CN111985150A - Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization - Google Patents
Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization Download PDFInfo
- Publication number
- CN111985150A CN111985150A CN202010639297.1A CN202010639297A CN111985150A CN 111985150 A CN111985150 A CN 111985150A CN 202010639297 A CN202010639297 A CN 202010639297A CN 111985150 A CN111985150 A CN 111985150A
- Authority
- CN
- China
- Prior art keywords
- electronic device
- design
- optimization
- input tolerance
- machine learning
- 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
- 238000005457 optimization Methods 0.000 title claims abstract description 112
- 238000000034 method Methods 0.000 title claims abstract description 57
- 238000010801 machine learning Methods 0.000 title claims abstract description 48
- 230000004044 response Effects 0.000 claims abstract description 24
- 238000012938 design process Methods 0.000 claims abstract description 4
- 238000004422 calculation algorithm Methods 0.000 claims description 48
- 238000005070 sampling Methods 0.000 claims description 26
- 238000004088 simulation Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 238000012952 Resampling Methods 0.000 claims description 4
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000003491 array Methods 0.000 abstract description 2
- 239000003795 chemical substances by application Substances 0.000 description 24
- 238000010586 diagram Methods 0.000 description 11
- 238000004364 calculation method Methods 0.000 description 7
- 239000000758 substrate Substances 0.000 description 6
- 238000012795 verification Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 2
- 238000012804 iterative process Methods 0.000 description 2
- 230000005855 radiation Effects 0.000 description 2
- 238000010845 search algorithm Methods 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000000342 Monte Carlo simulation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 238000010200 validation analysis 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]
- G06F30/20—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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses a multilayer electronic device robust optimization design method utilizing machine learning auxiliary optimization, which constructs a multilayer electronic device robust optimization design framework and introduces a machine learning auxiliary optimization method into each layer, thereby utilizing the existing information to predict information comprising electronic device performance response, worst performance under specific input tolerance, maximum input tolerance over capacity under specific output tolerance and the like, and greatly accelerating the robust optimization design process of the electronic device. The method can be used in the fields of robust optimization design of electronic devices including antennas, arrays, active and passive microwave millimeter wave devices and the like.
Description
Technical Field
The invention relates to a Multi-layer Electronic Device Robust Design Method (Multi-Stage Electronic Device Robust Design Method) utilizing Machine Learning Assisted Optimization (MLAO), which can be used in the fields of Electronic Device robustness Optimization Design and the like.
Background
In the past decade, the machine learning method has been widely introduced in the field of designing electronic devices such as antennas, passive devices, and circuit designs, and has achieved excellent results. The electronic device optimization method utilizing machine learning mostly adopts a mode of simplifying processes such as full-wave simulation with high calculation cost and the like into a proxy model with low calculation cost, and the problem of overlong optimization time caused by high simulation cost of full-wave simulation and multiple operation times of a metaheuristic algorithm is solved.
However, compared to the traditional electronic device optimization design problem, the robust optimization design problem of the electronic device is more complex: designers desire not only the desired target response of the electronic device, but also robustness with certain input tolerances. The most reliable solution is to find the design point that yields the volume of the largest input tolerance hypercube, given the output tolerances, and treat it as the most robust design point. However, this method usually requires a lot of simulations and optimization based on different design point parameter combinations, which brings far more burden on computing resources than the traditional electronic device optimization design, thereby limiting its application in electronic devices and other optimization fields requiring a lot of computing resources. An inexpensive proxy model of electronic device performance is established only by introducing a machine learning method, and the robustness optimization design of the electronic device still cannot be obtained within an acceptable time range.
Therefore, how to accelerate the robust optimization design of electronic devices is the key to solve the problem.
Disclosure of Invention
The purpose of the invention is as follows: aiming at the problems, the invention provides a multilayer electronic device robustness optimization design method using machine learning assisted optimization to accelerate the robustness optimization design of the electronic device.
The technical scheme is as follows: in order to achieve the above object, the robust optimization design method for a multilayer electronic device using machine learning assisted optimization according to the present invention includes the following steps:
constructing a multilayer robust optimization architecture, which comprises a first proxy model between electronic device design parameters and electronic device Performance response, a second proxy model between the electronic device design parameters and Input Tolerance and Worst Performance (WCP) of the electronic device, and a third proxy model between the electronic device design parameters and Maximum Input Tolerance hyper capacity (MITH) corresponding to the electronic device design parameters, wherein the first proxy model is established by using a machine learning method; the data set learned by the second agent model is obtained by searching the worst performance of the electronic device in the Input Tolerance Region (ITR) obtained by sampling by using the first agent model; aiming at a given design point, the data set learned by the third agent model is obtained by searching the maximum input excess capacity meeting the Output Tolerance Range (OTR) by using the second agent model;
the three agent models are utilized to predict, optimize and verify the performance response and robustness performance of the electronic device, and the models are updated through an iteration method, so that the robust optimization design process of the electronic device is accelerated.
Preferably, a first agent model between design parameters of the electronic device and performance response of the electronic device is established by a machine learning method based on a data set consisting of performance of the electronic device calculated by a sample point in a sampled design range and a full-wave simulation method, and the worst performance of the electronic device is searched in a range of values of different input tolerances obtained by sampling by combining the first agent model with an optimization algorithm, so that the worst performance corresponding to different input tolerance ranges is obtained by optimization, and the data set consisting of the design parameters of the electronic device and the input tolerances and the worst performance corresponding to the input tolerances is established.
Preferably, based on a data set composed of the established electronic device design parameters and input tolerance and the worst performance corresponding to the electronic device design parameters and the input tolerance, a second proxy model between the electronic device design parameters and the input tolerance and the worst performance of the electronic device is established by using a machine learning method, and the maximum input tolerance excess capacity corresponding to different electronic device design parameters is obtained by optimizing the second proxy model by using an optimization algorithm under the condition of the given output tolerance, so that the data set composed of the electronic device design parameters and the maximum input tolerance excess capacity corresponding to the electronic device design parameters is established.
Preferably, based on a data set composed of the established electronic device design parameters and the maximum input tolerance excess capacity corresponding to the electronic device design parameters, a third theoretic model between the electronic device design parameters and the maximum input tolerance excess capacity corresponding to the electronic device design parameters is established by a machine learning method, the third theoretic model is combined with an optimization algorithm to optimize the electronic device design parameters corresponding to the maximum input tolerance excess capacity under the condition of given output tolerance, and the electronic device design parameters are regarded as the most robust electronic device design points obtained through prediction.
Preferably, an iteration mode is used, accurate response of the performance of the electronic device is obtained by a full-wave simulation method continuously at the design point obtained by predicting the design target point and randomly sampling the design target point in the neighborhood range, and therefore the three agent models are continuously updated by a machine learning method in the algorithm process.
Preferably, the method for searching for the maximum input tolerance excess capacity using the established proxy model comprises the steps of:
(101) initializing and generating a first proxy model;
(102) sampling a plurality of input tolerance ranges within a value range of a design parameter of an electronic device;
(103) searching worst performance in a plurality of sampled input tolerance ranges, and establishing a data set formed by the design parameters of the electronic device, the input tolerance ranges and the corresponding worst performance;
(104) establishing or updating an existing second proxy model by using a machine learning method based on the obtained data set;
(105) optimizing the input tolerance range to maximize the input tolerance excess capacity;
(106) and (5) verifying whether the worst performance of the optimized input tolerance range corresponding to the maximum input tolerance excess capacity meeting the output tolerance range really meets the output tolerance range, putting the search result and the corresponding input tolerance excess capacity into the data set, outputting the maximum input tolerance excess capacity, the input tolerance range and the worst performance result if the worst performance meets and reaches the convergence condition, returning to the step (104) if the worst performance meets or does not reach the convergence condition, updating the second proxy model for the new data set, and continuing iteration.
Preferably, the step of designing the electronic device using the created proxy model includes:
(201) obtaining a first agent model and a second agent model;
(202) sampling a design point in a design space;
(203) calculating the maximum input tolerance excess capacity corresponding to the sampling design point, and establishing a data set formed by the design parameters of the electronic device and the maximum input tolerance excess capacity corresponding to the design parameters of the electronic device;
(204) establishing or updating a third mechanism model based on the obtained data set and a machine learning method; if the data set learned by the first agent model or the second agent model is updated, the first agent model or the second agent model is correspondingly updated;
(205) optimizing the performance, size or robustness characteristics of the designed electronic device by using the first proxy model and the third proxy model according to different optimization targets;
(206) resampling and verifying the response and the maximum input tolerance excess capacity near the optimization result so as to obtain accurate values of more sample points in the optimization design point and the neighborhood thereof, updating the data set, and improving the prediction accuracy of the proxy model near the predicted optimization value in a mode of updating the three proxy models in the subsequent iteration process; and if the termination condition is reached, outputting the response of the verified electronic device and the maximum input tolerance excess capacity, and if the termination condition is not reached, returning to the step (204), updating the first, second and third proxy models, and re-optimizing, wherein a new data set used for updating the third proxy model is obtained by utilizing the updated first and second proxy models in a maximum input tolerance excess capacity searching mode.
Has the advantages that: compared with the prior art, the invention has the following advantages: a machine learning method is introduced into a plurality of layers of robustness optimization design of the electronic device to establish a proxy model with low computational complexity for optimization, so that the design flow is greatly accelerated. Compared with the scheme of the proxy model which also utilizes the parameters of the electronic device and the performance response of the electronic device at present, the scheme of the invention is different in that the WCP is searched by utilizing the optimization algorithm, and the MITH is searched by utilizing the optimization algorithm on the basis, so that the obtained MITH is ensured to meet the OTR, namely the WCP is in the OTR, and the prior art scheme uses the MITH search based on the Monte Carlo idea, so that: 1) when the number of sampling points is not enough, the MITH may not be calculated; 2) because the WCP can not be obtained from the sampling point, the existing scheme can not ensure that the calculated MITH is the MITH meeting the OTR. The scheme of the invention searches WCP and MITH by using an optimization algorithm, ensures that the obtained MITH meets OTR, and is more reliable than the prior art; the increase of the calculation amount caused by the optimization algorithm is relieved by a technical method of generating an agent model with low calculation complexity by using a machine learning method; the precision of the proxy model used in multiple layers is improved by continuously expanding the data set in the iterative process of the algorithm and updating the proxy model.
Drawings
FIG. 1 is a diagram of a multi-layered robust optimization architecture according to an embodiment of the present invention.
Fig. 2 is a structural diagram of a first authentication example in the embodiment of the present invention.
Fig. 3 is a structural diagram of a second authentication example in the embodiment of the present invention.
Fig. 4 is a flowchart of an algorithm of the MITH search in the embodiment of the present invention.
FIG. 5 is a general algorithm flow diagram of the robustness optimization design of an embodiment of the present invention.
Fig. 6 is a diagram illustrating comparison between MITH obtained at different design points and results obtained by a conventional sampling (SS) method when MITH search is performed for the first verification example in the embodiment of the present invention.
Fig. 7 is a comparison diagram of the results of verifying the WCP of MITH calculated by the design method of the present invention and the conventional SS method at different design points with respect to the first verification example in the embodiment of the present invention.
FIG. 8 is a block diagram of a second example of authentication according to an embodiment of the present invention, passing through N0Resulting pareto frontier plots for the design target return loss, side lobes and MITH after 4 iterations.
FIG. 9 is a diagram of a second example of authentication, via N, in an embodiment of the present invention0Left and right of design points on the pareto frontier after 4 iterations11And comparing the predicted values of | and SLL with the verified values.
FIG. 10 is a graph comparing the computation time in the second validation example with the estimated consumption of computation time at each stage when no proxy model is used, in accordance with an embodiment of the present invention.
Detailed Description
The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.
The embodiment of the invention discloses a multilayer electronic device robust optimization design method by using machine learning auxiliary optimization, which constructs a multilayer electronic device robust optimization design framework and introduces a machine learning auxiliary optimization method into each layer, thereby predicting information including electronic device performance response, worst performance under specific input tolerance, maximum input tolerance excess capacity under specific output tolerance and the like by using the existing information and greatly accelerating the robust optimization design process of the electronic device. The established multilayer robust optimization architecture comprises a first proxy model between the design parameters of the electronic device and the performance response of the electronic device, a second proxy model between the design parameters of the electronic device, the input tolerance and the worst performance of the electronic device, and a third proxy model between the design parameters of the electronic device and the maximum input tolerance excess capacity corresponding to the design parameters of the electronic device, which are established by a machine learning method.
Fig. 1 details a specific structure of the multi-layer robust optimization architecture, taking the multi-layer antenna robust optimization design as an example.
Firstly, a first agent model between design parameters of the electronic device and performance response of the electronic device is learned by machine learning, and the WCPs corresponding to different ITRs are obtained by optimization by combining the agent model with an optimization algorithm, so that a data set consisting of the design parameters of the electronic device, the ITRs and the corresponding WCPs is established;
secondly, establishing a second proxy model between the design parameters of the electronic device and the ITR and the WCP of the electronic device by using a machine learning method based on the data set, and optimizing to obtain the MITH corresponding to different design parameters of the electronic device under the given OTR condition by using the proxy model and combining an optimization algorithm, thereby establishing a data set consisting of the design parameters of the electronic device and the MITH corresponding to the design parameters of the electronic device;
and finally, based on the data set, establishing a third proxy model between the electronic device design parameters and the corresponding MITH by using a machine learning method, optimizing to obtain the electronic device design parameters corresponding to the maximum MITH under the given OTR condition by using the proxy model and combining an optimization algorithm, and taking the electronic device design parameters as the most robust electronic device design points obtained through prediction.
Except that a certain number of sample points in a design range are obtained in an initialization stage by using a random sampling mode, the performance of the electronic device is calculated by using a full-wave simulation method, an iteration mode is used, accurate response of the performance of the electronic device is obtained by using the full-wave simulation method at a predicted design target point and the design point obtained by random sampling in a certain neighborhood range, and therefore the proxy model is continuously updated on line in the algorithm process.
Similarly, by using the obtained multiple proxy models, multi-objective optimization design can be performed on indexes including the performance, size, robustness and the like of the electronic device, namely, the multi-objective optimization obtains the pareto frontier including the performance, size, MITH and the like of the electronic device.
The practical applicability and the superiority of the proposed method are illustrated by two real antenna structure examples in the present embodiment. It will be appreciated by those skilled in the art that the method of the present invention is not limited to robust optimization of antennas, and can be used in the field of robust optimization design of electronic devices including antennas, arrays, active and passive microwave millimeter wave devices, and the like.
As shown in fig. 2, a schematic structural diagram of a conventional microstrip patch antenna unit is provided, where 1 is a feeding point, 2 is an antenna dielectric substrate with a thickness of 1mm, a dielectric constant of 2.2, and 3 is a metal patch printed on the surface of the dielectric substrate, and important characteristic parameters affecting the radiation performance of the antenna structure are shown in table 1, where l is a patch length, w is a patch width, and l is a patch widthfThe feed point is offset from the position of the patch center point. The optimization target of the antenna is the return loss | S at 10GHz11Robustness of l.
TABLE 1 microstrip patch antenna design parameters
| Parameter (mm) | Lower limit of | Upper limit of |
| |
0 | 20 |
| |
0 | 10 |
| |
0 | 20 |
As shown in the figureAnd 3, a structural schematic diagram of a conventional series-fed microstrip patch antenna array is provided, wherein 1 is a short circuit point, 2 is an antenna feed point, 3 is a series-fed microstrip patch antenna array printed on the surface of a dielectric substrate, the number of patches is 10, the structure is bilaterally symmetrical, 4 is an antenna dielectric substrate, the thickness of the antenna dielectric substrate is 1.5mm, the dielectric constant is 2.3, and important characteristic parameters influencing the radiation performance of the antenna structure are shown in table 2. l1To l5Representing the widths of the five patch antennas from left to right, respectively. Among other parameters, g017mm is the spacing between the patches, m02mm is the width of the microstrip line, w116.4mm is the length of the patch. The optimization target of the antenna is the return loss | S of the antenna at 5.8GHz11|, Side Lobe Level (SLL), and robustness.
Table 2 series feed microstrip patch antenna array design parameters
| Parameter (mm) | Lower limit of | Upper limit of |
| |
19 | 23 |
| |
18 | 22 |
| |
15 | 21 |
| l4 | 9 | 14 |
| |
6 | 12 |
Defining sets of sequence numbers to be used
Wherein P represents the number of design parameters and Q represents the number of performance responses; defining design parameters x and performance response y of an electronic deviceq(x) The relationship between is yq(x)＝yq(x1,x2,...,xP),
The uncertainty of the input of the electronic device design parameters, i.e., the input tolerance, can be defined as:
p≥0,
the input tolerance range ITR of an electronic device may be defined as
Wherein t is [ t ═ t1,...,tP]TPossible values for electronic device design parameters. For a given electronic device design point x, its robustness characteristics may be characterized by its ITH, also ITR ωx,Characterized by a hyper-volume, in the present invention, the value of ITH is defined as:
of the ITHThe larger the maximum value MITH, the more robust the design point x. The output tolerance of the electronic device performance is defined as
Δq≥0,
Whereby the output tolerance range OTR of the electronic device can be defined as
Wherein s ═ s1,...,sQ]TAs is the actual value possible for the performance of the electronic device.
FIG. 5 shows a general algorithm flowchart of robust optimization design in the design method of the present invention. Wherein the MITH search step is required in both steps S203 and S206. Fig. 4 shows a flowchart of the algorithm of the MITH search, wherein the WCP search step is required in both steps S103 and S106. The algorithms are all established on the basis of establishing a first agent model of electronic device parameters and performance response of the electronic device, in the invention, Gaussian Process Regression (GPR) is adopted as a machine learning method, and in the initialization and iteration processes, the first agent model is obtained by learning based on a data set formed by electronic device parameters obtained by sampling and electronic device performance obtained by utilizing a full-wave simulation solver HFSS and is used for WCP search, MITH search and final robustness optimization design. It should be noted that the first proxy model is continuously updated along with the expansion of the data set by resampling, verifying, and the like in the iterative process of the algorithm, so as to achieve the improvement of the accuracy.
On the basis of obtaining the first agent model, step S201 of the general algorithm flow of robust optimization design in fig. 5 further needs to obtain a second agent model, that is, an agent model between the electronic device design parameters and ITRs and the WCP of the electronic device. In the invention, the second proxy model is also established by using GPR as a machine learning method, and the establishing process of the data set is as follows: utilizing a first proxy model and optimization for given electronic device design parameters and ITRsThe algorithm is optimized by taking the worst electronic device performance in the ITR as an optimization target, the optimization result is taken as the WCP corresponding to the electronic device design parameters and the ITR, and the WCP corresponding to the value x of the electronic device design parameters is defined as
Wherein f isq(t) is yq(t) relative to yq(x) The amount of offset of (c). In the present invention, the Sampling method adopted is Latin Hypercube Sampling (LHS), and the optimization Algorithm adopted is Genetic Algorithm (GA). It should be noted that the second proxy model is also updated continuously during the iteration of the algorithm with the continuous update of the first proxy model to obtain higher prediction accuracy for the WCP.
The general algorithm flow diagram of the robust optimization design in fig. 5 includes the MITH search. As shown in FIG. 4, the MITH search may be expressed as maxT, given the design parameters of the electronic device and the OTRITH(x),s.t.F(x)∈Ωs,Δ， (2)
Wherein
The method comprises the following steps:
s101: initialization generates a first proxy model. The proxy model generation method here is the same as described above.
S102: sampling the ITR. Multiple ITRs are sampled over a range of values of design parameters of the electronic device, and in an example of the invention, 10 sample points are sampled using the LHS method.
S103: and calculating the WCP corresponding to the ITR. In the sampled ITR range, WCP search is carried out, namely, the GA and the first agent model are used for optimizing to obtain the worst performance of the electronic device in the ITR, so that a data set which is formed by the design parameters of the electronic device, the ITR and the corresponding WCP is established.
S104: and establishing or updating the second agent model. Using the data set obtained in step S103, an existing second proxy model, i.e., a proxy model between the electronic device design parameters and ITRs and the WCP of the electronic device, is established or updated using a machine learning method. In the present invention, GPR is adopted as the machine learning method used.
S105: ITR is optimized to maximize ITH. In the present invention, the ITR is optimized using GA to maximize ITH. The optimized cost function is:
wherein D (F (x), Ωs,Δ) Represents the difference between WCP and OTR, defined by the formula:
for the
Condition (f) ofMITHIs positive and decreases towards 0 as WCP approaches OTR, for F (x) e Ωs,Δ，fMITHBecomes negative and follows TITHIs increased and gradually decreased. By means of the constructed cost function, fMITHIs continuous, so that f can be more effectively controlledMITHAnd (6) optimizing.
S106: and verifying the ITR obtained by optimization. And for the optimized ITR corresponding to the MITH meeting the OTR, verifying whether the WCP of the ITR meets the OTR by utilizing WCP search, and putting the search result and the corresponding ITR into a data set. If the specified convergence condition is met or reached, outputting the results of MITH, ITR and WCP, and if the specified convergence condition is not met or reached, returning to the step S104, updating the second proxy model for the new data set, and continuing iteration. In the present invention, the specific convergence condition is set to have MITH at NuIf the number of times is not increased within 5 times, the optimization result is determined to be converged.
With the above MITH search algorithm flow, the overall algorithm of the robustness optimization design in fig. 5 has the following steps:
s201: the initialization obtains first and second proxy models.
S202: the design point was sampled. In the present invention, a LHS method is used to sample design points in a design space.
S203: the MITH of the design point was calculated. For the design points sampled in the previous step, their corresponding MITH is calculated using the MITH search algorithm described above. Thereby establishing a data set composed of the design parameters of the electronic devices and the corresponding MITH.
S204: and establishing or updating the first, second and third mechanism models. And establishing or updating a third theoretic model between the design parameters of the electronic device and the corresponding MITH by utilizing the data set and the machine learning method. Meanwhile, if the data set consisting of the design parameters and the performance of the electronic device and the data set consisting of the design parameters, the ITR and the WCP of the electronic device are updated, the first proxy model and the second proxy model are updated correspondingly. In the present invention, GPR is used here as a machine learning method.
S205: and (6) optimizing. With the first and third physics models, the designer can optimize the performance, size, robustness characteristics, etc. of the designed electronic device for different optimization objectives. In the example given by the patent of the invention, the robustness characteristic of the structure of the microstrip electronic device is subjected to single-target GA optimization; it is also possible to perform multi-objective optimization simultaneously to obtain the pareto frontier, including the different design objectives described above.
S206: re-sampling and verifying the response and MITH around the optimization results. Solving the optimization result obtained by using a single-target or multi-target optimization algorithm in a manner of resampling near the design point obtained by optimization and using a full-wave simulation solver to obtain the accurate values of more sample points in the optimization design point and the neighborhood thereof, updating the data set, and improving the prediction accuracy of the proxy model near the predicted optimization value in a manner of updating the first, second and third proxy models in the subsequent iteration process, wherein the new data set used for updating the third proxy model is obtained by using the updated data setThe first proxy model and the second proxy model are obtained in an MITH searching mode. And if the termination condition is reached, outputting the verified electronic device response and MITH, and if the termination condition is not reached, returning to the step S204, updating the first, second and third physical models, and optimizing again. In the present invention, the termination condition here may be set to the total number of iterations N0Or through N1The optimal MITH after the sub-iteration does not change.
By utilizing the algorithm flow, firstly, robust optimization design is carried out on the microstrip patch antenna shown in fig. 2, so as to verify the effectiveness and the advancement of the algorithm provided by the invention. Consider the OTR as | S11I is better than the-10 dB case. Fig. 6 and 7 are graphs showing comparison between the calculation of MITH and the verification of the calculation result thereof using the algorithm of the present invention and the conventional sampling (SS) algorithm. Wherein, the traditional SS algorithm is derived from J.A.Easum, J.Nagar, P.L.Werner, and D.H.Werner, effective multiple objective anti-optimization with tolerance analysis through the use of the sample of the substrate models, IEEE trans.antennas Propagg, vol.66, No.12, pp.6706-6715,2018, and the sampling point number of each iteration is NsThe number of iterations is 6 times recommended by its algorithm. It can be seen that the algorithm proposed by the present invention has two advantages:
(1) under the condition that the number of sampling points of the traditional SS algorithm is too small, the situation that the MITH cannot be obtained through calculation can occur, and the algorithm of the invention does not have the problems because of the adoption of the sampling and optimizing strategy.
(2) In the conventional SS algorithm, even if the number of sampling points is increased, the actual WCP may not satisfy the OTR because it is based on the monte carlo method in nature, whereas the present algorithm does not have such a problem because it utilizes the optimization algorithm.
Therefore, the algorithm provided by the invention has the characteristic of being more reliable than the traditional SS method when the MITH is calculated, and is more reliable than the traditional method in the subsequent robustness optimization process. Consider the OTR as | S11And if the I is better than-12 dB, directly performing single-target optimization on the MITH, and setting the total iteration number N 06 times. The resulting knotAs shown in table 3. Wherein the optimized antenna design parameters and input tolerance are in mm, and MITH is in mm3. As can be seen from table 3, the prediction of the MITH using the third theorem model learned by machine learning is quite accurate, and the reliability of the optimized maximum MITH can be reflected.
TABLE 3 | S for microstrip patch antenna using the algorithm of the present invention11Results of robustness optimization |
| Design parameters | l | lf | w | |
| Optimization results (antenna parameters) | 9.5687 | 2.3715 | 16.3756 | |
| Optimization results (input tolerance) | Δl | Δlf | Δw | MITH |
| Prediction value | 0.0444 | 0.2328 | 0.7706 | 0.0080 |
| Verified value | 0.0430 | 0.2413 | 0.7425 | 0.0077 |
And carrying out robustness optimization design on the series-fed microstrip patch antenna shown in the figure 3 by utilizing the algorithm flow. FIG. 8 shows a pass N0Resulting return loss | S for the design target after 4 iterations11|, Side Lobe (SLL), and robust pareto frontier. Wherein the unit of MITH is mm5. FIG. 9 shows MITH, | S in the above case11Predicted values and verified values for | and SLL. It can be seen that the proxy model learned by machine learning gives fairly accurate predictions of design goals. Table 4 shows the MITH, | S at 3 typical design points on the pareto frontier11Values of | and SLL, and parameters of design points and ITR, where | S11The units of | and SLL are dB, the units of antenna parameters and ITR are mm, and the units of MITH are mm5. It can be seen that although design points 2 and 3 are at | S11The overall performance of | and SLL is better than that of design point 1, but the MITH of the design point 1 is larger, so that the robustness is better. The optimized resulting pareto front may help designers to trade off between different design goals. FIG. 10 shows the comparison of the computation time in this example using the robust design method of the present invention with the pre-estimated computation time consumed when no proxy model is used at each level, and it can be seen that the robust design method of the present invention greatly saves the amount of antenna robust design required by the robust design by introducing the machine learning method at different levels of the algorithmTime of (d).
Table 4 typical design points on pareto frontier obtained by using the algorithm of the present invention to perform robustness optimization on a series fed microstrip patch antenna
The above description is only for the specific embodiments of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art can easily conceive of the changes or substitutions within the technical scope of the present invention, and all the changes or substitutions should be covered within the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the appended claims.
Claims (7)
1. A multilayer electronic device robustness optimization design method using machine learning assisted optimization is characterized by comprising the following steps:
constructing a multilayer robust optimization architecture, which comprises a first proxy model between the design parameters of the electronic device and the performance response of the electronic device, a second proxy model between the design parameters of the electronic device, the input tolerance and the worst performance of the electronic device, and a third proxy model between the design parameters of the electronic device and the maximum input tolerance excess capacity corresponding to the design parameters of the electronic device, wherein the first proxy model is established by using a machine learning method; the data set learned by the second agent model is obtained by searching the worst performance of the electronic device within the input tolerance range obtained by sampling by using the first agent model; aiming at a given design point, the data set learned by the third proxy model is obtained by searching the maximum input excess capacity meeting the output tolerance range by using the second proxy model;
the three agent models are utilized to predict, optimize and verify the performance response and robustness performance of the electronic device, and the models are updated through an iteration method, so that the robust optimization design process of the electronic device is accelerated.
2. The multilayer electronic device robust optimization design method using machine learning assisted optimization according to claim 1, wherein a data set composed of electronic device performance calculated based on sample points in a sampled design range and a full-wave simulation method is used, a first proxy model between electronic device design parameters and electronic device performance response is established using a machine learning method, and the worst performance of the electronic device is searched within the range of different sampled input tolerance values by using the first proxy model in combination with an optimization algorithm, so that the worst performance corresponding to different input tolerance ranges is optimized and obtained, and thus the data set composed of the electronic device design parameters and the input tolerance and the worst performance corresponding thereto is established.
3. The robust optimized design method for multilayer electronic devices using machine learning aided optimization according to claim 1, wherein based on the data set consisting of the established electronic device design parameters and input tolerance and worst performance corresponding thereto, a second proxy model between the electronic device design parameters and input tolerance and worst performance of the electronic device is established using a machine learning method, and the second proxy model is used in combination with an optimization algorithm to optimize to obtain the maximum input tolerance excess capacity corresponding to different electronic device design parameters under the condition of a given output tolerance, thereby establishing the data set consisting of the electronic device design parameters and the maximum input tolerance excess capacity corresponding thereto.
4. The multilayer electronic device robust optimization design method using machine learning aided optimization according to claim 3, wherein based on the data set composed of the established electronic device design parameters and the corresponding maximum input tolerance excess capacity, a third theoretic model between the electronic device design parameters and the corresponding maximum input tolerance excess capacity is established using a machine learning method, and the third theoretic model is combined with an optimization algorithm to optimize the electronic device design parameters corresponding to the maximum input tolerance excess capacity under the condition of the given output tolerance and is regarded as the most robust predicted electronic device design point.
5. The robust optimization design method for multi-layer electronic devices using machine learning aided optimization according to claim 1, wherein the precise response of the electronic device performance is obtained by using full-wave simulation method at the design point obtained by random sampling of the predicted design target point and its neighborhood range in an iterative manner, so that the three agent models are continuously updated by using machine learning method during the algorithm process.
6. The method of robust optimal design for multilayer electronics using machine learning assisted optimization of claim 1, wherein the method step of searching for maximum input tolerance excess capacity using the established proxy model comprises:
(101) initializing and generating a first proxy model;
(102) sampling a plurality of input tolerance ranges within a value range of a design parameter of an electronic device;
(103) searching worst performance in a plurality of sampled input tolerance ranges, and establishing a data set formed by the design parameters of the electronic device, the input tolerance ranges and the corresponding worst performance;
(104) establishing or updating an existing second proxy model by using a machine learning method based on the obtained data set;
(105) optimizing the input tolerance range to maximize the input tolerance excess capacity;
(106) and (5) verifying whether the worst performance of the optimized input tolerance range corresponding to the maximum input tolerance excess capacity meeting the output tolerance range really meets the output tolerance range, putting the search result and the corresponding input tolerance excess capacity into the data set, outputting the maximum input tolerance excess capacity, the input tolerance range and the worst performance result if the worst performance meets and reaches the convergence condition, returning to the step (104) if the worst performance meets or does not reach the convergence condition, updating the second proxy model for the new data set, and continuing iteration.
7. The method of robust optimal design of multi-layered electronic devices using machine learning assisted optimization of claim 1, wherein the step of designing the electronic device using the established proxy model comprises:
(201) obtaining a first agent model and a second agent model;
(202) sampling a design point in a design space;
(203) calculating the maximum input tolerance excess capacity corresponding to the sampling design point, and establishing a data set formed by the design parameters of the electronic device and the maximum input tolerance excess capacity corresponding to the design parameters of the electronic device;
(204) establishing or updating a third mechanism model based on the obtained data set and a machine learning method; if the data set learned by the first agent model or the second agent model is updated, the first agent model or the second agent model is correspondingly updated;
(205) optimizing the performance, size or robustness characteristics of the designed electronic device by using the first proxy model and the third proxy model according to different optimization targets;
(206) resampling and verifying the response and the maximum input tolerance excess capacity near the optimization result so as to obtain accurate values of more sample points in the optimization design point and the neighborhood thereof, updating the data set, and improving the prediction accuracy of the proxy model near the predicted optimization value in a mode of updating the three proxy models in the subsequent iteration process; and if the termination condition is reached, outputting the response of the verified electronic device and the maximum input tolerance excess capacity, and if the termination condition is not reached, returning to the step (204), updating the first, second and third proxy models, and re-optimizing, wherein a new data set used for updating the third proxy model is obtained by utilizing the updated first and second proxy models in a maximum input tolerance excess capacity searching mode.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010639297.1A CN111985150A (en) | 2020-07-06 | 2020-07-06 | Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010639297.1A CN111985150A (en) | 2020-07-06 | 2020-07-06 | Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111985150A true CN111985150A (en) | 2020-11-24 |
Family
ID=73438325
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010639297.1A Pending CN111985150A (en) | 2020-07-06 | 2020-07-06 | Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111985150A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113051836A (en) * | 2021-04-19 | 2021-06-29 | 东南大学 | Quick antenna modeling method under array environment assisted by machine learning |
| CN113609677A (en) * | 2021-08-05 | 2021-11-05 | 东南大学 | Multipath-based machine learning auxiliary antenna design method |
| WO2023115760A1 (en) * | 2021-12-20 | 2023-06-29 | 东南大学 | Motor multi-objective robustness optimization method based on local agent model |
| CN116542074A (en) * | 2023-06-13 | 2023-08-04 | 西安电子科技大学 | Robust design method for feed excitation of microwave energy transmission antenna system |
| US11775720B2 (en) | 2021-07-02 | 2023-10-03 | International Business Machines Corporation | Integrated circuit development using machine learning-based prediction of power, performance, and area |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110162847A (en) * | 2019-05-05 | 2019-08-23 | 东南大学 | Machine learning auxiliary antenna design method based on addition feature policy |
| CN110728034A (en) * | 2019-09-24 | 2020-01-24 | 东南大学 | Antenna rapid multi-target modeling method using multistage cooperative machine learning |
| US20200184131A1 (en) * | 2018-06-27 | 2020-06-11 | Dalian University Of Technology | A method for prediction of key performance parameter of an aero-engine transition state acceleration process based on space reconstruction |
-
2020
- 2020-07-06 CN CN202010639297.1A patent/CN111985150A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200184131A1 (en) * | 2018-06-27 | 2020-06-11 | Dalian University Of Technology | A method for prediction of key performance parameter of an aero-engine transition state acceleration process based on space reconstruction |
| CN110162847A (en) * | 2019-05-05 | 2019-08-23 | 东南大学 | Machine learning auxiliary antenna design method based on addition feature policy |
| CN110728034A (en) * | 2019-09-24 | 2020-01-24 | 东南大学 | Antenna rapid multi-target modeling method using multistage cooperative machine learning |
Non-Patent Citations (1)
| Title |
|---|
| 蒲保兴;杨路明;肖潇;: "基于进化计算的多目标鲁棒优化方法", 系统工程与电子技术, no. 05, 15 May 2009 (2009-05-15) * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113051836A (en) * | 2021-04-19 | 2021-06-29 | 东南大学 | Quick antenna modeling method under array environment assisted by machine learning |
| CN113051836B (en) * | 2021-04-19 | 2024-04-26 | 东南大学 | Rapid modeling method for antenna under machine learning-assisted array environment |
| US11775720B2 (en) | 2021-07-02 | 2023-10-03 | International Business Machines Corporation | Integrated circuit development using machine learning-based prediction of power, performance, and area |
| CN113609677A (en) * | 2021-08-05 | 2021-11-05 | 东南大学 | Multipath-based machine learning auxiliary antenna design method |
| CN113609677B (en) * | 2021-08-05 | 2024-01-23 | 东南大学 | Multipath-based machine learning auxiliary antenna design method |
| WO2023115760A1 (en) * | 2021-12-20 | 2023-06-29 | 东南大学 | Motor multi-objective robustness optimization method based on local agent model |
| CN116542074A (en) * | 2023-06-13 | 2023-08-04 | 西安电子科技大学 | Robust design method for feed excitation of microwave energy transmission antenna system |
| CN116542074B (en) * | 2023-06-13 | 2023-12-19 | 西安电子科技大学 | Robust design method for feed excitation of microwave energy transmission antenna system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111985150A (en) | Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization | |
| Pietrenko‐Dabrowska et al. | Computationally‐efficient design optimisation of antennas by accelerated gradient search with sensitivity and design change monitoring | |
| Koziel et al. | Efficient multi-objective simulation-driven antenna design using co-kriging | |
| CN110162847B (en) | Machine learning auxiliary antenna design method based on feature adding strategy | |
| CN106021813B (en) | A kind of fast multi-target antenna design method based on dynamic renewal agency model | |
| Wu et al. | Multilayer machine learning-assisted optimization-based robust design and its applications to antennas and array | |
| Qiu et al. | An adaptive robustness evolution algorithm with self-competition for scale-free internet of things | |
| CN109271695B (en) | Multi-target antenna design method based on neural network | |
| Dong et al. | Fast multi-objective optimization of multi-parameter antenna structures based on improved MOEA/D with surrogate-assisted model | |
| CN110728034B (en) | Antenna rapid multi-target modeling method using multistage cooperative machine learning | |
| Koziel et al. | Improved modeling of microwave structures using performance-driven fully-connected regression surrogate | |
| Pietrenko-Dabrowska et al. | Expedited antenna optimization with numerical derivatives and gradient change tracking | |
| Bose et al. | Design of an aperture-coupled microstrip antenna using a hybrid neural network | |
| Wei et al. | Fully automated design method based on reinforcement learning and surrogate modeling for antenna array decoupling | |
| Khan et al. | Modeling of microstrip antennas using neural networks techniques: a review | |
| CN116595867A (en) | Agent model aided antenna design optimization method based on variable space data constraint | |
| CN113609677B (en) | Multipath-based machine learning auxiliary antenna design method | |
| Merad et al. | Design and resonant frequency calculation of rectangular microstrip antennas | |
| Hammoodi et al. | Elliptical printed dipole antenna design using ann based on levenberg-marquardt algorithm | |
| Koziel et al. | Cost-efficient electromagnetic-simulation-driven antenna design using co-Kriging | |
| Jiang et al. | Improved algorithm for de‐interleaving radar signals with overlapping features in the dynamically varying electromagnetic environment | |
| MUSHAIB et al. | Designing of Microstrip Patch Antenna Using Artificial Neural Network: A Review | |
| Guney et al. | Adaptive‐network‐based fuzzy inference system models for input resistance computation of circular microstrip antennas | |
| Cao et al. | An Improved GMRES Method for Solving Electromagnetic Scattering Problems by MoM | |
| CN113051836B (en) | Rapid modeling method for antenna under machine learning-assisted array environment |
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 |