensive survey paper that systematically collects and categorizes the advancement of CVNNs.In this paper,we discuss and summarize the recent advances based on their learning algorithms,activation functions,which is the most challenging part of building a CVNN,and applications.Besides,we outline the structure and applications of complex-valued convolutional,residual and recurrent neural networks.Finally,we also present some challenges and future research directions to facilitate the exploration of the ability of CVNNs.
基金Project supported by the National Natural Science Foundation of China(Grant Nos.61503338,61573316,61374152,and 11302195)the Natural Science Foundation of Zhejiang Province,China(Grant No.LQ15F030005)
文摘In this paper, a novel design procedure is proposed for synthesizing high-capacity auto-associative memories based on complex-valued neural networks with real-imaginary-type activation functions and constant delays. Stability criteria dependent on external inputs of neural networks are derived. The designed networks can retrieve the stored patterns by external inputs rather than initial conditions. The derivation can memorize the desired patterns with lower-dimensional neural networks than real-valued neural networks, and eliminate spurious equilibria of complex-valued neural networks. One numerical example is provided to show the effectiveness and superiority of the presented results.
文摘In this paper, the singularity and its effect on learning dynamics in the complex-valued neural network are elucidated. It has learned that the linear combination structure in the updating rule of the complex-valued neural network increases the speed of moving away from the singular points, and the complex-valued neural network cannot be easily influenced by the singular points, whereas the learning of the usual real-valued neural network can be attracted in the neighborhood of singular points, which causes a standstill in learning. Simulation results on the learning dynamics of the three-layered real-valued and complex-valued neural networks in the neighborhood of singularities support the analytical results.
基金support from the National Natural Science Foundation of China (Grant Nos.62027820,61975143,61735012,and 62205380).
文摘Recently,deep learning has been used to establish the nonlinear and nonintuitive mapping between physical structures and electromagnetic responses of meta-atoms for higher computational efficiency.However,to obtain sufficiently accurate predictions,the conventional deep-learning-based method consumes excessive time to collect the data set,thus hindering its wide application in this interdisciplinary field.We introduce a spectral transfer-learning-based metasurface design method to achieve excellent performance on a small data set with only 1000 samples in the target waveband by utilizing open-source data from another spectral range.We demonstrate three transfer strategies and experimentally quantify their performance,among which the“frozen-none”robustly improves the prediction accuracy by∼26%compared to direct learning.We propose to use a complex-valued deep neural network during the training process to further improve the spectral predicting precision by∼30%compared to its real-valued counterparts.We design several typical teraherz metadevices by employing a hybrid inverse model consolidating this trained target network and a global optimization algorithm.The simulated results successfully validate the capability of our approach.Our work provides a universal methodology for efficient and accurate metasurface design in arbitrary wavebands,which will pave the way toward the automated and mass production of metasurfaces.
文摘Deep neural networks(DNNs)are effective in solving both forward and inverse problems for nonlinear partial differential equations(PDEs).However,conventional DNNs are not effective in handling problems such as delay differential equations(DDEs)and delay integrodifferential equations(DIDEs)with constant delays,primarily due to their low regularity at delayinduced breaking points.In this paper,a DNN method that combines multi-task learning(MTL)which is proposed to solve both the forward and inverse problems of DIDEs.The core idea of this approach is to divide the original equation into multiple tasks based on the delay,using auxiliary outputs to represent the integral terms,followed by the use of MTL to seamlessly incorporate the properties at the breaking points into the loss function.Furthermore,given the increased training dificulty associated with multiple tasks and outputs,we employ a sequential training scheme to reduce training complexity and provide reference solutions for subsequent tasks.This approach significantly enhances the approximation accuracy of solving DIDEs with DNNs,as demonstrated by comparisons with traditional DNN methods.We validate the effectiveness of this method through several numerical experiments,test various parameter sharing structures in MTL and compare the testing results of these structures.Finally,this method is implemented to solve the inverse problem of nonlinear DIDE and the results show that the unknown parameters of DIDE can be discovered with sparse or noisy data.
基金supported by the Beijing Natural Science Foundation(Grant No.L223013)。
文摘For the diagnostics and health management of lithium-ion batteries,numerous models have been developed to understand their degradation characteristics.These models typically fall into two categories:data-driven models and physical models,each offering unique advantages but also facing limitations.Physics-informed neural networks(PINNs)provide a robust framework to integrate data-driven models with physical principles,ensuring consistency with underlying physics while enabling generalization across diverse operational conditions.This study introduces a PINN-based approach to reconstruct open circuit voltage(OCV)curves and estimate key ageing parameters at both the cell and electrode levels.These parameters include available capacity,electrode capacities,and lithium inventory capacity.The proposed method integrates OCV reconstruction models as functional components into convolutional neural networks(CNNs)and is validated using a public dataset.The results reveal that the estimated ageing parameters closely align with those obtained through offline OCV tests,with errors in reconstructed OCV curves remaining within 15 mV.This demonstrates the ability of the method to deliver fast and accurate degradation diagnostics at the electrode level,advancing the potential for precise and efficient battery health management.
文摘The ability to accurately predict urban traffic flows is crucial for optimising city operations.Consequently,various methods for forecasting urban traffic have been developed,focusing on analysing historical data to understand complex mobility patterns.Deep learning techniques,such as graph neural networks(GNNs),are popular for their ability to capture spatio-temporal dependencies.However,these models often become overly complex due to the large number of hyper-parameters involved.In this study,we introduce Dynamic Multi-Graph Spatial-Temporal Graph Neural Ordinary Differential Equation Networks(DMST-GNODE),a framework based on ordinary differential equations(ODEs)that autonomously discovers effective spatial-temporal graph neural network(STGNN)architectures for traffic prediction tasks.The comparative analysis of DMST-GNODE and baseline models indicates that DMST-GNODE model demonstrates superior performance across multiple datasets,consistently achieving the lowest Root Mean Square Error(RMSE)and Mean Absolute Error(MAE)values,alongside the highest accuracy.On the BKK(Bangkok)dataset,it outperformed other models with an RMSE of 3.3165 and an accuracy of 0.9367 for a 20-min interval,maintaining this trend across 40 and 60 min.Similarly,on the PeMS08 dataset,DMST-GNODE achieved the best performance with an RMSE of 19.4863 and an accuracy of 0.9377 at 20 min,demonstrating its effectiveness over longer periods.The Los_Loop dataset results further emphasise this model’s advantage,with an RMSE of 3.3422 and an accuracy of 0.7643 at 20 min,consistently maintaining superiority across all time intervals.These numerical highlights indicate that DMST-GNODE not only outperforms baseline models but also achieves higher accuracy and lower errors across different time intervals and datasets.
文摘The increasing popularity of the Internet and the widespread use of information technology have led to a rise in the number and sophistication of network attacks and security threats.Intrusion detection systems are crucial to network security,playing a pivotal role in safeguarding networks from potential threats.However,in the context of an evolving landscape of sophisticated and elusive attacks,existing intrusion detection methodologies often overlook critical aspects such as changes in network topology over time and interactions between hosts.To address these issues,this paper proposes a real-time network intrusion detection method based on graph neural networks.The proposedmethod leverages the advantages of graph neural networks and employs a straightforward graph construction method to represent network traffic as dynamic graph-structured data.Additionally,a graph convolution operation with a multi-head attention mechanism is utilized to enhance the model’s ability to capture the intricate relationships within the graph structure comprehensively.Furthermore,it uses an integrated graph neural network to address dynamic graphs’structural and topological changes at different time points and the challenges of edge embedding in intrusion detection data.The edge classification problem is effectively transformed into node classification by employing a line graph data representation,which facilitates fine-grained intrusion detection tasks on dynamic graph node feature representations.The efficacy of the proposed method is evaluated using two commonly used intrusion detection datasets,UNSW-NB15 and NF-ToN-IoT-v2,and results are compared with previous studies in this field.The experimental results demonstrate that our proposed method achieves 99.3%and 99.96%accuracy on the two datasets,respectively,and outperforms the benchmark model in several evaluation metrics.
基金supported by National key research and development program(No.2022YFA1602404)the National Natural Science Foundation of China(Nos.12388102,12275338,12005280)the Key Laboratory of Nuclear Data foundation(No.JCKY2022201C152)。
文摘This study investigates photonuclear reaction(γ,n)cross-sections using Bayesian neural network(BNN)analysis.After determining the optimal network architecture,which features two hidden layers,each with 50 hidden nodes,training was conducted for 30,000 iterations to ensure comprehensive data capture.By analyzing the distribution of absolute errors positively correlated with the cross-section for the isotope 159Tb,as well as the relative errors unrelated to the cross-section,we confirmed that the network effectively captured the data features without overfitting.Comparison with the TENDL-2021 Database demonstrated the BNN's reliability in fitting photonuclear cross-sections with lower average errors.The predictions for nuclei with single and double giant dipole resonance peak cross-sections,the accurate determination of the photoneutron reaction threshold in the low-energy region,and the precise description of trends in the high-energy cross-sections further demonstrate the network's generalization ability on the validation set.This can be attributed to the consistency of the training data.By using consistent training sets from different laboratories,Bayesian neural networks can predict nearby unknown cross-sections based on existing laboratory data,thereby estimating the potential differences between other laboratories'existing data and their own measurement results.Experimental measurements of photonuclear reactions on the newly constructed SLEGS beamline will contribute to clarifying the differences in cross-sections within the existing data.
基金funded by the Youth Fund of the National Natural Science Foundation of China(Grant No.42261070).
文摘Spectrum-based fault localization (SBFL) generates a ranked list of suspicious elements by using the program execution spectrum, but the excessive number of elements ranked in parallel results in low localization accuracy. Most researchers consider intra-class dependencies to improve localization accuracy. However, some studies show that inter-class method call type faults account for more than 20%, which means such methods still have certain limitations. To solve the above problems, this paper proposes a two-phase software fault localization based on relational graph convolutional neural networks (Two-RGCNFL). Firstly, in Phase 1, the method call dependence graph (MCDG) of the program is constructed, the intra-class and inter-class dependencies in MCDG are extracted by using the relational graph convolutional neural network, and the classifier is used to identify the faulty methods. Then, the GraphSMOTE algorithm is improved to alleviate the impact of class imbalance on classification accuracy. Aiming at the problem of parallel ranking of element suspicious values in traditional SBFL technology, in Phase 2, Doc2Vec is used to learn static features, while spectrum information serves as dynamic features. A RankNet model based on siamese multi-layer perceptron is constructed to score and rank statements in the faulty method. This work conducts experiments on 5 real projects of Defects4J benchmark. Experimental results show that, compared with the traditional SBFL technique and two baseline methods, our approach improves the Top-1 accuracy by 262.86%, 29.59% and 53.01%, respectively, which verifies the effectiveness of Two-RGCNFL. Furthermore, this work verifies the importance of inter-class dependencies through ablation experiments.
文摘Identifying cyberattacks that attempt to compromise digital systems is a critical function of intrusion detection systems(IDS).Data labeling difficulties,incorrect conclusions,and vulnerability to malicious data injections are only a few drawbacks of using machine learning algorithms for cybersecurity.To overcome these obstacles,researchers have created several network IDS models,such as the Hidden Naive Bayes Multiclass Classifier and supervised/unsupervised machine learning techniques.This study provides an updated learning strategy for artificial neural network(ANN)to address data categorization problems caused by unbalanced data.Compared to traditional approaches,the augmented ANN’s 92%accuracy is a significant improvement owing to the network’s increased resilience to disturbances and computational complexity,brought about by the addition of a random weight and standard scaler.Considering the ever-evolving nature of cybersecurity threats,this study introduces a revolutionary intrusion detection method.
基金supported by the National Key R&D Program of China(Grant No.:2022YFC3501805)the National Natural Science Foundation of China(Grant No.:82374030)+2 种基金the Science and Technology Program of Tianjin in China(Grant No.:23ZYJDSS00030)the Tianjin Outstanding Youth Fund,China(Grant No.:23JCJQJC00030)the China Postdoctoral Science Foundation-Tianjin Joint Support Program(Grant No.:2023T030TJ).
文摘Metabolomics covers a wide range of applications in life sciences,biomedicine,and phytology.Data acquisition(to achieve high coverage and efficiency)and analysis(to pursue good classification)are two key segments involved in metabolomics workflows.Various chemometric approaches utilizing either pattern recognition or machine learning have been employed to separate different groups.However,insufficient feature extraction,inappropriate feature selection,overfitting,or underfitting lead to an insufficient capacity to discriminate plants that are often easily confused.Using two ginseng varieties,namely Panax japonicus(PJ)and Panax japonicus var.major(PJvm),containing the similar ginsenosides,we integrated pseudo-targeted metabolomics and deep neural network(DNN)modeling to achieve accurate species differentiation.A pseudo-targeted metabolomics approach was optimized through data acquisition mode,ion pairs generation,comparison between multiple reaction monitoring(MRM)and scheduled MRM(sMRM),and chromatographic elution gradient.In total,1980 ion pairs were monitored within 23 min,allowing for the most comprehensive ginseng metabolome analysis.The established DNN model demonstrated excellent classification performance(in terms of accuracy,precision,recall,F1 score,area under the curve,and receiver operating characteristic(ROC))using the entire metabolome data and feature-selection dataset,exhibiting superior advantages over random forest(RF),support vector machine(SVM),extreme gradient boosting(XGBoost),and multilayer perceptron(MLP).Moreover,DNNs were advantageous for automated feature learning,nonlinear modeling,adaptability,and generalization.This study confirmed practicality of the established strategy for efficient metabolomics data analysis and reliable classification performance even when using small-volume samples.This established approach holds promise for plant metabolomics and is not limited to ginseng.
基金supported by the National Key Research and Development Program of China (Nos.2022YFC3702000 and 2022YFC3703500)the Key R&D Project of Zhejiang Province (No.2022C03146).
文摘Severe ground-level ozone(O_(3))pollution over major Chinese cities has become one of the most challenging problems,which have deleterious effects on human health and the sustainability of society.This study explored the spatiotemporal distribution characteristics of ground-level O_(3) and its precursors based on conventional pollutant and meteorological monitoring data in Zhejiang Province from 2016 to 2021.Then,a high-performance convolutional neural network(CNN)model was established by expanding the moment and the concentration variations to general factors.Finally,the response mechanism of O_(3) to the variation with crucial influencing factors is explored by controlling variables and interpolating target variables.The results indicated that the annual average MDA8-90th concentrations in Zhejiang Province are higher in the northern and lower in the southern.When the wind direction(WD)ranges from east to southwest and the wind speed(WS)ranges between 2 and 3 m/sec,higher O_(3) concentration prone to occur.At different temperatures(T),the O_(3) concentration showed a trend of first increasing and subsequently decreasing with increasing NO_(2) concentration,peaks at the NO_(2) concentration around 0.02mg/m^(3).The sensitivity of NO_(2) to O_(3) formation is not easily affected by temperature,barometric pressure and dew point temperature.Additionally,there is a minimum IRNO_(2) at each temperature when the NO_(2) concentration is 0.03 mg/m^(3),and this minimum IRNO_(2) decreases with increasing temperature.The study explores the response mechanism of O_(3) with the change of driving variables,which can provide a scientific foundation and methodological support for the targeted management of O_(3) pollution.
基金The National Natural Science Foundation of China(W2431048)The Science and Technology Research Program of Chongqing Municipal Education Commission,China(KJZDK202300807)The Chongqing Natural Science Foundation,China(CSTB2024NSCQQCXMX0052).
文摘This paper addresses the consensus problem of nonlinear multi-agent systems subject to external disturbances and uncertainties under denial-ofservice(DoS)attacks.Firstly,an observer-based state feedback control method is employed to achieve secure control by estimating the system's state in real time.Secondly,by combining a memory-based adaptive eventtriggered mechanism with neural networks,the paper aims to approximate the nonlinear terms in the networked system and efficiently conserve system resources.Finally,based on a two-degree-of-freedom model of a vehicle affected by crosswinds,this paper constructs a multi-unmanned ground vehicle(Multi-UGV)system to validate the effectiveness of the proposed method.Simulation results show that the proposed control strategy can effectively handle external disturbances such as crosswinds in practical applications,ensuring the stability and reliable operation of the Multi-UGV system.
基金founded by the Open Project Program of Anhui Province Key Laboratory of Metallurgical Engineering and Resources Recycling(Anhui University of Technology)(No.SKF21-06)Research Fund for Young Teachers of Anhui University of Technology in 2020(No.QZ202001).
文摘Real-time prediction and precise control of sinter quality are pivotal for energy saving,cost reduction,quality improvement and efficiency enhancement in the ironmaking process.To advance,the accuracy and comprehensiveness of sinter quality prediction,an intelligent flare monitoring system for sintering machine tails that combines hybrid neural networks integrating convolutional neural network with long short-term memory(CNN-LSTM)networks was proposed.The system utilized a high-temperature thermal imager for image acquisition at the sintering machine tail and employed a zone-triggered method to accurately capture dynamic feature images under challenging conditions of high-temperature,high dust,and occlusion.The feature images were then segmented through a triple-iteration multi-thresholding approach based on the maximum between-class variance method to minimize detail loss during the segmentation process.Leveraging the advantages of CNN and LSTM networks in capturing temporal and spatial information,a comprehensive model for sinter quality prediction was constructed,with inputs including the proportion of combustion layer,porosity rate,temperature distribution,and image features obtained from the convolutional neural network,and outputs comprising quality indicators such as underburning index,uniformity index,and FeO content of the sinter.The accuracy is notably increased,achieving a 95.8%hit rate within an error margin of±1.0.After the system is applied,the average qualified rate of FeO content increases from 87.24%to 89.99%,representing an improvement of 2.75%.The average monthly solid fuel consumption is reduced from 49.75 to 46.44 kg/t,leading to a 6.65%reduction and underscoring significant energy saving and cost reduction effects.
基金supported by the National Natural Science Foundation of China(No.U21B2062).
文摘The isolated fracture-vug systems controlled by small-scale strike-slip faults within ultra-deep carbonate rocks of the Tarim Basin exhibit significant exploration potential.The study employs a novel training set incorporating innovative fault labels to train a U-Net-structured CNN model,enabling effective identification of small-scale strike-slip faults through seismic data interpretation.Based on the CNN faults,we analyze the distribution patterns of small-scale strike-slip faults.The small-scale strike-slip faults can be categorized into NNW-trending and NE-trending groups with strike lengths ranging 200–5000 m.The development intensity of small-scale strike-slip faults in the Lower Yingshan Member notably exceeds that in the Upper Member.The Lower and Upper Yingshan members are two distinct mechanical layers with contrasting brittleness characteristics,separated by a low-brittleness layer.The superior brittleness of the Lower Yingshan Member enhances the development intensity of small-scale strike-slip faults compared to the upper member,while the low-brittleness layer exerts restrictive effects on vertical fault propagation.Fracture-vug systems formed by interactions of two or more small-scale strike-slip faults demonstrate larger sizes than those controlled by individual faults.All fracture-vug system sizes show positive correlations with the vertical extents of associated small-scale strike-slip faults,particularly intersection and approaching fracture-vug systems exhibit accelerated size increases proportional to the vertical extents.
基金supported by Interdisciplinary Innova-tion Project of“Bioarchaeology Laboratory”of Jilin University,China,and“MedicineþX”Interdisciplinary Innovation Team of Norman Bethune Health Science Center of Jilin University,China(Grant No.:2022JBGS05).
文摘Severe acute respiratory syndrome coronavirus 2(SARS-CoV-2)mutations are influenced by random and uncontrollable factors,and the risk of the next widespread epidemic remains.Dual-target drugs that synergistically act on two targets exhibit strong therapeutic effects and advantages against mutations.In this study,a novel computational workflow was developed to design dual-target SARS-CoV-2 candidate inhibitors with the Envelope protein and Main protease selected as the two target proteins.The drug-like molecules of our self-constructed 3D scaffold database were used as high-throughput molecular docking probes for feature extraction of two target protein pockets.A multi-layer perceptron(MLP)was employed to embed the binding affinities into a latent space as conditional vectors to control conditional distribution.Utilizing a conditional generative neural network,cG-SchNet,with 3D Euclidean group(E3)symmetries,the conditional probability distributions of molecular 3D structures were acquired and a set of novel SARS-CoV-2 dual-target candidate inhibitors were generated.The 1D probability,2D joint probability,and 2D cumulative probability distribution results indicate that the generated sets are significantly enhanced compared to the training set in the high binding affinity area.Among the 201 generated molecules,42 molecules exhibited a sum binding affinity exceeding 17.0 kcal/mol while 9 of them having a sum binding affinity exceeding 19.0 kcal/mol,demonstrating structure diversity along with strong dual-target affinities,good absorption,distribution,metabolism,excretion,and toxicity(ADMET)properties,and ease of synthesis.Dual-target drugs are rare and difficult to find,and our“high-throughput docking-multi-conditional generation”workflow offers a wide range of options for designing or optimizing potent dual-target SARS-CoV-2 inhibitors.
基金supported by the National Natural Science Foundation of China(Grant No.42076214)Natural Science Foundation of Shandong Province(Grant No.ZR2024QD057).
文摘Timely and accurate forecasting of storm surges can effectively prevent typhoon storm surges from causing large economic losses and casualties in coastal areas.At present,numerical model forecasting consumes too many resources and takes too long to compute,while neural network forecasting lacks regional data to train regional forecasting models.In this study,we used the DUAL wind model to build typhoon wind fields,and constructed a typhoon database of 75 processes in the northern South China Sea using the coupled Advanced Circulation-Simulating Waves Nearshore(ADCIRC-SWAN)model.Then,a neural network with a Res-U-Net structure was trained using the typhoon database to forecast the typhoon processes in the validation dataset,and an excellent storm surge forecasting effect was achieved in the Pearl River Estuary region.The storm surge forecasting effect of stronger typhoons was improved by adding a branch structure and transfer learning.
基金the North Dakota Industrial Commission (NDIC) for their financial supportprovided by the University of North Dakota Computational Research Center。
文摘Accurate estimation of mineralogy from geophysical well logs is crucial for characterizing geological formations,particularly in hydrocarbon exploration,CO_(2) sequestration,and geothermal energy development.Current techniques,such as multimineral petrophysical analysis,offer details into mineralogical distribution.However,it is inherently time-intensive and demands substantial geological expertise for accurate model evaluation.Furthermore,traditional machine learning techniques often struggle to predict mineralogy accurately and sometimes produce estimations that violate fundamental physical principles.To address this,we present a new approach using Physics-Integrated Neural Networks(PINNs),that combines data-driven learning with domain-specific physical constraints,embedding petrophysical relationships directly into the neural network architecture.This approach enforces that predictions adhere to physical laws.The methodology is applied to the Broom Creek Deep Saline aquifer,a CO_(2) sequestration site in the Williston Basin,to predict the volumes of key mineral constituents—quartz,dolomite,feldspar,anhydrite,illite—along with porosity.Compared to traditional artificial neural networks (ANN),the PINN approach demonstrates higher accuracy and better generalizability,significantly enhancing predictive performance on unseen well datasets.The average mean error across the three blind wells is 0.123 for ANN and 0.042 for PINN,highlighting the superior accuracy of the PINN approach.This method reduces uncertainties in reservoir characterization by improving the reliability of mineralogy and porosity predictions,providing a more robust tool for decision-making in various subsurface geoscience applications.
基金supported by Macao Science and Technology Development Fund,Macao SAR,China(Grant No.:0043/2023/AFJ)the National Natural Science Foundation of China(Grant No.:22173038)Macao Polytechnic University,Macao SAR,China(Grant No.:RP/FCA-01/2022).
文摘Accurate prediction of molecular properties is crucial for selecting compounds with ideal properties and reducing the costs and risks of trials.Traditional methods based on manually crafted features and graph-based methods have shown promising results in molecular property prediction.However,traditional methods rely on expert knowledge and often fail to capture the complex structures and interactions within molecules.Similarly,graph-based methods typically overlook the chemical structure and function hidden in molecular motifs and struggle to effectively integrate global and local molecular information.To address these limitations,we propose a novel fingerprint-enhanced hierarchical graph neural network(FH-GNN)for molecular property prediction that simultaneously learns information from hierarchical molecular graphs and fingerprints.The FH-GNN captures diverse hierarchical chemical information by applying directed message-passing neural networks(D-MPNN)on a hierarchical molecular graph that integrates atomic-level,motif-level,and graph-level information along with their relationships.Addi-tionally,we used an adaptive attention mechanism to balance the importance of hierarchical graphs and fingerprint features,creating a comprehensive molecular embedding that integrated hierarchical mo-lecular structures with domain knowledge.Experiments on eight benchmark datasets from MoleculeNet showed that FH-GNN outperformed the baseline models in both classification and regression tasks for molecular property prediction,validating its capability to comprehensively capture molecular informa-tion.By integrating molecular structure and chemical knowledge,FH-GNN provides a powerful tool for the accurate prediction of molecular properties and aids in the discovery of potential drug candidates.
