A method of ai-assisted design optimization
The AI-assisted design optimization method using RDE and a surrogate model addresses the inefficiencies of traditional design optimization by auto-tuning hyperparameters and performing comprehensive searches, resulting in faster and more accurate optimal design parameter identification.
Patent Information
- Application Number
- PCT/MY2024/050041
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-25
- Filing Date
- 2024-05-07
- Publication Date
- 2024-11-28
AI Technical Summary
Traditional engineering design optimization is time-consuming, prone to human error, and inefficient due to the vast number of parameters involved, limiting the ability to explore all potential solutions for optimal design.
A method of AI-assisted design optimization using reinforced differential evolution (RDE) with a trained surrogate model, which auto-tunes hyperparameters and performs both local and global searches to find optimal design parameters efficiently.
This approach significantly reduces the time and effort required for design optimization, achieving faster and more accurate results by leveraging AI to explore a broader solution space and converge on optimal design parameters.
Smart Images

Figure MY2024050041_28112024_PF_FP_ABST
Abstract
Description
[0001] A METHOD OF AI-ASSISTED DESIGN OPTIMIZATION
[0002] 1. TECHNICAL FIELD OF THE INVENTION
[0003] The present invention relates to a method of artificial intelligence (Al)-assisted design optimisation, used to optimise engineering design problems, whereby said method is built on top of at least one any suitable artificial intelligence (Al) algorithm model 1505 such as reinforcement learning (RL) model, supervised learning (SL), polynomial approximation and / or differential equations approximation consolidated with statistical approaches. After at least one design's goal is predefined, the method of the present invention finds non-linear relationships between the design structures and the design's goal(s), and controls how a set of design's parameters is generated. These sets of design's parameters are evaluated and the current state of gap with the design's goal is updated back to said to the optimizer. Iteratively, the desired design's parameters are approximated. In order to speed up the optimization process, especially for complex 3D designs, the evaluation process time per iteration is addressed. In simulating the 3D design's response in order to evaluate the current state of gap with the design's goal, the 3D design's response is approximated by an Al approximation function, instead of computed in Finite Element Analysis (FEA) which consumes more time for more complex 3D designs. 2. BACKGROUND OF THE INVENTION
[0004] Engineering design optimization is part of the engineering design methodology, whereby said engineering design optimization is generally referred to as a process of performing at least one iteration of mathematical formulations or algorithms in finding the best design parameters or variables in the design in a given constraint, that can meet the requirements, objectives or goals of a project.
[0005] Traditionally, engineering design optimization was generally being done manually, which is time consuming and requires hands-on experience, prone to human error and inefficient. This is due to the vast number of parameters or variables involved, which makes the optimization possibilities to be too many to be tried one- by-one. Another problem of manual engineering design optimization is the inability to explore the full potential possibilities of solution fast enough to come up with the optimum design that meets the objectives or goals of the design.
[0006] Albert Ludwigs et al, DE202020104238U1, disclosed a device for learning and operating a strategy that is set up for parametrizing an evolutionary algorithm. However, said prior art does not include usage of surrogate model and predictive model.
[0007] Adriaenssen Steven et al, US20220027743A1, disclosed a method and device for learning a strategy and operating the strategy, which also does not include usage of surrogate model and predictive model. Hence, it would be advantageous to alleviate the shortcomings by having a method of Al-assisted design optimization which is self-adaptive, and able to autotune the optimizer7s hyperparameters based on different design's structure.
[0008] 3. SUMMARY OF THE INVENTION
[0009] Accordingly, it is the primary aim of the present invention to provide a method of Al-assisted design optimization which is self-adaptive and auto-tunes said optimizer's hyperparameters based on different design's structure.
[0010] It is yet another objective of the present invention to provide a method of Al- assisted design optimization which is fast and accurate.
[0011] It is yet another objective of the present invention to provide a method of Al- assisted design optimization which can perform generic design's structure optimization to both local search and global search for optimal solutions, whereby local search is done to search optimal solutions locally in the nearest minima or maxima while global search is done to search optimal solution globally in the whole minima or maxima. Optimal solution is defined as a set of design's parameters that will generate a design output response that respects the predefined design's goal (highest fitness).
[0012] It is yet another objective of the present invention to provide a method of Al- assisted design optimization which is generic and global. Additional objects of the invention will become apparent with an understanding of the following detailed description of the invention or upon employment of the invention in actual practice.
[0013] According to the preferred embodiment of the present invention the following is provided:
[0014] A method of artificial intelligence (Al)-assisted design optimization, comprising the following steps: i. obtaining at least one design; ii. defining at least one design goal; iii. training at least one surrogate model; iv. optimizing parameters of said design with reinforced differential evolution (RDE) and said trained surrogate model to achieve said design goal.
[0015] 4. BRIEF DESCRIPTION OF THE DRAWINGS Other aspect of the present invention and their advantages will be discerned after studying the Detailed Description in conjunction with the accompanying drawings in which: Fig. 1 is a flowchart showing the method of Al-assisted design optimization of the present invention.
[0016] Fig. 2A is a 3D design structure while Fig. 2B is a 2D design in schematic diagram.
[0017] Fig. 2C is a graph showing the design goal versus the original design output responses and optimised design output responses.
[0018] Fig. 3 is a flowchart showing the method of training surrogate model.
[0019] Fig. 4 shows an example of the samples of the multiple sets of design parameters.
[0020] Fig. 5 shows an example of how the multiple sets of design parameters are simulated to product design's output response.
[0021] Fig. 6 is a table showing an example of the inputs and outputs of the dataset.
[0022] Fig. 7 is a flowchart showing the method of perform RDE optimization.
[0023] Fig. 8 is a flowchart showing the method of adjusting the population.
[0024] Fig. 9A shows an example of how the target and candidates are chosen.
[0025] Fig. 9B shows an example of how the candidates undergo the mutation process.
[0026] Fig. 9C shows an example of how the trial is produced.
[0027] Fig. 10 shows an example of how to obtain the respected design output responses in numerical form from the trial and target.
[0028] Fig. 11 shows an example of how the target is replaced by the trial.
[0029] Fig. 12 shows an example of how the next target is being processed. Fig. 13 is a flowchart showing the method of selecting DE algorithm.
[0030] Fig. 14 is a block diagram showing the surrogate model.
[0031] Fig. 15 is a block diagram showing the RDE process, whereby the RL model is integrated into the DE algorithm.
[0032] Fig. 16 is a block diagram showing the RL model.
[0033] Fig. 17 is a block diagram showing the predictive model.
[0034] Fig. 18A shows an example of the main interface of the graphics user interface (GUI) of the implementation of the method of Al-assisted design optimization of the present invention.
[0035] Fig. 18B shows an example of the GUI of the implementation of the method of Al-assisted design optimization of the present invention, during optimization process.
[0036] Fig. 19 is a bar chart showing the comparison of cost functions between optimizing done using the method of the present invention, Generic Algorithm (GA) and Particle Swarm Optimization (PSO).
[0037] Fig. 20 A is a table showing comparison between various 3D design simulation's output responses, whereby Fig. 20B is a graph of the best response using GA, Fig. 20C is a graph of the best response using Multi-Objectives Genetic Algorithm (MOGA), Fig. 20D is a graph of the best response using Adaptive Multiple Objectives (AMO) and Fig. 20E is a graph of the best response using the method of the present invention. 5. DETAILED DESCRIPTION OF THE DRAWINGS
[0038] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be understood by the person having ordinary skill in the art that the invention may be practised without these specific details. In other instances, well known methods, procedures and / or components have not been described in detail so as not to obscure the invention.
[0039] The invention will be more clearly understood from the following description of the embodiments thereof, given by way of example only with reference to the accompanying drawings, which are not drawn to scale.
[0040] The invention presents a method of artificial intelligence (Al)-assisted design optimization 101, as shown in Fig. 1. Step (i) of the method of the present invention is obtaining 103 at least one design 201, whereby the design 201 can be circuitry (as shown in Fig. 2B), 2-dimentional (2D) structures or 3-dimentional (3D) structures (as shown in Fig. 2A) in at least one computer-aided design (CAD) software that comprises of subcategories not limited to computer-aided engineering (CAE), 2D CAD, 3D CAD, or printed circuit board (PCB) design. The design's 201 elements should preferably be predefined to cover the whole design structure so that the design's output response can be successfully generated without errors. However, it is not compulsory to cover the whole design structure, so long as the design is valid (can be successfully simulated) or the design's output response can be captured. The design's output response, for example, can be S-parameters response in 2D / 3D
[0041] Microwave filter structures, or blade's temperature of a 3D turbine blade structure, or cost of a structure, or electrical performance of a PCB design, in the form of numerical data. In order to examine whether the design structure has met the design's goal(s) or requirement(s), the design's response is evaluated using the method of the present invention. The design's response generally changes with the change of the design's parameters 203. The design's parameters 203, for example, can be width and length (dimensions) of the 3D design's structures, or capacitance and inductance of a 2D circuit design, or the arrangement of the PCB components. Thus, the objective of the method of the present invention is to find the exact value of the design's parameters 203 such that the output design's response can satisfy the design's goal(s), by continuously adjusting the value of the design parameters 203, said process is referred to as optimisation. In an example of step (i) of the method of the present invention, a complete design 201 is utilized and the set of the design parameters 203 to be optimised is identified by the user.
[0042] Step (ii) of the present invention is defining 105 at least one design goal 205 or at least one set of design goal. The design goal(s) 205 (one goal or multiple goals) must be defined by the user. The design goal(s) 205, for example, can be the y-axis value of a certain range in the x-axis in the x-y graph with the conditions (bigger than, smaller than, or equal to), or the dimensions of a 3D structures, or the cost of the design structure after being manufactured, or the locations of the PCB components placed on the PCB board. The design 201 prepared may or may not have already satisfied or achieved the design's goal(s) 205. The achievements of the design's goal(s) 205 are the requirements that determines whether the design 201 is good or bad. Nevertheless, the design 201 can still be running through the design optimisation process. By evaluating the design's output response 207 using a mathematical analysis in the automatic design optimization method of the present invention, said method can identify if the design's output response 207 satisfies, not satisfies, very close to, or very far away from the design's goal(s) 205 in numerical form. These design goal(s) or set of design goal(s) 205 are used to evaluate the design 201 in said method of the present invention. An example of an illustration of the comparison between the design's output response 207 and the design goal 205 can be shown in Fig. 2C, whereby the dotted lines are the goals 205, whereby in this example, from 3.416GHz to 3.786GHz, the value of design's output response must be smaller than -17dB. The dashed curves are the original design responses before optimization 207A while the solid curves are the optimised design output responses 207B. These curves are called S-parameters.
[0043] Step (iii) is training 107 at least one surrogate model 1401. The objective of the surrogate model 1401 is to approximate the design's output response 207. For example, in microwave filter's 3D design, surrogate model 1401 is used to approximate the S-parameters in numerical form, or in the example of a turbine's blade design, surrogate model 1401 is used to approximate the blade's temperature in graph form. Surrogate model 1401 can also be used to approximate the cost of a structure design, or to approximate the dimensions of a PCB layout. The function approximator algorithms used in this surrogate model 1401 are not limited to Bayesian network, neural network, decision tree, or other function approximator algorithms (either an Al machine learning algorithm or otherwise). As shown in Fig. 3, step (iii) of training 107 surrogate model 1401 comprises of the following sub-steps. Step (iii)(a) samples 301 multiple sets of design parameters 203. The surrogate model 1401 requires surrogate model input in numerical form to produce output of an approximated design's response 207. The surrogate model input in numerical form in the method of the present invention is the set of design's 201 design parameters 203 that are to be optimised. It is preferable to have more than five design parameters 203. The sampling of multiple sets of the design parameters 203 can be done using at least one sampling technique not limited to Latin Hypercube, Quasi-Monte Carlo (QMC) with Sobol / Halton sequence, simple random sampling, Monte-Carlo sampling, Orthogonal array sampling or combination thereof. These multiple sets of the design parameters can be referred to as "population" in DE's terminology. "Sample" is used here to stress that the design parameters 203 are obtained with a certain sequence using said sampling techniques or mathematical approaches. An example of said sampling process can be illustrated in Fig. 4. The dots are the samples 401, wherein each dot represents a set of design parameters 203. The set of design parameters (one dot) can be pictured as: [ pl, p2, ... pN ] where N is the number of total design parameters 203 in a sample (one dot). The "N" can be 6, 10, 15 or any other suitable amount of design parameters 203. "pl" can be width with "p2" being the length. These dots (multiple sets of the design parameters 203) are being sampled in a latent space 403, as shown in Fig. 4, whereby it can be illustrated as a sphere in Fig. 4. Latent space refers to an abstract space to visualise high-dimensional population in the 2D image of Fig. 4. Step (iii)(b) involves performing simulation 303 on said design 201 based on said sampled sets of design parameters 203 using at least one simulation software, before collecting multiple sets of simulated design's output responses 207 for each set of design parameter 203. As shown in Fig. 5, for each set of design parameters 203 in the multiple sampled sets being updated to said simulation model, the respected design's output response 207 is collected from said simulation model. This process is automated without needing the user to key-in the said design parameters 203 and run the simulation to obtain the design's output response 207.
[0044] Step (iii)(c) is constructing 305 at least one data set 501 comprising said multiple sets of design parameters 203 and said multiple sets of simulated design's output responses 207, whereby said data set 501 can be constructed in at least one suitable form, such as numerical tabular data, Touchstone format, JavaScript Object Notation (JSON) format, or Image format. An example in numerical tabular data form is as shown in Fig. 6. The construction process is not limited to using any pre-processing techniques, including feature engineering such as feature reduction, selection / generation or normalisation and standardisation. As shown in Fig. 6, design parameters / input of the data set 501 can be constructed such that each row is for each set of the design parameters, whereas the output responses / output of the data set can be constructed such that the same row is constructed with the output responses
[0045] 207 of the respected design parameters 203. There can be a plurality of output responses because the output responses can be made up of multiple curves. Step (iii)(d) is fitting 307 said data set 501 into said surrogate model 1401. The data set from step (iii)(c) is fit into the defined surrogate model 1401, whereby said surrogate model is not limited to extreme gradient boosting (xgboost), neural network model, probabilistic model (Bayesian network, Gaussian process, Monte-Carlo Tree Search (MCTS), Markov Chain Monte Carlo (MCMC)), fuzzy logic model or decision tree model (gradient boost, random forest). The "f it" is to address that the surrogate model 1401 is finding the underlying pattern between the input and the output of said data set 501, thus building up or obtain the function that can be used to approximate the output response 207 from the design parameter 203 / input. Said function approximator is the surrogate model 1401, as shown in Fig. 14. The objective of training a surrogate model 1401 is to construct or obtain an accurate function in said surrogate model 1401 to generate the output responses 207 accurately with respect to said design parameter 203, to eventually replace the simulation. The main advantage of the surrogate model 1401 is it saves time to approximate the design's response 207 compared to conventional ways of simulating the design's response, which uses solely the existing 2D or 3D simulation software to simulate the design's output response.
[0046] Step (iii)(e) involves evaluating 309 said trained surrogate model 1401 to identify said surrogate model's performance by analysing said surrogate model's 1401 uncertainty using at least one mathematical analysis. The mathematical analysis is not limited to the standard deviation, variance, or the entropy of said surrogate model's output. This evaluation is the automated process. The surrogate model's 1401 actual output response 207 is generated by said surrogate model 1401 based on the set of design parameters 203 from the data set in step (iii)(c), and the surrogate model's (1401) actual output response 207 will be compared with the expected output response as in the data set 501. The uncertainty analysis is represented in the numerical form. Types of uncertainty analysis that can be done is cross-validation, Monte-Carlo methods, bootstrap, Bayesian inference, or polynomial chaos expansion.
[0047] Sub-steps (iii)(a) to (iii)(e) are repeated 311 if said uncertainty does not meet a predetermined level or termination criterion, within a margin of acceptance or tolerance. If the criterion is not satisfied, outside the margin of acceptance, or not within the tolerance, the new multiple sets of the design's parameters 203 are resampled and sub-steps (iii)(a) to (iii)(e) are repeated. The training process as in step (iii) is repeated until the surrogate model's 1401 the uncertainty criteria are satisfied.
[0048] As shown in Fig. 7, step (iv) is optimizing 109 parameters 203 of said design 201 with reinforced differential evolution (RDE) optimization algorithm 1501 and said trained surrogate model 1401 to achieve the intended design goal 205. Said RDE 1501 is using a global optimization algorithm, called differential evolution (DE) algorithm 1503, while the DE's hyperparameters are controlled by at least one any suitable Al algorithm model 1505 not limited to reinforcement learning (RL), supervised learning (SL), polynomial approximation, differential equations approximation or combination thereof. The DE algorithm's hyperparameters 607 that are controlled by Al including mutation rate (or scaling factor), crossover rate, selection strategy or combination thereof. In short, RDE 1501 will first sample 601 an initial multiple sets of design parameters 203. The RDE 1501 will iteratively (go through a series of the repeated steps) adjust the multiple sets of design's parameters 203, approximate the design's output responses 207 using the trained surrogate model 1401 of each multiple sets of adjusted design parameters 203, evaluate the multiple design's output responses 207 with the design's goals 205 and re-adjusting the multiple sets of design's parameters 203. Step (iv) is terminated if at least one termination criterion has been satisfied, whereby said termination criterion is not limited to a satisfactory of the design's goal 205, maximum iteration reached, or a convergence of multiple sets of design's parameters achieved. The satisfaction of the termination criterion is determined by at least one computation using at least one mathematical analysis.
[0049] As shown in Fig. 7, step (iv) of optimizing 109 design parameters 203 with the RDE algorithm 1501 and said trained surrogate model 1401 comprises of the following sub-steps. In sub-step (iv)(a), multiple sets of design parameters 203 are sampled 601. This process will sample multiple sets of the design parameters 203 using at least one sampling technique which are not limited to Latin Hypercube, QMC with Sobol / Halton sequence, or simple random sampling, Monte-Carlo sampling, orthogonal array sampling or combination thereof. "Sample" is used here to stress that the design parameters 203 are obtained with a certain sequence using said sampling techniques or mathematical approaches addressed above. These multiple sets of the design parameters can be referred to as "population" in DE's terminology.
[0050] An example of said sampling process can be illustrated in Fig. 4. The dots are the samples 401, wherein each dot represents a set of design parameters 203. All dots are classified as a population in said RDE algorithm 1501. The set of design parameters (one dot) can be pictured as: [ pl, p2, ... pN] where N is the number of total design parameters 203 in a sample (one dot). The "N" can be 6, 10, 15 or any other suitable amount of design parameters 203. "pl" can be width with "p2" being the length. These dots (multiple sets of the design parameters 203) are being sampled in a latent space 403, as shown in Fig. 4, whereby it can be illustrated as a sphere in this Fig. 4.
[0051] In step (iv)(b), multiple sets of simulated design's approximate output responses 207 are collected 603 from said trained surrogate model 1401, for each set of design parameters 203. This can be shown in Fig. 14.
[0052] In step (iv)(c), pre-processing is performed 605 on said multiple sets of design parameters 203. The process of pre-processing the population is crucial to ensure generic use cases of the RDE algorithm 1501 for all types of design's parameters 203 optimization, because the number of design's parameters 203 are different from one another. The method of perform said pre-processing 605 is not limited to any preprocessing techniques, including feature engineering such as feature reduction, feature selection, feature generation, normalisation, standardisation, and autoencoder. Feature reduction can include Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), Locally Linear Embedding (LLE) or t- Distributed Stochastic Neighbour Embedding (t-SNE).
[0053] In step (iv)(d), as shown in Fig. 16, said pre-processed multiple sets of design parameters 1511 are supplied 607 to at least one any suitable artificial intelligence (Al) algorithm model 1505 such as reinforcement learning (RL) model, supervised learning (SL), polynomial approximation and / or differential equations approximation, to obtain at least one differential evolution (DE) algorithm's hyperparameters 1507. One of the objectives of said Al algorithm model 1505 such as RL model is to use the pre- processed design parameters 1511 in said population as the input, and generate the DE algorithm's 1503 hyperparameters 1507. Conventionally, the DE algorithm hyperparameters 1507 are either fixed or adjusted by mathematical approaches without using Al machine learning methods during the optimization process. The unique DE algorithm of the present invention is enhanced with said RL model to control the DE algorithm's hyperparameters 1507, all being part of said RDE optimization algorithm 1501. The block diagram of said RDE algorithm 1501 is shown in Fig. 15.
[0054] In step (iv(e), said hyperparameters 1507 are updated 609 in said DE algorithm 1503. Said update can be done automatically. The DE's hyperparameter 1507 is not limited to mutation rate / scaling rate, crossover rate, selection strategy, population size or selection strategy. These hyperparameters 1507 are crucial in adjusting the population of design parameters 203, in the design optimization method, to search for the right design's parameters that can give the correct design's output response to achieve the main objective, which is achieving the design's goals 205.
[0055] In step (iv)(f), multiple sets of design parameters 203 are adjusted 611 by said DE algorithm 1503 using the DE's hyperparameters 1507 generated by the RL model 1505. The DE's hyperparameters 1507 such as crossover rate and scaling factor, are scalar values that can be used in any suitable equation or algorithm in the formation of mutate 905 and trial 907. The objective of this process is to gradually (in faster rate than the conventional DE optimisation process) converge the set of the design's parameters 203 in the population 1509 to the optimal state, where the optimal state is the state when at least one set of the design's parameters 203 can give the fitness that satisfies the design's goals 205.
[0056] As shown in Fig. 8, sub-step (iv)(f) of adjusting 611 multiple sets of design parameters by said DE algorithm comprises of the following sub-steps. In sub-step (iv)(f)(i), multiple distinct sets of design parameters are selected 801 from the current set of design parameters 203 in the current population 1509, using at least one predictive model 1701. The adjusted population workflow of step (iv)(f) is an iterative process, whereby said predictive model 1701 iterates through the population 1509 and performs major processes such as selection, mutation, and crossover, in the DE algorithm 1503. For each set of the design parameters in the population (referred to as the "target" 901), the distinct multiple set of design parameters in the population (referred to as the "candidates" 903) are selected from said predictive model 1701 (not limited to Al machine learning model or any other mathematical model). The candidates 903 must be distinct to the target 901. As shown in Fig. 9A, in the current turn (for example, turn 1), a sample (a set of design parameters) is targeted / focussed, referring to said target 901. There may be three candidates 903 being chosen.
[0057] Sub-step (iv)(f)(i) comprises of the following sub-steps, as shown in Fig. 13. In sub-step (iv)(f)(i)(a), the fitness value of the whole sample of multiple sets of design parameters is computed 1301 based on said multiple sets of simulated design's approximate output responses. The computation of fitness value is done using at least one mathematical approach. Using the population's design responses from step (iv)(b), the respected fitness for each set of design parameters in the population is computed using the same method in sub-step (iv)(f)(iv). In sub-step (iv)(f)(i)(b), the set of design parameters with the highest fitness value is identified 1303, said highest fitness is referred to as "best set" 1703. In sub-step (iv)(f)(i)(c), said current set (target) 901 of design parameters and said set of design parameters with the highest fitness value (best set) 1703 are pre-processed 1305 using the same method as in step (iv)(c). In sub-step (iv)(f)(i)(d), said pre-processed current set (target) 901A of design parameters and pre-processed set of design parameters with the highest fitness value (best set) 1703 are used 1307 on said predictive model 1701 to obtain an acceptable number (typically less than 5) of distinct sets of design parameters. The block diagram of this predictive model 1701 is shown in Fig. 17.
[0058] In sub-step (iv)(f)(ii), a first distinct set of design parameters is obtained 803 from said multiple distinct sets of design parameters using at least one mutation algorithm. The candidates 903 are processed using said mutation algorithm, as shown in Fig. 9B, to generate a new set of design parameters (referred to as "mutate" 905) using the mutation rate / scaling rate obtained from the RL model. The mutation algorithm is not limited to the common mutation algorithm in scope of DE such as best / 1, rand / 1, best / 2, rand / 2, or current-to-best / 1.
[0059] Sub-step (iv)(f)(iii) includes performing 805 at least one crossover process on said sets of design parameters in said population (mutate 905) and design parameter derived from said mutation algorithm (target 901) using a crossover rate obtained from said RL model 1505 and sensitivity analysis, to obtain a second set of design parameters (referred to as "trial" 907). The sensitivity analysis is not limited to eigenvectors analysis, Sobol analysis, Morris analysis or Monte-Carlo analysis. As shown in Fig. 9C, target 901 and the mutate 905 will undergo a crossover process to generate said trial 907.
[0060] In sub-step (iv)(f)(iv), the fitness value of said current set (target) 901 of design parameters and said second set (trial) 907 of design parameters are computed 807, using at least one mathematical approach such as LI norm, L2 norm, mean squared error (MSE), mean absolute error (MAE) or Kullback-Leibler Divergence. For target 901, said mathematical approach is done between said current set of design parameter's (the target's) output response 207 and the design goals 205. For trial 907, said mathematical approach is done between the second set of design parameter's (trial's) output response 207 and the design goals 205. The fitness is represented in numerical form. Fig. 10 shows the respected design's output responses in numerical form of the trial 907 and the target 901.
[0061] Sub-step (iv)(f)(v) involves comparing said target's 901 fitness with said trial's 907 fitness, and replacing 809 the current set (target) 901 of design parameters with said second set (trial) 907 of design parameters in said multiple sets of design parameters, if said second set (trial) 907 of design parameters' fitness is higher than the fitness value of said current set (target) 901 of design parameters, thereby adjusting the multiple sets of design parameters. Otherwise, the target 901 remains and the trial 907 is discarded. The population 1509 is adjusted when the trial 907 is successfully accepted and replaces the target 901, to become adjusted population 1509A. As shown in Fig. 11, the target 901 is replaced by trial 907 if trial 907 has a higher fitness.
[0062] Otherwise, the target 901 remains unchanged.
[0063] Sub-step (iv)(f)(vi) includes repeating 811 sub-steps (iv)(f)(i) to (iv)(f)(v) for the subsequent current set (target) 901 of design parameters until there are no more current sets 901 of design parameters, whereby the last current set 901 of design parameters will be used in step (iv)(g). This is shown in Fig. 12.
[0064] Step (iv)(g), involves identifying 613 set of design parameters in said adjusted multiple sets of design parameters with the highest fitness value. For each set of design's parameters 203 in the population 1509, the respected fitness is evaluated to identify which set of the design's parameters has the highest fitness. The fitness is computed automatically using mathematical approach.
[0065] The set of the design's parameters generated with the highest fitness, using substep (iv)(g), is used to perform sub-step (iv)(h), whereby sub-steps (iv)(a) to (iv)(g) are repeated 615 if said multiple sets of design parameters with the highest fitness value does not meet said design goal 205. The determination of the meeting of said design goal can be within a margin of acceptance or tolerance. The termination criteria include if the highest fitness (from step (iv)(g)) respects the design goals 205, or if the maximum iterations of the optimization process are reached, or if the population convergence is reached. The population convergence is high means all sets of design's parameters 203 in the population 1509 are very much similar to each other, while low convergence means all sets of design's parameters 203 in the population 1509 are very much different to each other. A population converging or reaching convergence means the population has high convergence. The convergence is represented in numerical form using mathematical approach not limited to computing population's standard deviation or variance. If the criteria are not satisfied, outside the margin of acceptance, or not within the tolerance, then the population is replaced by the adjusted population (from step (iv)(f)) and repeats the process from step(iv)(b) to step (iv)(h).
[0066] Fig. 18A shows an example of the main interface of the graphics user interface (GUI) of the implementation of the method of Al-assisted design optimization of the present invention. The GUI is able to assist the user to implement the workflows of said method. In this example, the GUI comprises of design output response plotting 1801, optimization report 1803, component attributes panel 1805, variable attributes panel 1807, goal settings panel 1809, operating system information 1811, optimizer settings panel 1813 and optimization cost chart 1815. In the design output response plotting 1801, the design goals, design's output response before optimization and design's optimised output responses are plotted. The optimization report 1803 shows the cost value of each design's goals. The component attributes panel 1805 allow the users to observe the design's parameters available to be optimised, select which design's parameters to be optimized, change the design's parameters and / or plot the design's output responses. The variable attributes panel 1807 allows users to set up the selected design's parameters to be optimized, such as range of design's parameters. The goal settings panel 1809 allows users to set up design's goals. The operating system information 1811 displays the operating system's information to the user. The optimizer settings panel 1813 allows users to configure the method of the present invention, such as using a surrogate model, using multiprocessing and / or setting up how many evaluations / iterations to be run. The optimization cost chart 1815 plots the cost value per evaluation or iteration, besides showing a table of the previous results. Fig. 18B shows an example of the GUI of the implementation of the method of Al-assisted design optimization of the present invention, during optimization process.
[0067] Fig. 19 is a bar chart showing the comparison of cost functions between optimizing done using the method of the present invention, Generic Algorithm (GA) and Particle Swarm Optimization (PSO). These cost functions, including Ackley, Sphere, I2norm, etc are the standard functions available to be used to benchmark optimizers. The lower the cost function value, the better the performance of the optimizer, because the lower the cost, the better the fitness.
[0068] Fig. 20 A is a table showing comparison between various 3D design simulation's output responses, whereby the dashed curves are the design's output responses before optimization, the solid curves are the design's output responses after optimization and the dotted lines are the design's goals. The optimizing method of the present invention is compared with the Genetic Algorithm (GA), Multi-Objectives Genetic Algorithm (MOG A) and Adaptive Multiple Objectives (AMO). Fig. 20B is a graph of the best response using GA, Fig. 20C is a graph of the best response using MOGA, Fig. 20D is a graph of the best response using AMO and Fig. 20E is a graph of the best response using the method of the present invention. From the tabulated results in Fig.
[0069] 20A, the method of the present invention is able to reduce the cost to absolute zero (highest fitness) with a considerably shorter number of evaluations or iterations. While the present invention has been shown and described herein in what are considered to be the preferred embodiments thereof, illustrating the results and advantages over the prior art obtained through the present invention, the invention is not limited to those specific embodiments. Thus, the forms of the invention shown and described herein are to be taken as illustrative only and other embodiments may be selected without departing from the scope of the present invention, as set forth in the claims appended hereto.
Claims
WHAT IS CLAIMED IS:
1. A method of artificial intelligence (Al)-assisted design optimization (101), comprising the following steps: i. obtaining (103) at least one design (201); ii. defining (105) at least one design goal (205); hi. training (107) at least one surrogate model (1401); iv. optimizing (109) parameters (203) of said design (201) with reinforced differential evolution (RDE) algorithm (1501) and said trained surrogate model (1401) to achieve said design goal (205).
2. The method of Al-assisted design optimization (101) as claimed in Claim1, wherein said step (iii) of training (107) surrogate model (1401) comprises of the following sub-steps: a. sampling (301) multiple sets of design parameters (203); b. performing simulation (303) on said design (201) based on said sampled sets of design parameters (203) using at least one simulation software, before collecting multiple sets of simulated design's output responses (207) for each set of design parameter (203);c. constructing (305) at least one data set (501) comprising said multiple sets of design parameters (203) and said multiple sets of simulated design's output responses (207); d. fitting (307) said data set (501) into said surrogate model (1401) to obtain function of said output response (207) with respect to said design parameter (203); e. evaluating (309) said surrogate model (1401) to identify said surrogate model's performance by analysing said surrogate model's (1401) uncertainty using at least one mathematical analysis; f. repeating sub-steps (iii)(a) to (iii)(e) if said uncertainty does not meet a predetermined level (311) using another set of design parameters (203).
3. The method of Al-assisted design optimization (101) as claimed in Claim 2, wherein said mathematical analysis is standard deviation, variance, or the entropy of said surrogate model's output.
4. The method of Al-assisted design optimization (101) as claimed in Claim1, wherein said step (iv) of optimizing (109) parameters (203) of said design (201) with RDE algorithm (1501) and said trained surrogate model (1401) comprises of the following sub-steps: a. sampling (601) multiple sets of design parameters (203);b. collecting (603) multiple sets of simulated design's approximate output responses (207) from said trained surrogate model (1401), for each set of design parameters (203); c. performing pre-processing (605) on said multiple sets of design parameters (203); d. supplying (607) said pre-processed multiple sets of design parameters (1511) to at least one any suitable artificial intelligence (Al) algorithm model (1505) to obtain at least one differential evolution (DE) algorithm's hyperparameters (1507); e. updating (609) said hyperparameters (1507) in said DE algorithm(1503); f. adjusting (611) multiple sets of design parameters (203) by said DE algorithm (1503); g. identifying (613) set of design parameters in said adjusted multiple sets of design parameters with the highest fitness value; h. repeating (615) sub-steps (iv)(a) to (iv)(g) if said multiple sets of design parameters with the highest fitness value does not meet said design goal (205).
5. The method of Al-assisted design optimization (101) as claimed in Claim 4, wherein said pre-processing (605) is any suitable feature engineering method such as feature reduction, feature selection, feature generation, normalisation, standardization and autoencoder.
6. The method of Al-assisted design optimization (101) as claimed in Claim4, wherein said differential evolution (DE) algorithm's hyperparameters (607) is mutation rate, crossover rate, selection strategy or combination thereof.
7. The method of Al-assisted design optimization (101) as claimed in Claim 4, wherein said Al algorithm model (1505) is reinforcement learning(RL), supervised learning (SL), polynomial approximation, differential equations approximation or combination thereof.
8. The method of Al-assisted design optimization (101) as claimed in Claim 4, wherein said sub-step (iv)(f) of adjusting (611) multiple sets of design parameters by said DE algorithm comprises of the following sub-steps: i. selecting (801) multiple distinct sets of design parameters from the current set of design parameters (203), using at least one predictive model (1701);ii. obtaining (803) a first distinct set of design parameters from said multiple distinct sets of design parameters using at least one mutation algorithm; iii. performing (805) at least one crossover process on said sets of design parameters in said population and design parameter derived from said mutation algorithm, using a crossover rate obtained from said RL model (1505) and sensitivity analysis, to obtain a second set of design parameters; iv. computing (807) the fitness value of said current set (901) of design parameters and said second set (907) of design parameters; v. replacing (809) the current set (901) of design parameters with said second set (907) of design parameters in said multiple sets of design parameters, if said second set (907) of design parameters' fitness is higher than the fitness value of said current set (901) of design parameters, thereby adjusting the multiple sets of design parameters; vi. repeating (811) sub-steps (iv)(f)(i) to (iv)(f)(v) for the subsequent current set (901) of design parameters until there are no more current sets (901) of design parameters, wherebythe last current set (901) of design parameters will be used in step (iv)(g).
9. The method of Al-assisted design optimization (101) as claimed in Claim8, wherein said mutation algorithm is best / 1, rand / 1, best / 2, rand / 2, or current-to-best / 1.
10. The method of Al-assisted design optimization (101) as claimed in Claim 8, wherein said sub-step (iv)(f)(iv) of computing fitness value is done using at least one mathematical approach.
11. The method of Al-assisted design optimization (101) as claimed in Claim 8, wherein said sub-step (iv)(f)(i) comprises of the following sub-steps: a. computing (1301) fitness value of the whole sample of multiple sets of design parameters based on said multiple sets of simulated design's approximate output responses; b. identifying (1303) the set of design parameters with the highest fitness value; c. performing pre-processing (1305) of said current set (901) of design parameters and said set of design parameters with the highest fitness value;d. obtaining (1307) an acceptable number of distinct sets of design parameters using said pre-processed current set of design parameters and pre-processed set of design parameters with the highest fitness value on said predictive model (1701).
12. The method of Al-assisted design optimization (101) as claimed in Claim11, wherein said sub-step (iv)(f)(i)(a) of computing fitness value is done using at least one mathematical approach.
13. The method of Al-assisted design optimization (101) as claimed in Claim 10 or 12, wherein said mathematical approach is LI norm, L2 norm, or mean squared error (MSE), mean absolute error (MAE) or Kullback-Leibler Divergence between said current set of design parameters' output response (207) and said design goals (205) of said current set of design parameters; and between said second set of design parameters' output response (207) and said design goals (205) of said second set of design parameters.
Citation Information
Patent Citations
Multi-target seawater intrusion management model based on adaptive DNN substitution model
CN111723516A
Structural Topology Optimization Using Numerical Derivatives
US20160078161A1
Method and device for processing data associated with a simulation model for at least one aspect of a technical system
US20220092237A1
Systems and methods for generating molded parts comprising reinforced composite materials
US20220108053A1
Transforming a model in a first language to a surrogate in a second language for simulation
WO2022240893A1
Cited By
Model training method and device, chip design method and device and electronic equipment
CN119416666A
Wind turbine generator part topology-parameter collaborative optimization method fused with artificial intelligence
CN121598698A
