CN111723420A - Structural topology optimization method based on deep learning - Google Patents
Structural topology optimization method based on deep learning Download PDFInfo
- Publication number
- CN111723420A CN111723420A CN202010430938.2A CN202010430938A CN111723420A CN 111723420 A CN111723420 A CN 111723420A CN 202010430938 A CN202010430938 A CN 202010430938A CN 111723420 A CN111723420 A CN 111723420A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- topology optimization
- learning model
- layer
- decoding
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 63
- 238000005457 optimization Methods 0.000 title claims abstract description 39
- 238000013135 deep learning Methods 0.000 title claims abstract description 12
- 238000012549 training Methods 0.000 claims abstract description 29
- 238000013136 deep learning model Methods 0.000 claims abstract description 19
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000013461 design Methods 0.000 claims description 13
- 230000006870 function Effects 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 3
- MKXZASYAUGDDCJ-NJAFHUGGSA-N dextromethorphan Chemical compound C([C@@H]12)CCC[C@]11CCN(C)[C@H]2CC2=CC=C(OC)C=C21 MKXZASYAUGDDCJ-NJAFHUGGSA-N 0.000 claims description 3
- 238000006073 displacement reaction Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 2
- 230000000717 retained effect Effects 0.000 claims 1
- 238000004364 calculation method Methods 0.000 abstract description 10
- 238000004904 shortening Methods 0.000 abstract 1
- 238000007796 conventional method Methods 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 230000007547 defect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 235000001968 nicotinic acid Nutrition 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/18—Network design, e.g. design based on topological or interconnect aspects of utility systems, piping, heating ventilation air conditioning [HVAC] or cabling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Geometry (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Hardware Design (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Structural Engineering (AREA)
- Civil Engineering (AREA)
- Architecture (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Networks & Wireless Communication (AREA)
- Complex Calculations (AREA)
Abstract
The invention relates to a structural topology optimization method based on deep learning, which comprises the following steps: 1) generating training data; 2) preprocessing training data; 3) constructing a deep learning model for training; 4) and optimizing by adopting a trained deep learning model to obtain an output result, namely a topological optimization structure. Compared with the prior art, the method has the advantages of improving the calculation efficiency, greatly shortening the time, being good in robustness and generalization, being strong in practicability and the like.
Description
Technical Field
The invention relates to the field of structural optimization design of building bridges and the like, in particular to a structural topology optimization method based on deep learning.
Background
People are always pursuing novel design to improve the production efficiency, and the high-efficiency design method still has a wide prospect in the modern society. The design goal of structural designers has been to meet the requirements of service performance with minimal material (cost), and the more traditional approach is to make a number of initial designs based on experience and intuition, then calculate and analyze each solution, and decide the final solution based on the results, which has obvious limitations:
first, the quality of the initial solution depends heavily on the level of the designer;
second, the entire process takes a significant amount of time and labor cost, and the optimal solution may not exist in the set of initial solutions.
In view of the various defects of the traditional method, scholars at home and abroad try to find a method capable of automatically finding out an optimal design scheme based on a given design target, and with the improvement of computer performance and the deep research of optimization theory in the field of mathematics in recent years, the structure topology optimization technology is rapidly developed and is widely applied to the fields of civil engineering, automobiles, machinery and the like.
The topological optimization can find the design scheme which best meets the target based on the given load and boundary conditions under the condition of no initial scheme, the optimization result of the method can have any shape, size and topological form in the design domain, and therefore the method has high probability of finding the structural form with the highest structural material utilization rate, and in addition, the method does not need the initial design scheme and gets rid of the dependence on the level and experience of a designer.
Many topology optimization methods have been developed, which can be mainly classified into the following four categories: density-based methods, level-set methods, evolutionary methods, and intelligent algorithms. The density-based method is the most popular method in the field of topology optimization at present, has very good stability after rigorous mathematical verification, but has limitations, on one hand, the boundary of the final result optimized by the method often comprises an intermediate density unit, and the result needs to be post-processed to meet the requirement; on the other hand, the result obtained by the method is a local optimal solution, but cannot be guaranteed to be a global optimal solution, the level set method determines the parts of the structural material through the boundary or contour line of a higher one-dimensional implicit function, a clear interface can be created at the boundary of the structure, a fuzzy phenomenon cannot occur, but the complexity in mathematics is higher, so a large amount of calculation time is consumed in the optimization process, the evolution method is a structural optimization method based on a biological evolution theory, and some units with the worst force transfer efficiency are continuously deleted in the optimization process until the expected volume fraction is reached, however, the method cannot be recovered after some units are easily deleted by mistake in the initial optimization stage, and finally the method falls into the local optimal solution.
With the development of computer science, intelligent algorithms are applied to the field of topology optimization, most of the algorithms adopt the principle of bionics and have the capability of searching for a global optimal solution, but compared with other algorithms, the algorithms are too huge in calculation cost and cannot be used for solving the actual engineering problem, and in recent years, due to the vigorous development of machine learning algorithms and the calculation capability of computer chips (CPU and GPU), the machine learning algorithms are gradually used for solving the topology optimization problem. At present, the generalization capability of network models which spend a lot of time for training in the methods is relatively poor, the methods can only be suitable for specific initial conditions, and the calculation efficiency is a common problem in any topology optimization method, especially the optimization problem for large structures.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a structural topology optimization method based on deep learning.
The purpose of the invention can be realized by the following technical scheme:
a structure topology optimization method based on deep learning comprises the following steps:
1) generating training data;
2) preprocessing training data;
3) constructing a deep learning model for training;
4) and optimizing by adopting a trained deep learning model to obtain an output result, namely a topological optimization structure.
In the step 1), the training data comprises design domain size, boundary conditions, volume fraction, penalty factor, filter radius, load quantity and load direction.
In the step 2), the preprocessed input tensor serving as the deep learning model is a 6-channel input tensor.
The input tensor of the 6 channels is specifically:
displacement in the direction of node X, Y as the first and second paths of the input tensor;
positive strain of nodex、yAnd shear strain gammaxyA third, fourth and fifth channel as input tensors;
all numbers of the last channel of the input tensor are all the same and equal to the volume fraction.
In the step 3), the deep learning model adopts a network form of U-Net.
The deep learning model comprises:
and an encoding part: the method comprises three continuous coding blocks, wherein each coding block comprises 2 convolution layers, 2 batch regularization layers and 1 pooling layer;
a decoding part: the decoding device is composed of a plurality of decoding blocks, and each decoding block comprises a connecting layer, an anti-convolution layer, a batch regularization layer and a convolution layer.
In the coding part, the output of the last coding block enters the decoding part after being subjected to convolution of two additional convolution kernels with different sizes.
In the decoding part, the input of each decoding block consists of two parts: one part from the previous convolutional layer and the other part from the corresponding coding block.
In order to make the output of the network correspond to the two-dimensional structure, the last layer of the output of the network adopts a tensor of 1 channel 40 x 80, and the activation function of the last convolution adopts a Sigmoid function, so as to ensure the probability that the result corresponds to whether each unit is finally reserved.
The hyper-parameters used in the training process of the deep learning model are specifically as follows:
training a framework: keras;
l2 regularization coefficients: 1 e-5;
initial learning rate: 0.001;
learning rate attenuation coefficient: 0.1;
learning rate decay conditions: the loss of the verification set is not reduced in 10 continuous training rounds;
loss function: cross entropy;
and (3) an optimization algorithm: adam.
Compared with the prior art, the invention has the following advantages:
firstly, the invention carries out structural topology optimization by taking a machine learning network model as a core, thereby obviously improving the calculation efficiency, and the method can obtain basically the same result by only using 0.38 percent of time of the traditional method.
And secondly, because the input of the network is elaborately designed and has good robustness, the model trained based on the method provided by the patent has strong generalization capability, and the model trained by the boundary condition of the cantilever beam can better solve the problem of structural topology optimization corresponding to the boundary condition of the simply supported beam and the continuous beam.
And thirdly, a deep learning model network structure suitable for solving the structural topology optimization problem is provided, and the practicability is high.
Drawings
FIG. 1 is a flow chart of the method of the present invention.
Figure 2 is the model input tensor.
Fig. 3 is a network model structure.
FIG. 4 is a comparison graph of the calculation results of three examples, wherein FIG. 4a is the comparison between the conventional method and the method of the present invention in the conventional case, FIG. 4b is the comparison between the conventional method and the method of the present invention in the structural disconnection case, and FIG. 4c is the comparison between the conventional method and the method of the present invention in the better than the expected case.
Fig. 5 is a comparison between the conventional method and the method of the present invention under different boundary conditions, wherein fig. 5a is a comparison between the performances under the boundary conditions of the simple beam, and fig. 5b is a comparison between the performances under the boundary conditions of the continuous beam.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments.
Examples
As shown in FIG. 1, the invention provides a structure topology optimization method based on deep learning, which comprises the steps of firstly generating a large amount of training data through a traditional variable density method, then converting original data into training data according to a set rule, training a deep learning model according to the training data, and finally performing efficient structure topology optimization calculation by using a trained neural network model.
The key steps are described in detail as follows:
(1) generating training data
80000 training samples were generated by the traditional variable density method, and the rules for generating training data are as follows:
TABLE 1 Generation of a training data rules Table
(2) Data pre-processing
How to effectively transmit initial information (design domain size, load condition and boundary condition) to the deep learning model is the key of the method. By comparing the performances of the models corresponding to different input tensors on the test set, the optimal input of the model is finally determined to be a 6-channel input tensor, and the tensor can transfer all initial information to the deep learning model in a manner convenient for understanding. As shown in FIG. 2, the first two channels of the input tensor correspond to the displacement in the direction of the node X, Y, and the next 3 channels represent the positive strain of the nodex、yAnd shear strain gammaxyAll numbers of the last channel are all the same and equal to the volume fraction requirement.
(3) Deep learning model training
Aiming at the structural topology optimization problem, the network model structure with the best performance is shown in fig. 3, an integral framework structure of U-Net is adopted, the whole neural network generally adopts a network form of U-Net, and the network model structure is mainly divided into two parts:
1) and an encoding part: the input array is downsampled to return an array with a reduced dimension.
2) A decoding part: the input array is upsampled and returned to an array with an increasing dimension.
Considering that the input node information size is 41 × 81, and the output cell distribution information size is 40 × 80. Therefore, before the input tensor is formally transmitted to the encoding part, the input tensor needs to be adjusted to 40 × 80 through convolution, so that the connection of the U-Net is facilitated. The input tensor then passes through three consecutive coding blocks (Encoding blocks), each containing 2 convolutional layers, 2 bulk regularization layers, and 1 pooling layer. The output of the last encoded block enters the decoding section after undergoing convolution with two additional convolution kernels of different sizes.
Each Decoding Block (Decoding Block) includes a connection layer, a deconvolution layer, a batch regularization layer, and a convolution layer. Unlike the coding blocks, the input to each decoding block consists of two parts: one part is from the previous volume layer and the other part is from the corresponding coding block, which is also the biggest characteristic of the U-Net network. In order to make the output of the network correspond to the two-dimensional structure, the last layer of the network is output as a 1-channel 40 × 80 tensor, wherein the activation function of the last convolution adopts a Sigmoid function to ensure that the result corresponds to the probability whether each unit is finally reserved.
All of the hyper-parameters used in the model training process are shown in table 2.
TABLE 2 model training hyper-parameter table
The invention carries out structural topology optimization by taking a machine learning network model as a core, obviously improves the calculation efficiency, and compared with the traditional method, the calculation time is shown in a table 3, and the random selection settlement result is shown in a table 4, and the method provided by the invention can obtain basically the same result by only using 0.38% of the time of the traditional method.
TABLE 3 statistical table of computation time required for individual samples
Because the input of the network is designed and has good robustness, the model trained based on the method provided by the patent has strong generalization capability, and the model trained by the cantilever beam boundary condition can better solve the problem of structural topology optimization corresponding to the boundary conditions of the simply supported beam and the continuous beam, as shown in fig. 5.
Claims (10)
1. A structural topology optimization method based on deep learning is characterized by comprising the following steps:
1) generating training data;
2) preprocessing training data;
3) constructing a deep learning model for training;
4) and optimizing by adopting a trained deep learning model to obtain an output result, namely a topological optimization structure.
2. The method for structural topology optimization based on deep learning of claim 1, wherein in the step 1), the training data includes design domain size, boundary condition, volume fraction, penalty factor, filter radius, load amount and load direction.
3. The method for structural topology optimization based on deep learning of claim 1, wherein in the step 2), the preprocessed input tensor used as the deep learning model is a 6-channel input tensor.
4. The method as claimed in claim 3, wherein the 6-channel input tensor is specifically:
displacement in the direction of node X, Y as the first and second paths of the input tensor;
positive strain of nodex、yAnd shear strain gammaxyA third, fourth and fifth channel as input tensors;
all numbers of the last channel of the input tensor are all the same and equal to the volume fraction.
5. The method for structural topology optimization based on deep learning of claim 1, wherein in the step 3), the deep learning model is in a network form of U-Net.
6. The method according to claim 5, wherein the deep learning model comprises:
and an encoding part: the method comprises three continuous coding blocks, wherein each coding block comprises 2 convolution layers, 2 batch regularization layers and 1 pooling layer;
a decoding part: the decoding device is composed of a plurality of decoding blocks, and each decoding block comprises a connecting layer, an anti-convolution layer, a batch regularization layer and a convolution layer.
7. The method as claimed in claim 6, wherein the output of the last coding block in the coding part enters the decoding part after undergoing convolution of two additional convolution kernels with different sizes.
8. The method as claimed in claim 6, wherein the decoding part comprises two parts for each input of the decoding block: one part from the previous convolutional layer and the other part from the corresponding coding block.
9. The method as claimed in claim 6, wherein in order to make the output of the network correspond to the two-dimensional structure, the last layer of the output of the network uses a 1-channel 40 x 80 tensor, and the activation function of the last convolution uses Sigmoid function to ensure the probability that the result corresponds to whether each cell is finally retained.
10. The structural topology optimization method based on deep learning of claim 6, wherein the hyper-parameters used in the training process of the deep learning model are specifically:
training a framework: keras;
l2 regularization coefficients: 1 e-5;
initial learning rate: 0.001;
learning rate attenuation coefficient: 0.1;
learning rate decay conditions: the loss of the verification set is not reduced in 10 continuous training rounds;
loss function: cross entropy;
and (3) an optimization algorithm: adam.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010430938.2A CN111723420A (en) | 2020-05-20 | 2020-05-20 | Structural topology optimization method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010430938.2A CN111723420A (en) | 2020-05-20 | 2020-05-20 | Structural topology optimization method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111723420A true CN111723420A (en) | 2020-09-29 |
Family
ID=72564714
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010430938.2A Pending CN111723420A (en) | 2020-05-20 | 2020-05-20 | Structural topology optimization method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111723420A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113343580A (en) * | 2021-06-24 | 2021-09-03 | 大连理工大学 | Real-time topology optimization generation design method based on artificial intelligence technology |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140214370A1 (en) * | 2013-01-30 | 2014-07-31 | Honda Research Institute Europe Gmbh | Optimizing the design of physical structures/objects |
| CN108280305A (en) * | 2018-01-30 | 2018-07-13 | 西安交通大学 | Radiating element cooling duct rapid topology optimum design method based on deep learning |
| CN109919051A (en) * | 2019-02-22 | 2019-06-21 | 同济大学 | A kind of convolutional neural networks accelerated method for video image processing |
-
2020
- 2020-05-20 CN CN202010430938.2A patent/CN111723420A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140214370A1 (en) * | 2013-01-30 | 2014-07-31 | Honda Research Institute Europe Gmbh | Optimizing the design of physical structures/objects |
| CN108280305A (en) * | 2018-01-30 | 2018-07-13 | 西安交通大学 | Radiating element cooling duct rapid topology optimum design method based on deep learning |
| CN109919051A (en) * | 2019-02-22 | 2019-06-21 | 同济大学 | A kind of convolutional neural networks accelerated method for video image processing |
Non-Patent Citations (3)
| Title |
|---|
| YIQUAN ZHANG等: "A deep Convolutional Neural Network for topology optimization with strong generalization ability", 《ARXIV》 * |
| 王仁华等: "平面桁架结构拓扑优化方法研究", 《力学季刊》 * |
| 陈艾荣等: "结构拓扑优化理论及其在桥梁结构找型中的应用", 《同济大学学报(自然科学版)》 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113343580A (en) * | 2021-06-24 | 2021-09-03 | 大连理工大学 | Real-time topology optimization generation design method based on artificial intelligence technology |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111260211A (en) | Intelligent energy system evaluation method and device based on AHP-improved entropy weight method-TOPSIS | |
| CN110263873A (en) | A kind of power distribution network platform area classification method merging sparse noise reduction autoencoder network dimensionality reduction and cluster | |
| CN111209972A (en) | Image classification method and system based on hybrid connectivity deep convolution neural network | |
| CN109193756B (en) | Scene decoupling dynamic economic dispatching model solving method of wind power access system | |
| CN113343427B (en) | Structural topology configuration prediction method based on convolutional neural network | |
| CN117435145B (en) | Digital building information optimized storage method and system | |
| CN111985152B (en) | Event classification method based on dichotomy hypersphere prototype network | |
| CN106775705A (en) | A kind of software module division methods | |
| CN114880538A (en) | Attribute graph community detection method based on self-supervision | |
| CN111723420A (en) | Structural topology optimization method based on deep learning | |
| Li | Computer network connection enhancement optimization algorithm based on convolutional neural network | |
| CN109982095B (en) | CNN and GEP-based fractal image compression coding method | |
| CN114494284B (en) | Scene analysis model and method based on explicit supervision area relation | |
| CN117093885A (en) | Federal learning multi-objective optimization method integrating hierarchical clustering and particle swarm | |
| CN116303386A (en) | Intelligent interpolation method and system for missing data based on relational graph | |
| CN101882106B (en) | Dynamic optimization method of evolutionary testing based on catastrophe | |
| CN114332491A (en) | Saliency target detection algorithm based on feature reconstruction | |
| CN115186882A (en) | Clustering-based controllable load spatial density prediction method | |
| CN113807497A (en) | Non-paired image translation method for enhancing texture details | |
| CN112990336A (en) | Depth three-dimensional point cloud classification network construction method based on competitive attention fusion | |
| CN111860776A (en) | Lightweight time convolution network oriented to time sequence data rapid prediction | |
| CN117743870B (en) | Water conservancy data management system based on big data | |
| CN117035837B (en) | Method for predicting electricity purchasing demand of power consumer and customizing retail contract | |
| Jia et al. | An improved pointwise convolutional block for efficient model compression | |
| CN113343580A (en) | Real-time topology optimization generation design method based on artificial intelligence technology |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200929 |