AU2020102874A4 - A recommendation model for aero dynamic design of structures using deep recurrent neural network - Google Patents

Abstract

A RECOMMENDATION MODEL FOR AERO DYNAMIC DESIGN OF STRUCTURES USING DEEP RECURRENT NEURAL NETWORK ABSTRACT During the last 60 years, aerodynamic optimization seems to be an essential part of any aerodynamic architecture, with implementations for aircraft, vehicles, trains, bridges, wind turbines, internal pipe flows and cavities, among many others, and is also important to several areas of technology. With advances in computing resources, automatic design optimization systems have becoming more professional. As illustrated by the non-free lunch theorem, no algorithm will exceed each another on all levels of optimization issues. To resolve such concern, approaches have been suggested for proposing an emerging methodology for problem-solving. Even so, current recommendation models for continuous optimization suffer from poor enforceability and employability, primarily due to the difficulties in obtaining characteristics that can accurately define the major issue configuration and insufficient evidence for the learning of the recommendation model. This work suggests a generalized advice framework to resolve the two problems lay out above. First, a novel approach is suggested to describe the analytical objective function of a continuous optimization problem as a tree that is specifically employed as a component of the issue. In the case of black-box optimization problems which objective function is undefined, a symbolic regressor is used to approximate the tree structure. Using a deep recurrent neural network, a recommendation model is learned to suggest the most effective metaheuristic algorithm for white-or black-box optimization, allowing a big move forward towards a truly automatic algorithm recommendation for continuous optimization. Analytical results on 100,000 benchmark issues revealed that the optimal guidance approach ensures substantially higher efficiency than current models and shows high accessibility to real-world problems. 1| P a g e Algmthn 1 Algantm - - - AlpFnxm9m Sahie ,, Solve Solve ohPrTNm 1 Ponbem 2 ---- Prlo 1n A- ---------------- --------------------- Inde Of e Ilxofte xOof p t ha rn hm 1s t ah mh -.- t l:. n t rar rr I t f I l e IIfier Ferresdo Felmresof Festuses of Iqmt Feture extactionstrategy Problnm1 Problemi2 - - rbe sample I stree2 5amLeii Figure 3: A generic framework for algorithm recommendation IRandom nadchakproblems Train De eurn neural network wite-box p Tree Ip O Indexofthe pblem best algonthm Black-bOX mboli Esinted treel I"P problem [el I~ Figure 4: The design of recommendation system for aero dynamic design of structures 21Page

Description

Algmthn 1 Algantm - - - AlpFnxm9m

Sahie ,, Solve Solve ohPrTNm 1 Ponbem 2 ---- Prlo 1n

Inde Of e Ilxofte xOof p t ha rn A- hm ---------------- 1s t ah mh --------------------- -.- t l:. n t rar rr I t f I l e IIfier

Ferresdo Felmresof Festuses of Iqmt

Feture extactionstrategy

Problnm1 Problemi2 - - rbe sample I stree2 5amLeii

Figure 3: A generic framework for algorithm recommendation

IRandom nadchakproblems Train De eurn neural network

wite-box p Tree Ip O Indexofthe pblem best algonthm

Black-bOX mboli Esinted treel I"P problem I~ [el

Figure 4: The design of recommendation system for aero dynamic design of structures

21Page

A RECOMMENDATION MODEL FOR AERO DYNAMIC DESIGN OF STRUCTURES USING DEEP RECURRENT NEURAL NETWORK

Description

Field of the Invention:

This invention relates to make a recommendation model for aero dynamic design of structures using Deep Recurrent neural network. Aero dynamic data modeling contributes to the interpretation of the mapping mechanism among input and output data leveraging the relevant models. The deep recurrent neural network (Deep-RNN) utilized the idea to learn the classification process design.

Background of the invention:

Influenced by the dynamics of evolutionary processes and swarm behaviour in general, several metaheuristics have evolved extensively during the last generations, including genetic algorithms, particle swarm optimization, differential evolution, ant colony optimization, and among several others. These techniques are high-level strategies which do not depend on the distinct needs of the challenges and have demonstrated attractive competition in tackling various constrained optimization problems. Nevertheless, as shown by the non-free lunch theorem, there is no single methodology which can perform better anything else on all levels of computation problem. A great deal of work has therefore been invested to build strategies suited to particular kinds of problems. While the layouts of specialized applications which demand extensive field expertise, it is challenging for developers to customize evolutionary methods for variety of applications. Rather, current implementations are sometimes chosen for various technologies.

The synthesis of aerodynamics and artificial intelligence stretches prior to the 1940's. Kolmogorov has introduced mathematical training approaches to solve disruption issues. The neural networks are a research focused on mathematical learning. McCulloch and Pitts promoted the initial definition of MLPs. They aimed to understand the process by which the human brain performs complicated activities. Rosenblatt suggested a single-layer perceptron study explored the effects. Even then, it was discovered in 1969 that the perceptron for only one hidden units was unable to

11 P a g e recognize the XOR method. Stimulated by this downside, the configuration of the MLP was suggested by Minsky, yet how does one learn the variable of this template remains unresolved.

According to Zhang that the implementation of aero-elastic optimization with flush restrictions prevents consistency problems associated with dynamic mesh deformation by using a fixed over set mesh approach. Nevertheless, the challenge of constructing a mesh a priori that retains sufficient for the entire construction quest region is always unfeasible. Some metrics, like induced drag, are acutely vulnerable to grid size, and thus the aerodynamic characteristics of the system must be seen to be grid-independent for the core design region for the mesh used. If the aerodynamic features of the body are not representative of the grid, it is very simple for the framework for becoming unsuitable when the geometry and/or flow situations change the optimization technique. This will relate to undefined optimization knowledge and make the optimization phase worthless, as discretization failures of the mesh are manipulated contributing to false layouts.

In accordance with Zhao the Multi-Island GA (MIGA) is employed to refine the flow rate over the aerofoil. MIGA works by splitting the population into sub-populations insulating various locations of the construction area. Alternatives thereby evolve into individual categories with distinct criteria and are routinely encouraged to move among 'islands' and communicate with each generation. A population of 50 solutions distributed over 5 sub-populations was analyzed over 50 generations of approach replication needed each 4 decades. The high level of migrations needed to sustain complexity and low subgroups indicate that the method was quite vulnerable to genetic drift.

Rumelhart et al suggested a back propagation algorithm that tackled the learning issue of MLPs. As of then, MLPs have been turned into Neural Networks that signify the beginning of neural networks. Even then, owing to the lack of computing resources in those periods, and technical constraints, the total amount of secret layers in MLPs is two. The creation of neural networks has been limited and has still reached the winter era.

Nyan et al presented the use of accelerometers and gyroscopes for the device. In this job, the sensors are attached to a board mounted on the Intel PXA255 processor using Zigbee transceivers, in which the necessary data is being processed. Elsewhere in system, a specialized device built on FPGA technology develops information from accelerometers.

21Page

Yildirim et al explains the technique in which the solver reviews the configuration of every nonlinear execution leveraging the backward Euler step size equation. To ease the challenging initialization process, we employ an adaptive CFL methodology focusing on a pseudo-transient continuation. This begins with a limited time step towards reliability and then raises the step size quickly as it reaches a resolution that has the desirable durability features of the backward Euler method during the initial stage, thus attempting a Newton-type algorithm as the cycle time reaches infinite. For every time varying process, an upgrade solution is generated by incorrectly resolving a broad linear structure.

According to Chernukin and Zingg the limited experiments have performed about how the range of factors employed and the associated methodology can influence aerodynamic structures and point out that the distinction among computational models and weak optimizer consolidation can render troublesome. It can be predicted, although never assured, to raise the formulation of the objective function by growing the complexity of the design process. All 8 strategies reached both the optimum and the viability specifications with an objective value differing by about 5 per cent between the regional maxims. The outlines of the design are distinctive and thus indicate that the structural difference among the local optim has identical features of results.

Objects of the Invention:

The main objective is to implement a recommendation model for aero dynamic design of structures using Deep Recurrent Neural Network.

The second objective is to extract the features from continuous optimization. In the case of black box challenges whose objective function is uncertain, a number of solutions are extracted and their objective values are calculated; a symbolic regressor is then used to approximate the structure for knowing the configurations among the strategies and their empirical variables.

Another is to posit a deep learning classifier for recommend the most suitable metaheuristics for combinatorial computation challenges. In order to train the neural network, a large number of benchmark problems are randomly generated with the aid of the tree structure, and their classifications are gained by examining multiple metaheuristics from each concern.

Summary of the Invention:

31Page

The mentioned approach advocates an innovative tree structure to describe a given tree challenge, and then transform the tree to a reverse Polish phrase as a function of the issue. The suggested approach implies a tree-based technique that can produce an infinite amount of criterion issues of differing degrees of complexity. Compared to current approaches, the suggested testing specimen selection approach will include a wide variety of information for the testing of a high-performance recommendation model. This thesis embraces a deep recurrent neural network as a recommendation model that is far more efficient than that used in existing approaches.

The suggested approach employs the tree classification elements as input properties of the classification model for solving optimization problems. Especially, it translates the tree to the reverse Polish phrase, where the tree signifies the origin of the tree, tree: value denotes the representation of the node, and tree: left and tree: right represents the left and right node, accordingly. It should be remembered that the feature of the concern can also be expressed as an in-order phrase. In particular, the inverse Polish expression is a variable-length recurrent neural network that is used as a classifier.

Various characteristics are then determined on the basis of such quantitative values. It is assumed that the efficiency of the implementation is entirely dependent on the environments of the Problems, and these landscape-related features will widen the gap among method efficiency and challenge complexity. This theory is clear and rational, and it struggles from the computational complexity. Because ad approaches can be evaluated in the principle of fractions for which of the objective functions.

Detailed Description of the Invention:

Figure 1: Aircraft design framework Figure 2: Structure of Deep RNN Figure 3: A generic framework for algorithm recommendation

Figure 4: The design of recommendation system for aero dynamic design of structures

Detailed Description of the Invention:

41Page

Figure 1 shows the overall framework of designing the aircraft design. Figure 2 explains the concept of Deep learning or hierarchical learning is a segment of machine learning that primarily applies commonly adopted neural network principles. It represents a deep, recurrent neural network with hidden layers of L. Each secret phase is transferred periodically to both the next time interval of the current layer and the current time stage of the following level.

Figure 3 illustrates the Recommendation Algorithm that may be called a classification process, where each specimen comprises of attributes reflecting the complexity or layout of the optimization method and the subsequent mark is the dataset of the best fit optimization technique for the challenge. Evidently, the extraction of successful design elements for expressing the properties of the evolutionary algorithms and the development of a high-performance classification model for training the connection among the functionalities and the mark are two essential parts for the effectiveness of the recommendation process. Owing to the sophistication of the attributes in combinatorial optimization problems, most of the current approaches perceive these to be black box difficulties and obtain aspects that reflect environment properties. Particularly, a range of decision variables are examined in the decisions field of the Latin Hypercube sampling problem and its objective parameters are compared.

Figure 4 demonstrates the overall structure of the suggested approach and then a large number of benchmark issues are randomly generated by learning results. The selection technique and their marks are acquired by checking a variety of metaheuristics on them. Notice that the quantities of coefficients are assigned to random variables within [1; 10] and the distribution of random numbers varies within [1; 2]. These troubles are then employed to develop a deep recurrent neural network. In the case of a white-box continuous optimization technique, the tree structure is interpreted and then placed into the neural network. In the case of a black-box continuous optimization problem, the tree is calculated by a symbolic regressor and afterwards loaded into a neural network-based classifier.

51Page

Claims (6)

A RECOMMENDATION MODEL FOR AERO DYNAMIC DESIGN OF STRUCTURES USING DEEP RECURRENT NEURAL NETWORK CLAIMS

1. A computer-assisted approach for optimizing the configuration of a body surface that passes via a flow area, a process consisting of:

Formulation of a model of the computational fluid dynamics (CFD) of the said body requiring at least one CFD estimation;

Review of the measurement of the CFD for at least one defined state of the fluid domain and the achievement of at least one fluid flow on the said substrate of the body;

Assessment of the typical curvature of the surface relative to the position of the structure

2. As with claim 1, Computer-assisted approach for the recognition of system parameters for a predefined series of calculations comprising aerodynamic aircraft structures, including the detection of model parameters in the following steps. Archiving of aerodynamic data source pertaining to aircraft, Choosing a system for neural networking; Separate the identity of the model parameters into a multitude of limited identifiers by splitting linear and non-linear portions of the pre - defined set of coefficients representing aerodynamic models;

3. Following claim 2, claim 3 comprises of Choosing a mode to represent non-linear parts of the preset parameters from a collation or presentation as a neural network Consecutively performing partial identifiers, Calculate the difference among the quantities of the metric and the aerodynamic standard results, and verify the values of the parameters for which the deviation is less than the predetermined limit.

4. The method according to any of claim 2 and 3 comprising Estimation of aerodynamic flow variable according to the direction of flow line;

1 Pag e

Simultaneous visualization flow variable is a pressure gradient. normal curvature and aerodynamic flow variable are visualized graphically on the surface of the body.

5. As with claim 1 and 2, computer-implemented framework for optimizing the configuration of a body surface that passes through a flow area, a computer-implemented system includes: CAD unit for the representation of the said surface and the corresponding mathematical or functional characteristics; 3D computational fluid dynamics (CFD) module deployed on the device to acquire at least one flow line on the said surface for at least one defined fluid flow condition; And the estimation module is employed for evaluation process.

6. Deep recurrent neural network model is utilized for training the model for a classification task.

2|Page

A RECOMMENDATION MODEL FOR AERO DYNAMIC DESIGN OF STRUCTURES USING DEEP RECURRENT NEURAL NETWORK

Drawings 2020102874

Figure 1: Aircraft design framework

Figure 2: Structure of Deep RNN

1|Page

Figure 3: A generic framework for algorithm recommendation

Figure 4: The design of recommendation system for aero dynamic design of structures

2|Page