• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    MetaPINNs: Predicting soliton and rogue wave of nonlinear PDEs via the improved physics-informed neural networks based on meta-learned optimization

    2024-02-29 09:18:36YananGuo郭亞楠XiaoqunCao曹小群JunqiangSong宋君強(qiáng)andHongzeLeng冷洪澤
    Chinese Physics B 2024年2期
    關(guān)鍵詞:小群洪澤

    Yanan Guo(郭亞楠), Xiaoqun Cao(曹小群), Junqiang Song(宋君強(qiáng)), and Hongze Leng(冷洪澤)

    1College of Meteorology and Oceanography,National University of Defense Technology,Changsha 410073,China

    2College of Computer,National University of Defense Technology,Changsha 410073,China

    3Naval Aviation University,Huludao 125001,China

    Keywords: physics-informed neural networks, gradient-enhanced loss function, meta-learned optimization,nonlinear science

    1.Introduction

    In recent years, with the rapid development of artificial intelligence technology, especially deep learning, many fields have undergone significant transformations.Currently,deep learning technology has achieved remarkable milestones across various domains,including object detection,autonomous vehicles, and time series analysis.[1–3]Moreover,deep learning techniques are progressively shifting from being solely data-driven to being driven by the fusion of data and knowledge.It is worth paying attention to the rapid development of this research direction,and the related algorithms are beginning to penetrate into various fields.In the realm of scientific research,deep learning methods are becoming increasingly indispensable as research tools, profoundly advancing scientific discoveries,theoretical derivations,and other related research.They are exerting a rapid and profound influence on the development of disciplines such as physics, mathematics,materials science,and geosciences.[4–7]

    So far,scientific research driven by deep learning has resulted in a multitude of noteworthy accomplishments.For example, one notable application of deep learning methods has been in the field of weather and climate prediction, where it has substantially improved the accuracy of forecasts, particularly for highly impactful weather events like super typhoons and extreme precipitation.[8,9]Another exciting area of advancement is the utilization of deep neural networks for solving various types of partial differential equations (PDEs)and modeling complex natural physical phenomena.[10–13]In this context, Raissiet al.introduced the concept of physicsinformed neural networks (PINNs), which represents an improvement over traditional neural network-based methods for solving forward and inverse problems of partial differential equations.Numerous studies have shown that PINNs can effectively integrate the intrinsic constraint information in the physical equations.[14]Physics-informed neural networks serve as a comprehensive framework that combines datadriven and knowledge-driven approaches for modeling physical phenomena.PINNs differ from traditional neural networks in their unique ability to seamlessly incorporate the physical constraints of PDEs.This innovative approach reduces the need for extensive training data and enhances the interpretability of the model.PINNs are characterized by a flexible loss function that allows for the integration of PDE constraints, as well as boundary and initial condition constraints, in addition to observational data constraints.In recent years, a multitude of researchers have dedicated significant efforts to the study and enhancement of PINNs, introducing various innovative variants and applying them across diverse domains, including mechanics, geology, oceanography, and meteorology.[15–18]For example, Chenet al.proposed a variable coefficient physics-informed neural network(VC-PINN),[19]which specializes in dealing with the forward and inverse problems of variable coefficient partial differential equations.Caoet al.adopted an adaptive learning method to update the weight coefficients of the loss function for gradientenhanced PINNs and dynamically adjusted the weight proportion of each constraint term,[20]and numerically simulated localized waves and complex fluid dynamics phenomena,which achieved better prediction results than the traditional PINNs,and the improved method also performed well for the inverse problem under noise interference.

    Solitary waves and rogue waves represent the distinctive phenomenon frequently observed in various natural environments,[21–24]including the atmosphere and oceans.This research area is a cross-discipline of atmospheric science,ocean science,applied mathematics,and engineering technology and is considered an essential frontier research with farreaching theoretical significance and application prospects.[25]Consequently, solitary waves and rogue waves have emerged as a pivotal focus within the realm of geophysical fluid dynamics, spurring extensive theoretical exploration and highprecision simulation endeavors.However, traditional numerical computation methods encounter challenges such as low computational efficiency, substantial computational resource requirements, and error accumulation.Therefore, there is a growing imperative to explore novel numerical simulation techniques.For example, Chenet al.successfully solved the soliton solution of the nonlocal Hirota equation using a multilayer physics-informed neural network algorithm and investigated the parameter discovery problem of the nonlocal Hirota equation from the perspective of the soliton solution.[26]In addition, Chenet al.designed a two-stage PINN method based on conserved quantities that can impose physical constraints from a global perspective,[27]and applied it to multiple types of solitary wave simulations.Numerical results show that the two-stage PINN method has higher prediction accuracy and generalization ability compared with the original PINN method.Liet al.combined mixed-training and a priori information to propose mixed-training physicsinformed neural networks (PINNs) with enhanced approximation capabilities,[28]and used the model to simulate rogue waves for the nonlinear Schr?dinger (NLS) equation, and the numerical results show that the model can significantly improve the prediction accuracy compared with the original PINNs.Even when the spatio-temporal domain of the solution is enlarged or the solution has local sharp regions,the improved model still achieves accurate prediction results.Furthermore, Liet al.simulated higher-order rogue waves of the cmKdV equation using mixed-training physicsinformed neural networks (MTPINNs) and mixed-training physics-informed neural networks with a priori information(PMTPINNs),[29]which showed significantly better prediction accuracy compared to the original PINNs.In addition, these models have reliable robustness for the inverse problem under noisy data.In this study, we introduce a new structure and loss function specifically designed for modeling solitary and rogue wave phenomena using physics-informed neural networks.We conduct a comprehensive investigation into the simulation of solitary waves and rogue waves employing these enhanced PINNs and validate our approach through a series of numerical experiments.

    2.Methodology

    2.1.Physics-informed neural networks

    Physics-informed neural networks are a class of neural networks used for addressing supervised learning tasks while simultaneously adhering to the physical laws described by partial differential equations.As a machine learning method applied in the scientific computing, PINNs are employed to tackle a wide range of issues related to partial differential equations,encompassing approximate solutions of these equations, parameter inversion, model discovery, control, optimization, and more.After PINNs were proposed, a large amount of follow-up research work was triggered, and they gradually became a research hotspot in the emerging crosscutting field of scientific machine learning.From the perspective of mathematical function approximation theory,neural network can be regarded as a general nonlinear function approximator, and the modeling process of partial differential equations is also to find nonlinear functions that satisfy the constraints, which are quite consistent with each other.The neural network trained in this way not only approximates the observed data,but also automatically satisfies the physical properties of symmetry,invariance,and conservation followed by the partial differential equations.The following section describes the design method of PINNs using a partial differential equation in generalized form as an example.Let the functionu=u(x,t)satisfies a partial differential equation of the following form:

    whereN(u,λ)is a generalized function with parameterλthat performs differential operation onu,xis a spatial variable,tis a temporal variable,?is a subset of the Euclidean space,andTis the termination moment.The traditional physical model is that given the initial conditions, boundary conditions and physical parameterλofu(x,t), the value of theu(x,t)at any point in time and space can be predicted by solving the partial differential equation, and when an analytical solution is not available, it can be solved by using numerical methods, such as finite difference method and finite element method,etc.The PINNs consider the establishment of a neural networku(x,t)to approximate the solution of the partial differential equation.f(t,x)=ut+N(u,λ) is the residual of the partial differential equation, and the loss function of the PINN is defined as Loss=Lo+LIC+LBC+LPDE,where

    Fig.1.Schematic of physics-informed neural networks for solving partial differential equation.

    2.2.Meta learning

    While traditional deep learning methods use fixed learning algorithms to solve problems from scratch,meta-learning aims to improve the learning algorithm itself, which avoids many of the challenges inherent in deep learning, such as data and computational bottlenecks, and generalization problems.Meta-learning is divided into three main categories,namely metrics-based approaches, model-based approaches,and optimization-based approaches.Metric-based approaches learn a metric or distance function within a task,while modelbased approaches design an architecture or training process for fast generalization across tasks.Optimization-based methods achieve fast tuning by adjusting the optimization function using only a small number of samples.The framework of meta-learning is widely used in small-sample learning,and its core idea is to learn generalized prior knowledge on similar learning tasks so that only a few samples are needed, and the prior knowledge is used to quickly fine-tune on a new task to achieve the desired result.Therefore,the key of meta-learning is to discover the universal laws between different problems and utilize the universal laws to solve new tasks and problems.The universal laws need to achieve a balanced representation of the generalities and characteristics of the tasks.The search for universal laws mainly relies on the following points.

    (i) Discovering and summarizing the links between already solved problems and new problems, and extracting the universal laws of already solved problems for use in solving new problems.

    (ii)Decomposing and simplifying the new problem,finding universal laws in the already solved problem that are closely linked to the various subtasks of the new problem,and analyzing the scope of application of these laws.

    (iii)Learning logical reasoning in new tasks,using logical reasoning to make representations of new problems, finding laws in these representations, and finding solutions to complex new tasks through logical reasoning between the parts of the new problem.

    Meta-learning is essentially a bilevel optimization problem,where one optimization problem is nested within another optimization problem.The external optimization problem and the internal optimization problem are often referred to as the upper level optimization problem and the lower level optimization problem, respectively.The two-layer optimization problem involves two participants: the upper layer participant is the meta-learner and the lower layer participant is the baselearner.The optimal decision of the meta-learner depends on the response of the base-learner,which itself optimizes its own internal decisions.These two levels have their own different objective functions, constraints and decision variables.The objects and functions of the base-learner and meta-learner are shown in Fig.2.Base-learner is the model in the base layer.The data set on a single task is considered while training the base-learner each time.Meta-learner is the model in the metalayer that summarizes the training experience on all tasks.After training the base-learner each time, the meta-learner synthesizes the new experience and updates the parameters in the meta-learner.The main purpose of meta-learning is to find the meta-learnerF, under the guidance ofF, the base-learnerqcan get the optimal stateq*adapted to the current new task after a few steps of fine-tuning under the support setDtr.And the optimization ofFrequires the cumulative sum of the losses of all current tasks,i.e.,?∑Nn=1l(qn,Dten).The meta-learning works as shown in Fig.3.

    Fig.2.Schematic diagram of the base-learner and the meta-learner.

    Fig.3.The schematic of meta learning’s working mechanism.

    2.3.The improved physics-informed neural networks

    In the original PINNs,only the PDE residualsfare constrained to be zero.However, upon further analysis, it becomes evident that iff(x) is zero for anyx, then its derivative,?f(x),is also zero.Therefore,we make the assumption that the exact solution of the PDE possesses sufficient smoothness to warrant the existence of the gradient of PDE residuals, ?f(x).Consequently, we introduce gradient-enhanced PINNs, where not only are the PDE residuals zero, but also their derivatives, as described mathematically by the following equation:[30]

    Based on the aforementioned definition, it is inherent to derive the loss function for gradient-enhanced PINNs

    Here,Tgirepresents the set of residual points for the derivative?f/?xi.It is important to emphasize thatTgi(fori=1,2,...,d)may vary between different equations.To enhance the generalization capabilities of PINNs,this study integrates meta-learning into their training process.

    While the original optimizer uses manually designed optimization rules,such as the Adam optimizer for updating the weight vectorθof a physics-informed neural network with loss functionL(θ),meta-learning can improve upon these previously manually designed optimization rules.Specifically,the original Adam update rule is as follows:

    wheremtandvtare the first and second moments at time stept, respectively, withεbeing a small regularization constant used for numerical stability,β1,β2∈[0,1)being the exponential decay rates for the moment estimates, andηbeing the learning rate.The meta-learned optimizer’s parameters undergo updates through the following process:

    wherefis the parameter update function of the learnable optimizer,ztis the input feature, and?is the trainable metaparameter.For the component form,each parameterθiof the updated parameter vectorθis computed in the same way.The optimizer architecture used in this study utilizes a multilayer perceptron (MLP), which is simple, but better illustrates the potentiality of the approach in this paper.Specifically,the update formula for each parameterθiis as follows:

    (i)The neural network weightsθt.

    (ii)The gradients of loss function ?θL(θt).

    (iii) The second momentum accumulatorsvtwith decay ratesβ2∈{0.5,0.9,0.99,0.999}and inverse of the square root of these second momentum accumulators.

    (iv)The time stept.

    The present study improved physics-informed neural networks by utilizing the aforementioned input parameters.All input features, except for the time step, were normalized such that one of their second-order moment was equal to 1.The time step was transformed into a total of eleven features by computing tanh(t/x), wherex ∈{1,3,10,30,100,300,1000,3000,10k,30k,100k}.Subsequently, all features were concatenated and inputted into the MLP to produce the aforementioned three output heads.

    3.Numerical experiments and results

    In this section,we conduct experiments to investigate the solitary wave and rogue wave phenomena utilizing gradientenhanced PINNs with meta-learned optimization.Specifically,we seek to approximate the solitary wave solutions of the Kaup–Kuperschmidt (KK) equation and rogue waves of the nonlinear Schr?dinger equation.At the end of model training,we quantitatively evaluate the performance of the enhanced PINN model for solitary wave and rogue wave simulation.This assessment encompasses an in-depth analysis of its convergence rate and the accuracy of the prediction results.The accuracy of the approximate solutions generated by the enhanced PINN model is quantified using the L2 norm relative error,defined as follows:

    whereNrepresents the number of grid points, ?u(xi,ti) indicates the approximate solution generated by the neural network model,andu(xi,ti)represents the exact solution.

    3.1.Numerical simulation of solitary wave of the Kaup–Kuperschmidt equation

    Next, we conduct experimental research on the Kaup–Kuperschmidt equation,[31,32]which is mathematically defined as follows:

    The KK equation is commonly used to describe the evolution and nonlinear effects of water waves.It has a wide range of applications, including oceanography, fluid dynamics, and nonlinear optics,making it an important tool for studying various nonlinear wave phenomena.Meanwhile,researchers have found that the KK equation can support multiple solutions,including solitary waves and other nonlinear wave solutions,which play a crucial role in the propagation and evolution of water waves.However,the study and analysis of the KK equation are typically very complex because it is a highly nonlinear partial differential equation.

    For Eq.(13),we meticulously defined our numerical simulation parameters, limiting our spatial exploration to the interval [-20,20] and our temporal investigation to the period[-5,5].These parameter choices were guided by careful consideration of the following set of initial condition and boundary condition:

    In accordance with the previously stated conditions, the solitary wave solution of the KK equation is expressed as follows:

    whereθ1=k1x-k51tandu(x,t)describes the amplitude of the wave at positionxand timet.

    To effectively approximate the precise solution of the KK equation,we employed a neural network architecture consisting of five hidden layers,each comprising 64 neurons.In terms of the initial and boundary conditions,we generated a training dataset through random sampling, incorporating 400 samples.We selected random spatiotemporal locations to calculate residuals, facilitating the formulation of physical constraints.A total of 15000 data points were sampled during the experiment.Furthermore, a limited set of observations, consisting of 400 data points,was employed.Utilizing these data samples, we devised a composite loss function.We utilized a learnable optimizer to optimize the loss function, with the optimizer’s constant set asλj=10-4.Following the training process, the neural network model was employed to approximate the solitary wave solution of the Kaup–Kuperschmidt equation.In Fig.4,both the exact solution and the predictions generated by the trained model are presented.From Fig.4,it is evident that the enhanced PINNs can accurately predict the precise solution of the KK equation with high precision.Subsequently,for a detailed analysis of the experimental outcomes, we selectively examine the prediction results at different moments, comparing them with the exact solution to assess the accuracy of the improved PINNs.Figure 5 provides a visual representation of the comparison between the exact solution and the approximated solution at different times.The prediction values generated by the enhanced PINNs closely match the exact solutions, signifying a noteworthy degree of accuracy.Furthermore, following multiple experiments, we quantify the disparities between the approximate solution and the exact solution.The L2 norm relative error between the predictions generated by the enhanced PINNs and the exact solution is approximately 8.20×10-4.With the same training data and parameter configurations, the L2 norm relative error between the predictions made by the original PINNs and the true solution is approximately 5.22×10-3.Figure 6 displays the error statistics for both the improved PINNs and the original PINNs.Furthermore, we conduct an analysis of the model’s training time before and after the enhancements.The results indicate that the enhanced PINNs complete the training process in approximately 47.6%of the time required by the original PINNs.To more thoroughly analyze the effects of the number of neural network layers,the number of neurons and the number of collocation points on the prediction solution error, we chose different numbers of the hidden layers,different numbers of neurons and different numbers of collocation points to conduct numerical experiments and compared the prediction solution errors of the original PINNs and the improved PINNs.The experimental results are shown in Table 1.Analyzing the experimental results,it can be found that as the number of neural network layers increases,the number of neurons increases and the number of collocation points increases,the prediction solution error decreases.In addition,it is evident that the improved PINNs exhibit superior training efficiency and yield more accurate prediction outcomes compared to the original PINNs when simulating a single solitary wave of the KK equation.

    Table 1.For the KK equation,the errors of the predicted solutions are compared at different numbers of the neural network layers,different numbers of neurons,and different numbers of collocation points.

    Fig.4.Exact soliton solutions and predicted solutions of the KK equation: (a) two-dimensional visualization of the true solution, (b) threedimensional visualization of the true solution, (c) two-dimensional visualization of the predicted solution of the improved PINNs, (d) threedimensional visualization of the predicted solution of the improved PINNs.

    Fig.5.Comparison of the exact solitary wave solution to the predicted solitary wave solution of the KK equation at four different time instances:(a)t=-4.0,(b)t=-1.0,(c)t=1.0,and(d)t=4.0.

    Fig.6.Visual analysis of errors in the prediction results of the solitary wave solution of the KK equation: comparison of the errors of the improved PINNs and the original PINNs.

    3.2.Numerical simulation of first-order rogue wave of the nonlinear Schr?dinger equation

    In order to illustrate the applicability of the improved PINN to rogue wave prediction, we conduct experimental research on the nonlinear Schr?dinger equation,[33–35]which is mathematically defined as follows:

    The following solution is obtained by the Hirota transfor-mation:

    where

    In the above equations,?(0)nandφnare free complex parameters andkis a real parameter.In this subsection, the value ofkis zero.For the Eq.(18), the study domain for the numerical simulations is set to the spatial range[-5,5]and the time range[-5,5].When choosing these parameters,we carefully consider the following set of initial and boundary conditions:

    In the experiment,simulations are performed for the firstorder rogue wave of the nonlinear Schr?dinger equation.First,we construct a deep neural network with six hidden layers,each containing 64 neurons.During training,the loss function is optimized using a learnable optimizer with a constant set toλj=10-4.All weights and biases in the model are initialized using the Xavier method.For consistency and ease of performance comparison of subsequent methods, we still use the initial conditions as special Dirichlet boundary conditions on the spatio-temporal domain.It is worth noting that the number of observations is only 400, whereas 10000 collocation points were selected for the construction of the equation constraints and 400 sample points were selected for the construction of the initial condition constraints and boundary condition constraints.Figure 7 compares the difference between the prediction results of the improved PINNs and the true solution.The relatively good prediction results convince us that the improved PINNs can effectively predict the first-order rogue wave of the nonlinear Schr?dinger equation.Figure 8 shows the comparison of the prediction results between the exact solutions and the improved PINNs atdifferent moments,and we can find that the two solutions are in high consistency.Similarly,we calculate the L2 norm relative error of the prediction results of the improved PINNs with respect to the true solution,which has a value of 8.33×10-4,while the L2 norm relative error of the original PINNs with the same experimental setup is 6.41×10-3,and the comparison results are given in Fig.9.As for the computational efficiency,the time required for the improved PINNs to complete the training is only 51.6%of the training time of the original PINNs.In this study, we also quantitatively analyze the effects of the number of neural network layers, the number of neurons, and the number of collocation points on the prediction solution of the first-order rogue waves of the nonlinear Schr?dinger equation,and select different numbers of hidden layers, different numbers of neurons, and different numbers of collocation points for numerical experiments,and compare the prediction solution errors of the original PINNs and the improved PINNs.The specific experimental results are shown in Table 2.Analyzing the experimental results,it can be found that as the number of neural network layers increases, the number of neurons increases, and the number of collocation points increases,the prediction solution error is also reduced.The experiments can show that the improved PINNs outperform the original PINNs in terms of both computational efficiency and prediction accuracy.

    Table 2.For the first-order rogue wave of the nonlinear Schr?dinger equation,the errors of the predicted solutions are compared at different numbers of the neural network layers,different numbers of neurons,and different numbers of collocation points.

    Fig.7.Exact first-order rogue wave solutions and predicted solutions of the nonlinear Schr?dinger equation: (a)two-dimensional visualization of the true solution, (b) three-dimensional visualization of the true solution, (c) two-dimensional visualization of the predicted solution of the improved PINNs,(d)three-dimensional visualization of the predicted solution of the improved PINNs.

    Fig.8.Comparison of the exact first-order rogue wave solution to the predicted first-order rogue wave solution of the nonlinear Schr?dinger equation at four different time instances: (a)t=-3.0,(b)t=-1.0,(c)t=1.0,and(d)t=3.0.

    Fig.9.Visual analysis of errors in the prediction results of the first-orderrogue wave solution of the nonlinear Schr?dinger equation: comparison of the errors of the improved PINNs and the original PINNs.

    3.3.Numerical simulation of second-order rogue wave of the nonlinear Schro¨dinger equation

    In this subsection, we continue our study of the secondorder rogue wave of the nonlinear Schr?dinger equation.For the second-order rogue wave solution of Eq.(18), the study region for the numerical simulations is set to the spatial interval [-10,10] and the time interval [-10,10].In addition,the value ofkis likewise set to zero.With these parameters,we carefully consider the following set of initial and boundary conditions:

    In the above equation, the value ofCis 100.First, we construct a deep neural network that is with seven hidden layers and 64 neurons in each hidden layer.During the training process, the learnable optimizer is still utilized to optimize the loss function,and the constants of the optimizer are set toλj=10-4.In addition, we still use the initial conditions as special Dirichlet boundary conditions on the spatio-temporal domain.The number of observations in this experiment is 400,and 20000 collocation points were selected for the construction of the differential equation constraints, and 400 sample points were selected for the initial condition constraints and boundary condition constraints.Figure 10 compares the difference between the prediction results of the improved PINNs and the true solution, from which we find that the improved PINNs obtain prediction results close to the true solution.Figure 11 shows the comparison between the prediction results of the improved PINNs and the true solution at different moments, and again we can find a high consistency between the two.Similarly, we calculate the L2 norm relative error of the prediction results of the improved PINNs with respect to the true solution, which has a value of 7.10×10-4, whilethe L2 norm relative error of the original PINNs with the same experimental setup is 7.49×10-3,and the results of the comparison are given in Fig.12.Also analyzing the computational efficiency,the time required for the improved PINNs to complete the training is only 59.3% of the training time of the original PINNs.In this study, we have also taken different numbers of neural network layers, neurons, and collocation points as variables, and analyze the effects of these variables on the prediction solution of second-order rogue waves for the nonlinear Schr?dinger equation.In the specific experiments,we select different numbers of neural network layers,neurons,and collocation points for numerical experiments and compare the prediction solution errors of the original PINNs and the improved PINNs.The specific experimental results are shown in Table 3.Analyzing the experimental results, we can get the conclusion consistent with the previous experiments that with the increase of the number of neural network layers,the number of neurons,and the number of collocation points,the prediction solution error is also reduced.The experiment again demonstrates that for the second-order rogue wave prediction problem of the nonlinear Schr?dinger equation,the improved PINNs outperform the original PINNs in terms of both computational efficiency and prediction accuracy.

    Table 3.For the second-order rogue wave of the nonlinear Schr?dinger equation,the errors of the predicted solutions are compared at different numbers of the neural network layers,different numbers of neurons,and different numbers of collocation points.

    Fig.10.Exact second-order rogue wave solutions and predicted solutions of the nonlinear Schr?dinger equation: (a)two-dimensional visualization of the true solution, (b)three-dimensional visualization of the true solution, (c)two-dimensional visualization of the predicted solution of the improved PINNs,(d)three-dimensional visualization of the predicted solution of the improved PINNs.

    Fig.11.Comparison of the exact second-order rogue wave solution to the predicted second-order rogue wave solution of the nonlinear Schr?dinger equation at four different time instances: (a)t=-8.0,(b)t=-4.0,(c)t=0.0,and(d)t=4.0.

    Fig.12.Visual analysis of errors in the prediction results of the secondorder rogue wave solution of the nonlinear Schr?dinger equation: comparison of the errors of the improved PINNs and the original PINNs.

    4.Conclusion

    We have employed gradient-enhanced physics-informed neural networks(PINNs)with meta-learned optimization techniques for conducting numerical simulations of solitary waves and rogue waves.The enhanced PINNs incorporate gradient information, while the meta-learned optimization techniques significantly accelerate the training process of PINNs.As a result, the improved PINN model effectively harnesses more comprehensive physical insights and showcases superior generalization capabilities.Our experimental endeavors encompass the prediction of soliton solutions for the KK equation and rogue wave solution of the nonlinear Schr?dinger equation using these enhanced PINNs, followed by a comprehensive evaluation of the predictive outcomes.Notably, our experimental results clearly show that the training speed of these improved PINNs is greatly improved,while the prediction error is significantly reduced.This overall enhancement substantially elevates the accuracy and efficiency of solitary wave and rogue wave simulations.Given the enhanced capacity of PINNs to simulate solitary wave and rogue wave phenomena,our future research endeavors are aimed at extending their applicability to other intricate system simulations.These include ocean mesoscale eddy simulations,typhoon simulations,and atmospheric turbulence simulations, among others.Furthermore, we will explore the integration of PINNs with pretrained language models to further investigate the potential of its application in the field of complex physics simulation and to extend its application to physics problems at various scales.

    Acknowledgment

    Project supported by the National Natural Science Foundation of China(Grant Nos.42005003 and 41475094).

    猜你喜歡
    小群洪澤
    尖 叫
    百花園(2023年11期)2023-12-06 21:17:50
    以綠色高質(zhì)高效創(chuàng)建為抓手,助推洪澤區(qū)農(nóng)業(yè)高質(zhì)量跨越發(fā)展
    規(guī)范清潔高效機(jī)插精確生產(chǎn),強(qiáng)化“洪澤大米”區(qū)域品牌建設(shè)
    夏荷(新韻)
    洪澤『蟹』逅
    華人時刊(2020年13期)2020-09-25 08:21:52
    躲在樹上的雨
    罌粟花雖美不能栽
    明亮的眼睛
    長江圖
    民主(2017年3期)2017-05-12 09:48:20
    再為國夢譜華章
    ——江蘇省洪澤老年大學(xué)校歌
    亚洲一级一片aⅴ在线观看| 大又大粗又爽又黄少妇毛片口| 欧美精品啪啪一区二区三区| 人人妻,人人澡人人爽秒播| 久久精品影院6| 嫩草影院入口| 亚洲久久久久久中文字幕| 韩国av一区二区三区四区| 国产高潮美女av| 2021天堂中文幕一二区在线观| 悠悠久久av| 麻豆av噜噜一区二区三区| 日韩亚洲欧美综合| 最近在线观看免费完整版| 最近最新免费中文字幕在线| 丝袜美腿在线中文| 色播亚洲综合网| 嫩草影院新地址| 男女之事视频高清在线观看| 欧美性猛交╳xxx乱大交人| 欧美性猛交╳xxx乱大交人| 免费看美女性在线毛片视频| 偷拍熟女少妇极品色| 性插视频无遮挡在线免费观看| 99九九线精品视频在线观看视频| 精品不卡国产一区二区三区| 免费av不卡在线播放| 在线观看免费视频日本深夜| 亚州av有码| 少妇猛男粗大的猛烈进出视频 | 日韩高清综合在线| 精品一区二区三区av网在线观看| 人人妻人人看人人澡| 一级黄色大片毛片| 波多野结衣高清无吗| 久久精品夜夜夜夜夜久久蜜豆| 国产伦人伦偷精品视频| 人妻久久中文字幕网| 日韩,欧美,国产一区二区三区 | 日韩欧美国产在线观看| 午夜福利在线观看免费完整高清在 | 黄色配什么色好看| 国产伦精品一区二区三区视频9| 狠狠狠狠99中文字幕| 色视频www国产| 国产亚洲精品av在线| 成人国产一区最新在线观看| 偷拍熟女少妇极品色| 色视频www国产| 国产午夜精品论理片| 最近视频中文字幕2019在线8| 麻豆成人午夜福利视频| 波多野结衣巨乳人妻| 亚洲电影在线观看av| 久久中文看片网| 日韩高清综合在线| 国内精品久久久久精免费| 日本-黄色视频高清免费观看| 日韩欧美三级三区| 熟女人妻精品中文字幕| 国产伦一二天堂av在线观看| 一a级毛片在线观看| 欧美zozozo另类| 精品午夜福利视频在线观看一区| 亚洲经典国产精华液单| 国产精品免费一区二区三区在线| 韩国av一区二区三区四区| 在线天堂最新版资源| 国产91精品成人一区二区三区| 97热精品久久久久久| 久久久久久久久久久丰满 | 午夜亚洲福利在线播放| 一级av片app| 两性午夜刺激爽爽歪歪视频在线观看| 久久亚洲精品不卡| 久久久国产成人免费| 99精品在免费线老司机午夜| 99国产极品粉嫩在线观看| 久久久久久久午夜电影| 国产视频内射| 欧美激情国产日韩精品一区| 婷婷六月久久综合丁香| 美女cb高潮喷水在线观看| 大又大粗又爽又黄少妇毛片口| 国产精品综合久久久久久久免费| 欧美日韩乱码在线| 亚洲久久久久久中文字幕| 一级黄色大片毛片| 久久久久久国产a免费观看| 亚洲人与动物交配视频| 91在线观看av| 男女下面进入的视频免费午夜| 亚洲五月天丁香| 日韩欧美国产在线观看| 亚洲成人久久性| 色哟哟·www| 少妇人妻精品综合一区二区 | 又爽又黄无遮挡网站| 国产精品乱码一区二三区的特点| 高清日韩中文字幕在线| 51国产日韩欧美| 午夜日韩欧美国产| 一a级毛片在线观看| 国模一区二区三区四区视频| 精品国内亚洲2022精品成人| 97人妻精品一区二区三区麻豆| 岛国在线免费视频观看| 亚洲不卡免费看| 精品不卡国产一区二区三区| 久久精品影院6| 免费在线观看成人毛片| 看十八女毛片水多多多| 日韩欧美三级三区| 国产亚洲欧美98| 人妻制服诱惑在线中文字幕| 久久久久久久久久黄片| 免费在线观看成人毛片| 变态另类成人亚洲欧美熟女| 国产精品一区二区三区四区久久| 成年免费大片在线观看| 熟女电影av网| 欧美性感艳星| 中文在线观看免费www的网站| 亚洲第一区二区三区不卡| 国产av不卡久久| 麻豆精品久久久久久蜜桃| 亚洲国产日韩欧美精品在线观看| 国内少妇人妻偷人精品xxx网站| 国产91精品成人一区二区三区| 国产亚洲精品综合一区在线观看| 18+在线观看网站| 一个人看的www免费观看视频| 国产精品久久久久久亚洲av鲁大| 嫩草影院新地址| 国产av一区在线观看免费| 国产午夜福利久久久久久| 亚洲国产欧美人成| 国产久久久一区二区三区| 一个人看的www免费观看视频| 欧美日韩黄片免| 国产精品久久电影中文字幕| 亚洲在线自拍视频| av天堂中文字幕网| 丰满乱子伦码专区| 别揉我奶头 嗯啊视频| 动漫黄色视频在线观看| 在线观看美女被高潮喷水网站| 亚洲人成网站在线播| 久久香蕉精品热| 十八禁网站免费在线| 一进一出抽搐动态| 白带黄色成豆腐渣| 久久久久久久午夜电影| 国产av一区在线观看免费| 永久网站在线| 最近中文字幕高清免费大全6 | 在线天堂最新版资源| 麻豆国产97在线/欧美| 国产视频内射| 亚洲成人免费电影在线观看| 亚洲最大成人av| 如何舔出高潮| 亚洲专区国产一区二区| 嫩草影院精品99| 亚洲av不卡在线观看| av.在线天堂| 琪琪午夜伦伦电影理论片6080| 免费看日本二区| 在线观看66精品国产| 国产一区二区激情短视频| 久久精品国产亚洲网站| 99视频精品全部免费 在线| 麻豆国产97在线/欧美| 老司机午夜福利在线观看视频| 麻豆成人av在线观看| 天堂√8在线中文| 亚洲一区高清亚洲精品| 成年女人看的毛片在线观看| 日本爱情动作片www.在线观看 | 亚洲一区二区三区色噜噜| 欧美性感艳星| 99热精品在线国产| 久久这里只有精品中国| 精品人妻熟女av久视频| 精品久久久久久久人妻蜜臀av| 国产一区二区三区av在线 | 免费观看人在逋| 级片在线观看| 日韩欧美精品v在线| a在线观看视频网站| 国产在线男女| 国产一区二区在线观看日韩| 午夜激情欧美在线| 在线天堂最新版资源| 国产精品精品国产色婷婷| 国产av一区在线观看免费| 日本黄大片高清| 国产精品女同一区二区软件 | 国产精品自产拍在线观看55亚洲| av在线老鸭窝| 国产真实伦视频高清在线观看 | 国产精品综合久久久久久久免费| 久久久久国内视频| 两性午夜刺激爽爽歪歪视频在线观看| 欧美日韩中文字幕国产精品一区二区三区| 老司机福利观看| 一进一出抽搐gif免费好疼| 精品免费久久久久久久清纯| 成人精品一区二区免费| 我的女老师完整版在线观看| 久久99热6这里只有精品| 女同久久另类99精品国产91| 日本爱情动作片www.在线观看 | 精品人妻偷拍中文字幕| 欧美在线一区亚洲| 在线播放无遮挡| 99久久中文字幕三级久久日本| 成年免费大片在线观看| 91久久精品电影网| 免费观看的影片在线观看| 欧美日本视频| 亚洲av.av天堂| 亚洲成a人片在线一区二区| 亚洲人成网站在线播放欧美日韩| 乱人视频在线观看| 亚洲人成网站高清观看| 精品一区二区免费观看| 成熟少妇高潮喷水视频| 国产爱豆传媒在线观看| 午夜福利视频1000在线观看| 国产精品亚洲美女久久久| 国产精品国产三级国产av玫瑰| 麻豆av噜噜一区二区三区| 午夜福利在线观看免费完整高清在 | 国产av在哪里看| 中国美女看黄片| 韩国av一区二区三区四区| 亚洲av美国av| 久久精品国产亚洲av天美| 国产三级中文精品| 简卡轻食公司| 三级国产精品欧美在线观看| 欧美成人一区二区免费高清观看| 成人国产综合亚洲| 亚洲av一区综合| 精品国内亚洲2022精品成人| 99热网站在线观看| 午夜福利成人在线免费观看| av中文乱码字幕在线| 高清毛片免费观看视频网站| 欧美不卡视频在线免费观看| a级毛片a级免费在线| av黄色大香蕉| 色噜噜av男人的天堂激情| bbb黄色大片| 又紧又爽又黄一区二区| 色播亚洲综合网| 欧美日本视频| a级毛片a级免费在线| 亚洲人成网站高清观看| 高清日韩中文字幕在线| 日本在线视频免费播放| 最好的美女福利视频网| 亚洲人与动物交配视频| 亚洲中文字幕一区二区三区有码在线看| 欧美成人性av电影在线观看| 久久久久久久久久成人| 露出奶头的视频| 九色国产91popny在线| 亚洲精品成人久久久久久| 亚洲人与动物交配视频| ponron亚洲| 俺也久久电影网| 特级一级黄色大片| 欧美日韩中文字幕国产精品一区二区三区| 在线观看舔阴道视频| 亚洲国产精品sss在线观看| 一进一出抽搐动态| ponron亚洲| 一本久久中文字幕| 国产亚洲精品久久久久久毛片| 观看美女的网站| 可以在线观看毛片的网站| 国产蜜桃级精品一区二区三区| 男人狂女人下面高潮的视频| 精品人妻1区二区| 午夜福利视频1000在线观看| 内射极品少妇av片p| 成人国产麻豆网| 久久九九热精品免费| 人妻丰满熟妇av一区二区三区| 老熟妇仑乱视频hdxx| 观看美女的网站| 又黄又爽又刺激的免费视频.| 少妇人妻一区二区三区视频| 国产日本99.免费观看| 国产高潮美女av| 男人的好看免费观看在线视频| 人妻制服诱惑在线中文字幕| 日本爱情动作片www.在线观看 | 国产av在哪里看| 别揉我奶头~嗯~啊~动态视频| 国产熟女欧美一区二区| 国产成人影院久久av| 成人高潮视频无遮挡免费网站| 中文字幕熟女人妻在线| 日韩在线高清观看一区二区三区 | 黄色丝袜av网址大全| 深夜a级毛片| 三级毛片av免费| 999久久久精品免费观看国产| 性插视频无遮挡在线免费观看| 日本熟妇午夜| 亚洲精品国产成人久久av| 成年女人毛片免费观看观看9| 欧美日韩黄片免| 久久久久久久精品吃奶| 成人欧美大片| 国产精品一及| 男女那种视频在线观看| 美女黄网站色视频| 国产成人a区在线观看| 久久久久久久久大av| 九九爱精品视频在线观看| 亚洲精品成人久久久久久| 美女高潮的动态| 久久欧美精品欧美久久欧美| 国产中年淑女户外野战色| 日本免费一区二区三区高清不卡| 亚洲国产色片| 精品人妻偷拍中文字幕| 少妇人妻精品综合一区二区 | 国产精品不卡视频一区二区| 国产av不卡久久| 成人美女网站在线观看视频| 国产成年人精品一区二区| 久久精品国产鲁丝片午夜精品 | 99精品在免费线老司机午夜| 麻豆国产97在线/欧美| 亚洲欧美日韩高清在线视频| 天堂动漫精品| 亚洲三级黄色毛片| 天堂√8在线中文| 亚洲欧美日韩高清在线视频| 日韩欧美三级三区| 国产 一区 欧美 日韩| 久久99热这里只有精品18| 校园春色视频在线观看| 九九久久精品国产亚洲av麻豆| 精品久久久久久成人av| 国产男靠女视频免费网站| 国产美女午夜福利| 嫩草影视91久久| 热99re8久久精品国产| 日韩一本色道免费dvd| 可以在线观看的亚洲视频| 身体一侧抽搐| 亚洲国产色片| 大型黄色视频在线免费观看| aaaaa片日本免费| 日韩精品青青久久久久久| 午夜激情欧美在线| 国产精品福利在线免费观看| 中出人妻视频一区二区| 黄片wwwwww| 老司机福利观看| 国产不卡一卡二| 午夜福利成人在线免费观看| 国产免费av片在线观看野外av| 欧美高清性xxxxhd video| 欧美bdsm另类| 尤物成人国产欧美一区二区三区| 91麻豆精品激情在线观看国产| 高清在线国产一区| 美女黄网站色视频| 联通29元200g的流量卡| 久久久久久久午夜电影| 色5月婷婷丁香| 国内毛片毛片毛片毛片毛片| 亚洲国产精品合色在线| 久久午夜福利片| 变态另类成人亚洲欧美熟女| 日本一二三区视频观看| 身体一侧抽搐| 美女 人体艺术 gogo| 午夜免费成人在线视频| 99久国产av精品| 国产不卡一卡二| 亚洲精品乱码久久久v下载方式| 男女做爰动态图高潮gif福利片| 亚洲久久久久久中文字幕| 久久婷婷人人爽人人干人人爱| 国产av一区在线观看免费| 久久中文看片网| 精品久久久噜噜| 波多野结衣高清作品| 成人综合一区亚洲| 成人特级av手机在线观看| 久久久久久久久中文| 黄色视频,在线免费观看| 中文资源天堂在线| 99国产精品一区二区蜜桃av| 黄色日韩在线| 午夜精品在线福利| 亚洲专区中文字幕在线| 女人被狂操c到高潮| 一级a爱片免费观看的视频| 亚洲av不卡在线观看| 国产高清激情床上av| 日韩欧美国产在线观看| 床上黄色一级片| 看片在线看免费视频| 综合色av麻豆| 内射极品少妇av片p| 69av精品久久久久久| 国语自产精品视频在线第100页| 最近最新免费中文字幕在线| 国内精品美女久久久久久| 亚洲国产色片| 久久人人精品亚洲av| 亚洲精品456在线播放app | 国产又黄又爽又无遮挡在线| 亚洲国产欧洲综合997久久,| 精品久久久久久久人妻蜜臀av| 亚洲性夜色夜夜综合| 国产精品99久久久久久久久| 色精品久久人妻99蜜桃| 国产一区二区三区视频了| 国产麻豆成人av免费视频| 人人妻,人人澡人人爽秒播| 免费看日本二区| 草草在线视频免费看| av黄色大香蕉| 高清日韩中文字幕在线| 亚洲美女搞黄在线观看 | 久久精品国产亚洲av天美| 久久欧美精品欧美久久欧美| 99久久精品国产国产毛片| 亚洲专区国产一区二区| 啪啪无遮挡十八禁网站| 国产探花极品一区二区| 久久精品综合一区二区三区| 国产欧美日韩精品一区二区| 九九久久精品国产亚洲av麻豆| 亚洲精品456在线播放app | 无人区码免费观看不卡| 国产精品一区二区性色av| 日韩高清综合在线| 伦理电影大哥的女人| 亚洲av一区综合| 韩国av一区二区三区四区| 精品久久国产蜜桃| 亚洲欧美日韩高清在线视频| 亚洲av.av天堂| 欧洲精品卡2卡3卡4卡5卡区| 精品一区二区免费观看| 国产成人av教育| 国产极品精品免费视频能看的| 久久久久国内视频| 成年版毛片免费区| 两个人视频免费观看高清| 赤兔流量卡办理| 女生性感内裤真人,穿戴方法视频| 国产在视频线在精品| 少妇被粗大猛烈的视频| 桃色一区二区三区在线观看| 99久久久亚洲精品蜜臀av| 春色校园在线视频观看| 又紧又爽又黄一区二区| av视频在线观看入口| 中文字幕av在线有码专区| 亚洲国产欧洲综合997久久,| 老女人水多毛片| 久久精品久久久久久噜噜老黄 | 欧美高清成人免费视频www| 美女高潮喷水抽搐中文字幕| 欧美国产日韩亚洲一区| 欧美日韩精品成人综合77777| 成人二区视频| 2021天堂中文幕一二区在线观| 热99re8久久精品国产| 成人毛片a级毛片在线播放| 俺也久久电影网| 麻豆成人av在线观看| 亚洲精品影视一区二区三区av| 一个人看视频在线观看www免费| 极品教师在线免费播放| 成人精品一区二区免费| 国产精品三级大全| 联通29元200g的流量卡| 亚洲熟妇中文字幕五十中出| 午夜久久久久精精品| 黄色女人牲交| 久久精品国产亚洲av天美| 少妇丰满av| 欧美极品一区二区三区四区| 国产三级在线视频| 亚洲成a人片在线一区二区| 欧美最新免费一区二区三区| 美女xxoo啪啪120秒动态图| 国产欧美日韩精品一区二区| 国产私拍福利视频在线观看| 尤物成人国产欧美一区二区三区| 国产主播在线观看一区二区| 一卡2卡三卡四卡精品乱码亚洲| 国产男靠女视频免费网站| 亚洲人成伊人成综合网2020| 此物有八面人人有两片| 久久久久久久精品吃奶| 少妇熟女aⅴ在线视频| 少妇裸体淫交视频免费看高清| 韩国av一区二区三区四区| a级毛片a级免费在线| 波多野结衣巨乳人妻| 九九热线精品视视频播放| 国产在线男女| 露出奶头的视频| 午夜日韩欧美国产| 狂野欧美激情性xxxx在线观看| 99久久精品一区二区三区| 国产成人福利小说| 亚洲精品粉嫩美女一区| 精品久久久久久久久久免费视频| 他把我摸到了高潮在线观看| 国产精品久久电影中文字幕| 国产高清激情床上av| 国产在线精品亚洲第一网站| 亚洲熟妇熟女久久| 麻豆一二三区av精品| 色综合站精品国产| 在线国产一区二区在线| 国产精品日韩av在线免费观看| 国产av麻豆久久久久久久| 欧美性猛交黑人性爽| 免费av观看视频| 嫩草影院精品99| 亚洲性久久影院| 69人妻影院| 精品福利观看| 欧美成人一区二区免费高清观看| 久久香蕉精品热| 成年人黄色毛片网站| 一区二区三区高清视频在线| 国产亚洲av嫩草精品影院| 一区二区三区激情视频| 成年人黄色毛片网站| 人妻夜夜爽99麻豆av| 午夜精品久久久久久毛片777| ponron亚洲| 午夜激情福利司机影院| 国产男靠女视频免费网站| 干丝袜人妻中文字幕| www日本黄色视频网| 一卡2卡三卡四卡精品乱码亚洲| 久久精品夜夜夜夜夜久久蜜豆| 尤物成人国产欧美一区二区三区| 他把我摸到了高潮在线观看| 老女人水多毛片| АⅤ资源中文在线天堂| 午夜视频国产福利| 美女xxoo啪啪120秒动态图| 老司机深夜福利视频在线观看| 天美传媒精品一区二区| 少妇高潮的动态图| 成人无遮挡网站| 亚洲电影在线观看av| 别揉我奶头~嗯~啊~动态视频| 日韩精品青青久久久久久| 国产亚洲欧美98| 美女免费视频网站| 欧美极品一区二区三区四区| 久久99热这里只有精品18| 国产综合懂色| 哪里可以看免费的av片| 午夜免费男女啪啪视频观看 | 久久久久久久精品吃奶| 亚洲专区国产一区二区| 美女cb高潮喷水在线观看| 午夜福利高清视频| 国产日本99.免费观看| av国产免费在线观看| 黄色丝袜av网址大全| 十八禁网站免费在线| 中文字幕av在线有码专区| 欧美日韩乱码在线| 亚洲经典国产精华液单| 春色校园在线视频观看| 国内少妇人妻偷人精品xxx网站| 日本撒尿小便嘘嘘汇集6| 婷婷色综合大香蕉| 老师上课跳d突然被开到最大视频| 亚洲av成人av| 韩国av一区二区三区四区| 欧美在线一区亚洲| 日本黄色视频三级网站网址| 在线免费十八禁| 国产男靠女视频免费网站| 校园春色视频在线观看| 国产91精品成人一区二区三区| 美女被艹到高潮喷水动态| 偷拍熟女少妇极品色| 精品福利观看| 国产精品嫩草影院av在线观看 | 精品免费久久久久久久清纯| 三级男女做爰猛烈吃奶摸视频| 国产欧美日韩一区二区精品| 亚洲真实伦在线观看| 中文字幕免费在线视频6| 99久久中文字幕三级久久日本| 国产精品野战在线观看| 我要搜黄色片| 小说图片视频综合网站| 一进一出抽搐动态| 日本色播在线视频| 观看免费一级毛片|