WO2020253055A1 - 一种基于遗传算法和机器学习的并行模拟电路优化方法 - Google Patents
一种基于遗传算法和机器学习的并行模拟电路优化方法 Download PDFInfo
- Publication number
- WO2020253055A1 WO2020253055A1 PCT/CN2019/117984 CN2019117984W WO2020253055A1 WO 2020253055 A1 WO2020253055 A1 WO 2020253055A1 CN 2019117984 W CN2019117984 W CN 2019117984W WO 2020253055 A1 WO2020253055 A1 WO 2020253055A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- circuit
- machine learning
- optimization
- spice
- model
- Prior art date
Links
- 238000005457 optimization Methods 0.000 title claims abstract description 153
- 238000010801 machine learning Methods 0.000 title claims abstract description 86
- 238000000034 method Methods 0.000 title claims abstract description 77
- 238000004422 calculation algorithm Methods 0.000 title claims abstract description 50
- 230000002068 genetic effect Effects 0.000 title claims abstract description 45
- 238000004088 simulation Methods 0.000 claims abstract description 123
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000004364 calculation method Methods 0.000 claims abstract description 10
- 235000013599 spices Nutrition 0.000 claims abstract 20
- 238000013461 design Methods 0.000 claims description 62
- 238000012360 testing method Methods 0.000 claims description 22
- 238000005070 sampling Methods 0.000 claims description 20
- 238000005516 engineering process Methods 0.000 claims description 16
- 230000035772 mutation Effects 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 238000012795 verification Methods 0.000 claims description 11
- 238000013528 artificial neural network Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 10
- 238000007637 random forest analysis Methods 0.000 claims description 7
- 238000012706 support-vector machine Methods 0.000 claims description 7
- 238000003066 decision tree Methods 0.000 claims description 5
- 238000013178 mathematical model Methods 0.000 claims description 4
- 238000003062 neural network model Methods 0.000 claims 1
- 230000000694 effects Effects 0.000 description 5
- 230000009286 beneficial effect Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 4
- 230000008878 coupling Effects 0.000 description 4
- 238000010168 coupling process Methods 0.000 description 4
- 238000005859 coupling reaction Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000007935 neutral effect Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 239000003990 capacitor Substances 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 238000011960 computer-aided design Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000012545 processing 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]
- G06F30/30—Circuit design
- G06F30/36—Circuit design at the analogue level
- G06F30/373—Design optimisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/30—Circuit design
- G06F30/36—Circuit design at the analogue level
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/30—Circuit design
- G06F30/36—Circuit design at the analogue level
- G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
Definitions
- the invention relates to a parallel analog circuit optimization method based on genetic algorithm and machine learning, belonging to the technical field of artificial intelligence and integrated circuit design or computer-aided design.
- SoC system-on-chip
- SoC generally includes two parts of analog circuit and digital circuit.
- the design of the digital circuit part can be quickly realized with the help of mature EDA auxiliary tools.
- the design of the analog part mainly depends on the designer's manual design and debugging with the help of simulation software such as SPICE. Because there are many non-ideal factors in the design of analog circuits, manual debugging faces difficult and time-consuming problems when there are many design variables and large design space. In order to solve the problems faced by manual debugging of analog circuits, it is urgent to realize automatic optimization of analog circuits.
- Model-based optimization methods generally use analytical models of circuits or other proxy models to evaluate circuit performance, and optimize circuit parameters in the design space based on these models to find the optimal design. Since the calculation time based on the model is much shorter than the SPICE simulation time, the advantage of this method is fast. However, because the accuracy of analytical models or other proxy models is lower than that of SPICE simulation, this method has accuracy problems and often requires further processing of optimization results.
- SPICE simulator is an industry-standard integrated circuit simulation tool that can accurately evaluate circuit performance.
- the present invention provides a parallel simulation circuit optimization method based on genetic algorithm and machine learning.
- the invention combines the advantages of model-based and simulation-based optimization methods, takes into account the efficiency and accuracy of analog circuit optimization, and proposes an analog circuit optimization method that can achieve SPICE-level optimization accuracy and higher efficiency.
- the cost value represents the pros and cons of the circuit performance, which is calculated according to the circuit performance index value obtained by SPICE simulation.
- the circuit performance index can be one or more, which means that this method can achieve multi-objective optimization.
- Orthogonal Latin square is a special square matrix, that is, each element appears only once in each row and column. Orthogonal Latin square means that in a compound square matrix composed of s Latin squares, there are no repeating tuples, and the s square matrices are said to be orthogonal to each other.
- SPICE simulator SPICE (Simulation program with integrated circuit emphasis) is a powerful analog circuit simulator.
- Parallel simulation technology uses multiple CPUs to perform multiple SPICE simulations at the same time.
- Circuit design variables refer to circuit parameter values that need to be changed in the circuit, such as resistance, capacitance, and the width (W) of the MOS tube.
- Circuit performance parameters circuit performance values. For example, the gain, bandwidth, and phase margin of an operational amplifier; the passband ripple of a complex filter, etc.
- a parallel simulation circuit optimization method based on genetic algorithm and machine learning including global optimization based on genetic algorithm and local optimization based on machine learning.
- the global optimization and the local optimization are performed alternately.
- the genetic algorithm-based global optimization uses SPICE simulation to evaluate circuit performance to achieve higher accuracy, and at the same time combines parallel computing technology to improve optimization efficiency.
- a circuit model based on machine learning technology is established near the best point obtained by global optimization, and further optimized in a local area based on the model to obtain better circuit parameters. The steps are as follows:
- (1) Generate initial population According to the principle of orthogonal Latin square, sampling is performed in the design space, and the obtained data forms the initial population;
- the design space refers to the value range of the circuit design variables, which is determined by the user;
- the sampling refers to the design In the space, select multiple sets of variable values;
- the sampled data includes multiple sets of circuit design variable values, and each set of design variable values includes the value of each design variable;
- the initial population contains multiple sets of variable values.
- each Group design variable values are subjected to SPICE simulation, and multiple SPICE simulations are performed simultaneously to obtain the circuit performance parameters of each individual in the initial population; individual refers to the value of each group of circuit design variables in the initial population, and circuit performance parameters refer to the performance value of the circuit; For example, the gain and bandwidth of the operational amplifier, the passband ripple of the complex filter, etc. Circuit performance parameters are determined by the optimized target circuit.
- the initial population of the genetic algorithm is generated by orthogonal Latin square sampling, which reduces the number of individuals in the required population and makes the individuals as evenly distributed in the design space as possible.
- Combining SPICE simulation with parallel computing the circuit performance of many individuals in the population is evaluated through parallel SPICE simulation, which greatly improves efficiency.
- step (3) Use the training and test data set obtained in step (3) to train and generate a machine learning-based circuit performance model:
- the machine learning-based circuit performance model refers to a mathematical model established by the machine learning algorithm in a local range near the global optimum.
- the model can Reflect the relationship between circuit design variables and performance indicators. Since the machine learning model is established in a local area, the training data required is less than the training data required to establish the model in the global scope, which reduces the time cost. In addition, in a local area, the circuit performance is relatively stable, and the accuracy of the established machine learning model is relatively high, which is beneficial to the improvement of the accuracy of the analog circuit optimization method. In the local optimization process, the established machine learning model is used to replace the SPICE simulator to evaluate the circuit performance. Since the prediction speed of the machine learning model is much faster than SPICE simulation, it can greatly improve efficiency.
- the purpose of searching in the neighborhood of the optimal individual obtained by the global search is to further search for points with better performance and lower cost values.
- the circuit performance is no longer evaluated by SPICE simulation, but is predicted by the machine learning-based circuit performance model trained in step (4), that is, the circuit design variable value is used as the input of the circuit performance model, and the circuit performance model is run.
- the output result of the circuit performance model is the circuit performance parameter;
- the cost value is calculated according to the performance parameter of the circuit output by the circuit performance model, the cost value calculation formula is shown in formula (I), the local search is carried out in the direction of the decrease of the cost value;
- step (6) Perform SPICE simulation verification on the results obtained in step (5) of the partial optimization of the circuit performance model based on machine learning to improve the accuracy of the optimization results in the previous step, and update the initial value of the circuit design variables and simulation results after SPICE simulation verification The best individual in the population; update is to replace the original value with the current value.
- SPICE simulation verification is used to improve the accuracy of optimization results based on local optimization of machine models. The local optimization results verified by SPICE are added to the next-generation population in order to obtain a better population.
- the crossover is to randomly select a pair of individuals in the population according to a preset probability, that is, to randomly select two sets of parameter values of the circuit, and use the crossover operator to cross to generate a new individual; Mutation is the use of mutation operators to mutate certain values of individuals in the population to produce new individuals;
- step (8) For the new population generated in step (8), use SPICE simulator combined with parallel computing technology to execute SPICE simulation in parallel;
- the step (2) preferably refers to: substituting the SPICE simulation result obtained in step (1), that is, the circuit performance parameters of each individual in the initial population into the cost function, to calculate the population
- the cost value F cost ( cost value) of each individual is used to measure the performance of the circuit.
- the individual with the smallest cost value is the optimal individual in the population.
- the cost function F cost is shown in formula (I):
- N is the number of circuit performance indicators to be optimized
- n 1, 2, ... N
- P n is the square of the difference between the SPICE simulation results and the circuit performance parameters for the nth circuit performance parameter
- W n is the weight, representing the importance of the n-th circuit performance parameter, and the value of W n is set by the user as needed, and is a real number.
- the step (3) preferably refers to: in the neighborhood of the optimal individual selected in step (2), the neighborhood refers to 5% of the circuit performance parameters of the optimal individual or a user-defined range
- the neighborhood refers to 5% of the circuit performance parameters of the optimal individual or a user-defined range
- the circuit performance model based on machine learning in step (3) is artificial neural network model (artificial neutral networks, ANN), K-Nearest Neighbor (KNN) model, support vector machine (Support Vector).
- Machine SVM
- DNN Deep Neural Networks
- DT Decision Trees
- Random Forests Random Forests
- the present invention alternately executes global optimization based on genetic algorithm and local optimization based on machine learning, and iterates a certain number of times until the optimization goal is met or the preset number of iterations is reached.
- the present invention proposes a parallel simulation circuit optimization method based on genetic algorithm and machine learning, which not only achieves SPICE-level optimization accuracy, but also significantly improves optimization efficiency.
- the present invention proposes a parallel simulation circuit optimization method based on genetic algorithm and machine learning, which combines genetic algorithm, SPICE simulation and parallel technology to improve optimization efficiency while ensuring optimization accuracy.
- the present invention proposes a parallel simulation circuit optimization method based on genetic algorithm and machine learning.
- a machine learning model is established, and the machine learning model prediction is used to replace SPICE simulation, which improves optimization efficiency.
- the optimization results based on the machine learning model are further simulated and verified by SPICE to ensure the optimization accuracy.
- the present invention proposes a parallel simulation circuit optimization method based on genetic algorithm and machine learning.
- the training data of the machine learning model is generated by SPICE parallel simulation, which saves time.
- the present invention proposes a parallel simulation circuit optimization method based on genetic algorithm and machine learning.
- global optimization the evaluation of genetic algorithm population individuals is achieved through parallel SPICE simulation, and in local optimization, training data is obtained through parallel SPICE simulation.
- the parallelization of the optimization method is realized, and the optimization accuracy is guaranteed while the optimization efficiency is improved.
- Fig. 1 is a schematic flow diagram of a parallel analog circuit optimization method based on genetic algorithm and machine learning of the present invention.
- Embodiment 2 is a structural diagram of a fifth-order complex bandpass filter circuit in Embodiment 1 of a parallel analog circuit optimization method based on genetic algorithm and machine learning of the present invention.
- FIG. 3 is a cost value change curve of a parallel simulation circuit optimization method based on genetic algorithm and machine learning in the optimization process of Embodiment 1 of the present invention.
- FIG. 4 is a structural diagram of a second-order operational amplifier circuit in Embodiment 2 of a parallel analog circuit optimization method based on genetic algorithm and machine learning of the present invention.
- FIG. 5 is a cost value change curve of a parallel simulation circuit optimization method based on genetic algorithm and machine learning in the optimization process of Embodiment 2 of the present invention.
- the circuit is the fifth-order complex filter circuit shown in Figure 2.
- the circuit is composed of a first low-pass filter, a second low-pass filter, and a coupling connection unit.
- the first low-pass filter and the second low-pass filter are both 5th-order active RC low-pass filters
- the coupling connection unit includes 5 sets of coupling resistors. Table 1 lists the design goals of the fifth-order complex filter.
- the optimization goal of this embodiment is to make the passband ripple as small as possible on the premise that the center frequency is 12.24 MHz and the bandwidth is 9 MHz (the deviation does not exceed 5%).
- the resistance values of five groups of coupling resistors R 1 , R 2 , R 3 , R 4 , and R 5 are selected as the circuit design variables to be optimized. The specific implementation steps are as follows: as shown in Figure 1;
- sampling is performed in the design space, and the obtained data forms the initial population; sampling refers to selecting multiple sets of variable values in the design space; the data obtained by sampling includes multiple sets of circuits Design variable values, each group of design variable values includes the value of each design variable;
- the initial population contains multiple sets of variable values.
- design variables for each set Perform SPICE simulation and perform multiple SPICE simulations simultaneously to obtain the circuit performance parameters of each individual in the initial population; individual refers to the value of each group of circuit design variables in the initial population, and circuit performance parameters refer to the performance value of the circuit; for example, calculation The gain and bandwidth of the amplifier, the passband ripple of the complex filter, etc. Circuit performance parameters are determined by the optimized target circuit.
- the initial population of the genetic algorithm is generated by orthogonal Latin square sampling, which reduces the number of individuals in the required population and makes the individuals as evenly distributed in the design space as possible.
- Combining SPICE simulation with parallel computing the circuit performance of many individuals in the population is evaluated through parallel SPICE simulation, which greatly improves efficiency.
- step (1) The SPICE simulation results obtained in step (1), namely the circuit performance parameters of each individual in the initial population, are substituted into the cost function (cost function), and the cost value F cost (substitution value) of each individual in the population is calculated.
- the cost value is used for To measure the performance of the circuit, the individual with the smallest cost value is the optimal individual in the population, and the cost function F cost is shown in formula (I):
- N is the number of circuit performance indicators to be optimized
- n 1, 2, ... N
- P n is the square of the difference between the SPICE simulation results and the circuit performance parameters for the nth circuit performance parameter
- W n is the weight, which represents the importance of the performance of the n-th circuit. The value of W n is set by the user as needed, and is a real number.
- step (3) Generate training and test data sets for training the circuit performance model based on machine learning: refers to: in the neighborhood of the optimal individual selected in step (2), the neighborhood refers to the circuit performance parameters of the optimal individual Use the orthogonal Latin square principle to perform uniform sampling to obtain sampled data; use SPICE simulator and parallel technology to perform multiple SPICE simulations on the sampled data in the neighborhood of the optimal individual at the same time. That is, SPICE simulation is executed in parallel to obtain simulation results; sampling data and simulation results constitute training and test data sets; compared to serial simulation, parallel SPICE simulation can reduce time costs and improve optimization efficiency.
- step (3) Use the training and test data set obtained in step (3) to train and generate a machine learning-based circuit performance model:
- the machine learning-based circuit performance model refers to a mathematical model established by the machine learning algorithm in a local range near the global optimum.
- the model can Reflect the relationship between circuit design parameters and performance indicators. Since the machine learning model is established in a local area, the training data required is less than the training data required to establish the model in the global scope, which reduces the time cost. In addition, in a local area, the circuit performance is relatively stable, and the accuracy of the established machine learning model is relatively high, which is beneficial to the improvement of the accuracy of the analog circuit optimization method. In the local optimization process, the established machine learning model is used to replace the SPICE simulator to evaluate the circuit performance. Since the prediction speed of the machine learning model is much faster than SPICE simulation, it can greatly improve efficiency.
- the purpose of searching in the neighborhood of the optimal individual obtained by the global search is to further search for points with better performance and lower cost values.
- the circuit performance is no longer evaluated by SPICE simulation, but is predicted by the machine learning-based circuit performance model trained in step (4), that is, the circuit design variable value is used as the input of the circuit performance model, and the circuit performance model is run.
- the output result of the circuit performance model is the circuit performance parameter;
- the cost value is calculated according to the performance parameter of the circuit output by the circuit performance model, the cost value calculation formula is shown in formula (I), the local search is carried out in the direction of the decrease of the cost value;
- step (6) Perform SPICE simulation verification on the results obtained in step (5) of the partial optimization of the circuit performance model based on machine learning to improve the accuracy of the optimization results in the previous step, and update the initial value of the circuit design variables and simulation results after SPICE simulation verification The best individual in the population; update is to replace the original value with the current value.
- SPICE simulation verification is used to improve the accuracy of optimization results based on local optimization of machine models. The local optimization results verified by SPICE are added to the next-generation population in order to obtain a better population.
- the crossover is to randomly select a pair of individuals in the population according to a preset probability, that is, to randomly select two sets of parameter values of the circuit, and use the crossover operator to cross to generate a new individual; Mutation is the use of mutation operators to mutate certain values of individuals in the population to produce new individuals;
- step (8) For the new population generated in step (8), use SPICE simulator combined with parallel computing technology to execute SPICE simulation in parallel;
- the average relative error (Average Relative Error) and the correlation coefficient (Correlation Coefficient) are used to evaluate the accuracy of the circuit performance model based on machine learning.
- the formula calculation of average relative error and correlation coefficient is shown in formula (II) and formula (III):
- n, x and y are the size of the training data set, the predicted value of the machine learning model and the SPICE simulation result.
- the average relative error represents the error between the machine learning model output and the SPICE simulation value.
- the correlation coefficient is a statistical indicator that measures the degree of agreement between the output of the machine learning model and the SPICE simulation value. If the correlation coefficient is equal to 1.0, the model output value and the target value (SPICE simulation value) completely match.
- this embodiment compares six commonly used machine learning models. Including K-Nearest Neighbor (KNN) model, Support Vector Machine (SVM) model, Deep Neural Networks (DNN) model, Random Forests (Random Forests) model, Decision Trees , DT) model and artificial neural network model (artificial neutral networks, ANN).
- KNN K-Nearest Neighbor
- SVM Support Vector Machine
- DNN Deep Neural Networks
- Random Forests Random Forests
- Decision Trees Decision Trees
- DT artificial neural network model
- 10 training data sets and test data sets are used to train and test these six models. Among them, the training set and the test set are generated by SPICE simulation.
- Table 2 lists the minimum, maximum and average values of the average relative errors and correlation coefficients of the six models in this embodiment. Table 2 also lists the training time of each model.
- the relative error of the ANN model is less than 2%, which is smaller than the relative errors of other models.
- the correlation coefficient is very close to 1.0, which is larger than the correlation coefficients of other models. That is to say, the accuracy of the ANN model in this embodiment is the highest.
- the training time of the ANN model is also very small. Considering comprehensively, the ANN model is selected in this embodiment.
- This embodiment compares the following three optimization methods: a global optimization method based on genetic algorithm and SPICE parallel simulation (GA (SPICE)), a global optimization method based on genetic algorithm and SPICE parallel simulation plus a local optimization method based on SPICE simulation ( GA(SPICE)+LMS(SPICE)), global optimization based on genetic algorithm and local optimization method based on artificial neural network (GA(SPICE)+LMS(ANN)).
- GA genetic algorithm and SPICE parallel simulation
- GA(SPICE)+LMS(SPICE) global optimization method based on genetic algorithm and local optimization method based on artificial neural network
- the second method is based on SPICE simulation and combines global optimization and local optimization, so theoretically, it has the highest accuracy and the best optimization effect.
- the fifth-order complex bandpass filter adopts a 130nm CMOS process.
- the three optimization methods all run in a server environment with 80 Intel Xeon 1.9-GHz CPU cores and 125-GB storage.
- Figure 3 shows the change trend of the cost value of the three analog circuit optimization methods. Iteration on the abscissa indicates the number of iterations. The cost value is proportional to the passband ripple of the complex bandpass filter. The smaller the cost value, the better the circuit performance. Table 3 lists the optimization results of these three optimization methods.
- the circuit is the second-order differential operational amplifier circuit shown in Figure 4.
- This circuit is a fully differential second-order operational amplifier circuit for achieving high gain and high linearity .
- the compensation network includes Miller compensation capacitors and zero resistors used to improve phase margin.
- Table 4 lists the optimization indicators of this op amp.
- the optimization goal of this embodiment is to increase the open-loop gain and bandwidth as much as possible on the premise of satisfying the unity gain bandwidth and phase margin.
- the compensation capacitor C c and the compensation resistor R c also need to be optimized. So there are a total of seven design variables to be optimized, namely W 1 , W 3 , W 5 , W 7 , W 9 , C c , R c .
- sampling is performed in the design space, and the obtained data forms the initial population; sampling refers to selecting multiple sets of variable values in the design space; the data obtained by sampling includes multiple sets of circuits Design variable values, each group of design variable values includes the value of each design variable;
- the initial population contains multiple sets of variable values.
- design variables for each set The value of SPICE simulation is performed, and multiple SPICE simulations are performed in parallel to obtain the circuit performance parameters of each individual in the initial population; the individual refers to the value of each group of circuit design variables in the initial population, and the circuit performance parameter refers to the performance value of the circuit; for example, calculation The gain and bandwidth of the amplifier, the passband ripple of the complex filter, etc.
- Circuit performance parameters are determined by the optimized target circuit.
- the initial population of the genetic algorithm is generated by orthogonal Latin square sampling, which reduces the number of individuals in the required population and makes the individuals as evenly distributed in the design space as possible.
- Combining SPICE simulation with parallel computing the circuit performance of many individuals in the population is evaluated through parallel SPICE simulation, which greatly improves efficiency.
- step (1) The SPICE simulation results obtained in step (1), namely the circuit performance parameters of each individual in the initial population, are substituted into the cost function (cost function), and the cost value F cost (substitution value) of each individual in the population is calculated.
- the cost value is used for To measure the performance of the circuit, the individual with the smallest cost value is the best individual in the population, and the cost function F cost is shown in formula (I):
- N is the number of circuit performance indicators to be optimized
- n 1, 2, ... N
- P n is the square of the difference between the SPICE simulation results and the circuit performance parameters for the nth circuit performance parameter
- W n is the weight, which represents the importance of the performance of the n-th circuit. The value of W n is set by the user as needed, and is a real number.
- step (3) Generate training and test data sets for training the circuit performance model based on machine learning: refers to: in the neighborhood of the optimal individual selected in step (2), the neighborhood refers to the circuit performance parameters of the optimal individual Use the orthogonal Latin square principle to perform uniform sampling to obtain sampled data; use SPICE simulator and parallel technology to perform multiple SPICE simulations on the sampled data in the neighborhood of the optimal individual at the same time. That is, SPICE simulation is executed in parallel to obtain simulation results; sampling data and simulation results constitute training and test data sets; compared to serial simulation, parallel SPICE simulation can reduce time costs and improve optimization efficiency.
- step (3) Use the training and test data set obtained in step (3) to train and generate a machine learning-based circuit performance model:
- the machine learning-based circuit performance model refers to a mathematical model established by the machine learning algorithm in a local range near the global optimum.
- the model can Reflect the relationship between circuit design parameters and performance indicators. Since the machine learning model is established in a local area, the training data required is less than the training data required to establish the model in the global scope, which reduces the time cost. In addition, in a local area, the circuit performance is relatively stable, and the accuracy of the established machine learning model is relatively high, which is beneficial to the improvement of the accuracy of the analog circuit optimization method. In the local optimization process, the established machine learning model is used to replace the SPICE simulator to evaluate the circuit performance. Since the prediction speed of the machine learning model is much faster than SPICE simulation, it can greatly improve efficiency.
- the purpose of searching in the neighborhood of the optimal individual obtained by the global search is to further search for points with better performance and lower cost values.
- the circuit performance is no longer evaluated by SPICE simulation, but is predicted by the machine learning-based circuit performance model trained in step (4), that is, the circuit design variable value is used as the input of the circuit performance model, and the circuit performance model is run.
- the output result of the circuit performance model is the circuit performance parameter;
- the cost value is calculated according to the performance parameter of the circuit output by the circuit performance model, the cost value calculation formula is shown in formula (I), the local search is carried out in the direction of the decrease of the cost value;
- step (6) Perform SPICE simulation verification on the results obtained in step (5) of the partial optimization of the circuit performance model based on machine learning to improve the accuracy of the optimization results in the previous step, and update the initial value of the circuit design variables and simulation results after SPICE simulation verification The best individual in the population; update is to replace the original value with the current value.
- SPICE simulation verification is used to improve the accuracy of optimization results based on local optimization of machine models. The local optimization results verified by SPICE are added to the next-generation population in order to obtain a better population.
- the crossover is to randomly select a pair of individuals in the population according to a preset probability, that is, to randomly select two sets of parameter values of the circuit, and use the crossover operator to cross to generate a new individual; Mutation is the use of mutation operators to mutate certain values of individuals in the population to produce new individuals;
- step (8) For the new population generated in step (8), use SPICE simulator combined with parallel computing technology to execute SPICE simulation in parallel;
- this embodiment compares six commonly used machine learning models. Including K-Nearest Neighbor (KNN) model, Support Vector Machine (SVM) model, Deep Neural Networks (DNN) model, Random Forests (Random Forests) model, Decision Trees , DT) model and artificial neural network model (artificial neutral networks, ANN).
- KNN K-Nearest Neighbor
- SVM Support Vector Machine
- DNN Deep Neural Networks
- Random Forests Random Forests
- Decision Trees Decision Trees
- DT artificial neural network model
- This embodiment compares the accuracy of the six machine learning models with respect to four performance indicators of gain, bandwidth, unity gain bandwidth, and phase margin of the operational amplifier.
- this embodiment uses 10 training data sets and test data sets for training and testing. Among them, the training set and the test set are generated by SPICE simulation.
- Table 5 Table 6, Table 7, and Table 8 list the average relative errors and correlations of the six models established for the four performance indicators of the operational amplifier's gain, bandwidth, unity gain bandwidth, and phase margin in this embodiment. The minimum, maximum and average values of the coefficients. In addition, Table 5 to Table 8 also lists the training time of each model.
- the relative error of the ANN model is less than 2%, which is smaller than the relative errors of other models.
- the correlation coefficient is very close to 1.0, which is larger than the correlation coefficients of other models. That is to say, the accuracy of the ANN model in this embodiment is the highest.
- the training time of the ANN model is also very small. Considering comprehensively, the ANN model is selected in this embodiment.
- Example 1 also compares the three optimization methods in Example 1.
- This op amp uses 130nm CMOS process.
- the three optimization methods all run in a server environment with 80 Intel Xeon 1.9-GHz CPU cores and 125-GB storage.
- Figure 5 shows the change trend of the cost value of the three optimization methods for this op amp.
- the Cost value is used to evaluate circuit performance. The smaller the Cost value, the better the circuit performance.
- Table 9 lists the optimization results of these three optimization methods.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Geometry (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Genetics & Genomics (AREA)
- Physiology (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Description
Claims (4)
- 一种基于遗传算法和机器学习的并行模拟电路优化方法,其特征在于,包括步骤如下:(1)产生初始种群:根据正交拉丁方原理,在设计空间中进行采样,得到的数据形成初始种群;设计空间是指电路设计变量的取值范围,由用户定义决定;采样是指在设计空间中,选择多组变量值;采样得到的数据包括多组电路设计变量值,每组设计变量值包括各个设计变量的值;利用SPICE仿真器并结合并行技术,对初始种群同时进行多次SPICE仿真,并行执行SPICE仿真,即对每组设计变量值进行SPICE仿真并多次SPICE仿真同时进行,得到初始种群中各个个体的电路性能参数;个体是指初始种群中每组电路设计变量值,电路性能参数是指电路的性能值;(2)计算个体的cost值,并选择种群中cost值最小的个体为最优个体:(3)产生用于训练基于机器学习的电路性能模型的训练及测试数据集:(4)利用步骤(3)得到的训练及测试数据集,训练并生成基于机器学习的电路性能模型:利用步骤(3)得到的训练及测试数据集,进行训练模型,所述基于机器学习的电路性能模型,是指利用机器学习算法在全局最优点附近的局部范围内建立的数学模型,该模型能反映电路设计参数和性能指标之间的关系;(5)基于步骤(4)训练好的基于机器学习的电路性能模型,进行局部优化:在全局搜索得到的最优个体的邻域内进行搜索,利用步骤(4)训练好的基于机器学习的电路性能模型进行预测,即将电路设计变量值作为电路性能模型的输入,运行电路性能模型,电路性能模型输出的结果就是电路性能参数;根据电路性能模型输出的电路的性能参数计算cost值,cost值计算公式如式(Ⅰ)所示,局部搜索沿cost值减小的方向进行;(6)对步骤(5)基于机器学习的电路性能模型局部优化得到的结果,进行SPICE仿真验证,将SPICE仿真验证后的电路设计变量值及仿真结果更新初始种群中的最优个体;(7)判断是否满足优化终止条件,如果满足达到预先设定的迭代次数,或满足电路优化的目标,则优化结束,如果不满足,进入步骤(8);(8)进入遗传算法的进化流程,依次通过选择、交叉、变异,产生新的种群;所述选择,是根据cost值,选取种群中的优良个体,cost值越小的个体被选中的概率越大;所述交叉,是按照预先设定的概率,在群体中随机选取种群中的一对个体,即随机选取电路的两组参数值,利用交叉算子进行交叉,产生新的个体;所述变异,是对种群中的个体的某些值,利用变异算子进行突变,产生新的个体;(9)对步骤(8)产生的新种群,利用SPICE仿真器结合并行计算技术,并行执行SPICE仿真;(10)重复步骤(2)到步骤(7)。
- 根据权利要求1所述的一种基于遗传算法和机器学习的并行模拟电路优化方法,其特征在于,所述步骤(3),是指:在步骤(2)所选的最优个体的邻域内,邻域是指最优个体的电路性能参数的5%或者用户定义的范围内,利用正交拉丁方原理进行均匀采样,得到采样数据;利用SPICE仿真器并结合并行技术,对最优个体的邻域内的采样数据同时进行多次SPICE仿真,即并行执行SPICE仿真,得到仿真结果;采样数据和仿真结果构成训练及测试数据集。
- 根据权利要求1-3任一所述的一种基于遗传算法和机器学习的并行模拟电路优化方法,其特征在于,步骤(3)所述基于机器学习的电路性能模型为人工神经网络模型、K近邻模型、支持向量机模型、深度神经网络模型、决策树模型或随机森林模型。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US16/957,998 US11714943B2 (en) | 2019-06-19 | 2019-11-13 | Parallel analog circuit optimization method based on genetic algorithm and machine learning |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910530806.4A CN110245436B (zh) | 2019-06-19 | 2019-06-19 | 一种基于遗传算法和机器学习的并行模拟电路优化方法 |
CN201910530806.4 | 2019-06-19 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020253055A1 true WO2020253055A1 (zh) | 2020-12-24 |
Family
ID=67887985
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2019/117984 WO2020253055A1 (zh) | 2019-06-19 | 2019-11-13 | 一种基于遗传算法和机器学习的并行模拟电路优化方法 |
Country Status (3)
Country | Link |
---|---|
US (1) | US11714943B2 (zh) |
CN (1) | CN110245436B (zh) |
WO (1) | WO2020253055A1 (zh) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112668697A (zh) * | 2020-12-25 | 2021-04-16 | 西安电子科技大学 | 一种无人机飞控控制参数的模糊测试方法及系统 |
CN112712118A (zh) * | 2020-12-29 | 2021-04-27 | 银江股份有限公司 | 一种面向医疗文本数据的过滤方法及系统 |
CN112966447A (zh) * | 2021-03-25 | 2021-06-15 | 北京化工大学 | 一种基于自动机器学习的化学材料吸附性能预测方法及装置 |
CN113381700A (zh) * | 2021-06-29 | 2021-09-10 | 哈尔滨工业大学 | 一种高频无源rc积分放大器的rc参数设计方法 |
CN114429099A (zh) * | 2022-01-27 | 2022-05-03 | 西安交通大学 | 一种模拟集成电路参数优化方法及系统 |
CN114970431A (zh) * | 2022-06-29 | 2022-08-30 | 上海集成电路装备材料产业创新中心有限公司 | Mos管参数估计模型的训练方法和装置 |
CN116844642A (zh) * | 2023-07-03 | 2023-10-03 | 燕山大学 | 基于dna杂交反应技术的新型线性机器学习方法 |
Families Citing this family (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110245436B (zh) * | 2019-06-19 | 2021-03-30 | 山东大学 | 一种基于遗传算法和机器学习的并行模拟电路优化方法 |
CN111639471B (zh) * | 2020-06-01 | 2022-05-27 | 浙江大学 | 一种基于循环神经网络的电磁干扰滤波器设计方法 |
CN111797117A (zh) * | 2020-07-02 | 2020-10-20 | 北京润科通用技术有限公司 | 一种数据处理方法及装置 |
CN112383237B (zh) * | 2020-10-09 | 2022-03-22 | 河南科技大学 | 一种并网逆变器的模型预测控制方法 |
CN112417803B (zh) * | 2020-12-02 | 2024-02-06 | 苏州复鹄电子科技有限公司 | 一种基于人工智能算法的模拟集成电路设计参数自动优化方法 |
CN112836459B (zh) * | 2021-02-02 | 2022-09-20 | 山东大学 | 一种电路解析模型辅助的模拟电路自动优化方法、设备及存储介质 |
CN112906896A (zh) * | 2021-02-19 | 2021-06-04 | 阿里巴巴集团控股有限公司 | 信息处理方法及装置、计算设备 |
CN113051856B (zh) * | 2021-03-24 | 2024-01-19 | 西安电子科技大学 | 集成电路设计数据的优化方法 |
CN113283213B (zh) * | 2021-07-19 | 2021-11-09 | 苏州贝克微电子有限公司 | 一种基于机器学习的电路设计系统及方法 |
CN113627087B (zh) * | 2021-08-23 | 2023-11-03 | 中机中联工程有限公司 | 基于物流仿真与机器学习的生产系统参数自适应优化方法 |
CN114297934A (zh) * | 2021-12-30 | 2022-04-08 | 无锡雪浪数制科技有限公司 | 一种基于代理模型的模型参数并行仿真优化方法及装置 |
CN116312899B (zh) * | 2023-05-12 | 2023-08-01 | 中国人民解放军战略支援部队航天工程大学 | 一种基于激光雷达成像的材料表面brdf参数拟合方法和系统 |
CN116362174B (zh) * | 2023-05-24 | 2023-08-29 | 湖南大学 | Igbt设计参数全局优化方法及系统 |
CN117077508A (zh) * | 2023-07-14 | 2023-11-17 | 武汉理工大学 | 热电发电机的性能预测及优化方法 |
CN116702678B (zh) * | 2023-08-02 | 2024-01-30 | 华南理工大学 | 一种融合深度学习和群智能算法的dtco优化方法 |
CN117473932A (zh) * | 2023-10-27 | 2024-01-30 | 华南理工大学 | 一种敏捷设计库驱动的dtco高效优化方法 |
CN117272756A (zh) * | 2023-11-14 | 2023-12-22 | 江苏勇龙电气有限公司 | 一种gil组合式超高压套管用均压环的设计方法 |
CN117422005B (zh) * | 2023-12-19 | 2024-03-29 | 杭州四维映射软件有限公司 | 一种模拟电路仿真误差自动控制的方法及应用 |
CN117556775B (zh) * | 2024-01-10 | 2024-03-22 | 电子科技大学 | 多目标芯片电路参数优化设计方法 |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102289549A (zh) * | 2011-08-11 | 2011-12-21 | 复旦大学 | 一种模拟集成电路优化方法 |
CN104111871A (zh) * | 2013-04-16 | 2014-10-22 | 济南概伦电子科技有限公司 | 一种用于执行电路仿真中动态负载平衡的方法及装置 |
US9928324B1 (en) * | 2014-04-29 | 2018-03-27 | Cadence Design Systems, Inc. | System and method for accurate modeling of back-miller effect in timing analysis of digital circuits |
CN110245436A (zh) * | 2019-06-19 | 2019-09-17 | 山东大学 | 一种基于遗传算法和机器学习的并行模拟电路优化方法 |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8554530B1 (en) * | 2009-01-20 | 2013-10-08 | Cadence Design Systems, Inc. | Methods and systems for property assertion in circuit simulation |
US9047424B1 (en) * | 2013-10-25 | 2015-06-02 | Cadence Design Systems, Inc. | System and method for analog verification IP authoring and storage |
US10867091B1 (en) * | 2018-08-17 | 2020-12-15 | Synopsys, Inc. | Machine learning based power optimization using parallel training and localized data generation |
US11581700B1 (en) * | 2019-10-04 | 2023-02-14 | Freedom Photonics Llc | Multiple optoelectronic devices with thermal compensation |
US11861722B2 (en) * | 2020-02-18 | 2024-01-02 | BlueOwl, LLC | Systems and methods for generating and updating an inventory of personal possessions of a user for insurance purposes |
CN111931446B (zh) * | 2020-09-24 | 2021-01-01 | 中天弘宇集成电路有限责任公司 | 建模方法 |
-
2019
- 2019-06-19 CN CN201910530806.4A patent/CN110245436B/zh active Active
- 2019-11-13 US US16/957,998 patent/US11714943B2/en active Active
- 2019-11-13 WO PCT/CN2019/117984 patent/WO2020253055A1/zh active Application Filing
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102289549A (zh) * | 2011-08-11 | 2011-12-21 | 复旦大学 | 一种模拟集成电路优化方法 |
CN104111871A (zh) * | 2013-04-16 | 2014-10-22 | 济南概伦电子科技有限公司 | 一种用于执行电路仿真中动态负载平衡的方法及装置 |
US9928324B1 (en) * | 2014-04-29 | 2018-03-27 | Cadence Design Systems, Inc. | System and method for accurate modeling of back-miller effect in timing analysis of digital circuits |
CN110245436A (zh) * | 2019-06-19 | 2019-09-17 | 山东大学 | 一种基于遗传算法和机器学习的并行模拟电路优化方法 |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112668697A (zh) * | 2020-12-25 | 2021-04-16 | 西安电子科技大学 | 一种无人机飞控控制参数的模糊测试方法及系统 |
CN112668697B (zh) * | 2020-12-25 | 2024-04-09 | 西安电子科技大学 | 一种无人机飞控控制参数的模糊测试方法及系统 |
CN112712118A (zh) * | 2020-12-29 | 2021-04-27 | 银江股份有限公司 | 一种面向医疗文本数据的过滤方法及系统 |
CN112966447A (zh) * | 2021-03-25 | 2021-06-15 | 北京化工大学 | 一种基于自动机器学习的化学材料吸附性能预测方法及装置 |
CN113381700A (zh) * | 2021-06-29 | 2021-09-10 | 哈尔滨工业大学 | 一种高频无源rc积分放大器的rc参数设计方法 |
CN113381700B (zh) * | 2021-06-29 | 2024-01-30 | 哈尔滨工业大学 | 一种高频无源rc积分放大器的rc参数设计方法 |
CN114429099A (zh) * | 2022-01-27 | 2022-05-03 | 西安交通大学 | 一种模拟集成电路参数优化方法及系统 |
CN114429099B (zh) * | 2022-01-27 | 2024-01-19 | 西安交通大学 | 一种模拟集成电路参数优化方法及系统 |
CN114970431A (zh) * | 2022-06-29 | 2022-08-30 | 上海集成电路装备材料产业创新中心有限公司 | Mos管参数估计模型的训练方法和装置 |
CN114970431B (zh) * | 2022-06-29 | 2024-04-02 | 上海集成电路装备材料产业创新中心有限公司 | Mos管参数估计模型的训练方法和装置 |
CN116844642A (zh) * | 2023-07-03 | 2023-10-03 | 燕山大学 | 基于dna杂交反应技术的新型线性机器学习方法 |
CN116844642B (zh) * | 2023-07-03 | 2024-03-29 | 燕山大学 | 基于dna杂交反应技术的新型线性机器学习方法 |
Also Published As
Publication number | Publication date |
---|---|
US20230092630A1 (en) | 2023-03-23 |
CN110245436B (zh) | 2021-03-30 |
CN110245436A (zh) | 2019-09-17 |
US11714943B2 (en) | 2023-08-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2020253055A1 (zh) | 一种基于遗传算法和机器学习的并行模拟电路优化方法 | |
Li et al. | An artificial neural network assisted optimization system for analog design space exploration | |
Liu et al. | Filter design for autoregressive moving average graph filters | |
Shokouhifar et al. | Simplified symbolic transfer function factorization using combined artificial bee colony and simulated annealing | |
CN113222031A (zh) | 基于联邦个性化学习的光刻热区检测方法 | |
CN110705080B (zh) | 一种数字孪生模型精准组装方法及装置 | |
CN114841106A (zh) | 一种基于规则引导的遗传算法的集成电路优化方法及系统 | |
Shi et al. | Topological approach to automatic symbolic macromodel generation for analog integrated circuits | |
CN114595656A (zh) | 电子器件的算法模型和版图的生成方法、装置、设备及介质 | |
CN114004191A (zh) | 一种延迟电路宏模型提取方法、系统、设备以及介质 | |
CN111600734B (zh) | 一种网络故障处理模型的构建方法、故障处理方法及系统 | |
CN110675352A (zh) | 一种混合高低阶图卷积传播系统 | |
Biondi et al. | Multi-Objective Evolutionary Algorithms and Pattern Search Methods for Circuit Design Problems. | |
CN115906303A (zh) | 一种基于机器学习的平面微波滤波器设计方法及装置 | |
Rocha et al. | A new metaheuristc combining gradient models with NSGA-II to enhance analog IC synthesis | |
Dai et al. | Multi-objectives design optimization based on multi-objectives Gaussian processes for System-in-Package | |
Tseng et al. | A novel design space reduction method for efficient simulation-based optimization | |
CN110210691B (zh) | 资源推荐方法、装置、存储介质及设备 | |
Liu et al. | High-Efficiency Multiobjective Synchronous Modeling and Solution of Analog ICs | |
Berkol et al. | A hierarchical design automation concept for analog circuits | |
CN112836459B (zh) | 一种电路解析模型辅助的模拟电路自动优化方法、设备及存储介质 | |
Wan et al. | Tackling Over-Smoothing: Graph Hollow Convolution Network with Topological Layer Fusion | |
Yin et al. | An Efficient Reference-point Based Surrogate-assisted Multi-objective Differential Evolution for Analog/RF Circuit Synthesis | |
CN116992806A (zh) | 一种基于自注意力机制的大规模模拟集成电路自动优化方法 | |
Xu et al. | Efficient block pruning based on kernel and feature stablization |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19933294 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19933294 Country of ref document: EP Kind code of ref document: A1 |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19933294 Country of ref document: EP Kind code of ref document: A1 |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 21/09/2022) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19933294 Country of ref document: EP Kind code of ref document: A1 |