An automatic design method for analog integrated circuit size with high PVT robustness

Through the combination of Bayesian optimization and multi-task reinforcement learning, the problem of low PVT change processing efficiency in integrated circuit design is solved. Dynamic PVT angle pruning algorithm is adopted to improve sampling efficiency and convergence speed, ensuring PVT robustness circuit design.

CN118673863BActive Publication Date: 2024-11-26PEKING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410775869.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-17
Publication Date
2024-11-26
Estimated Expiration
2044-06-17

AI Technical Summary

Technical Problem

In the design of integrated circuits, it is difficult to efficiently handle PVT changes, resulting in low sampling efficiency, large calculation overhead and slow convergence speed, especially the problems of sensitivity at the initial sampling point and difficulty in taking into account the performance indicator changes of different PVT angles.

Method used

The Bayesian optimization algorithm is used to perform initial sampling under a single PVT angle, combined with the multi-task reinforcement learning algorithm to perform optimization under a multi-PVT angle, and a dynamic PVT angle pruning algorithm based on the latest sampling is proposed, and the PVT set is dynamically adjusted and streamlined to improve sampling Efficiency and convergence speed.

Benefits of technology

By improving the data quality of the initial data set, improving the success rate and convergence speed of the PVT robustness optimization process, reducing the number of simulations, improving sampling efficiency, and ensuring the robustness of the circuit performance at all PVT angles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118673863B_ABST
    Figure CN118673863B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for automatic design of analog integrated circuit size with high PVT robustness, which uses a Bayesian optimization algorithm to construct an initial data set, performs size optimization under a single PVT angle, and then uses a multi-task reinforcement learning algorithm to perform size optimization under multiple PVT angles, and performs dynamic PVT angle pruning based on the latest sampling. The present invention improves the success rate and convergence speed of the PVT robust size optimization process by improving the data quality of the initial data set; by constructing the latest sampling cache and streamlining the PVT set, the most recent sampling under each PVT angle is updated in batches, and then the PVT robustness of the size parameters is verified by full PVT angle testing, which greatly reduces the number of simulations required for size optimization and improves sampling efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention provides a method for automatically designing the size of an analog integrated circuit with high PVT robustness, and specifically relates to an automatic optimization technology based on Bayesian optimization and multi-task reinforcement learning, belonging to the technical field of integrated circuit design automation. Background Art

[0002] The existing integrated circuit design automation technologies include the following for the design of integrated circuit size:

[0003] 1) Incorporate PVT changes into the sizing process; for example, in the article "An Efficient Batch-Constrained Bayesian Optimization Approach for Analog Circuit Synthesis via Multiobjective Acquisition Ensemble" published in the IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (TCAD) journal in 2022 by Zeng Xuan's team from Fudan University, PVT changes were incorporated into the sizing process by traversing all PVT (Process, Voltage, Temperature) corners. Under this technology, the consideration of circuit PVT robustness requires dozens of times more circuit simulation times (i.e., sampling times), and the sampling efficiency is low, so the PVT corners need to be efficiently pruned.

[0004] 2) Highly robust size optimization by conquering PVT corners; for example, Kai-En Yang et al. published an article titled "Trust-Region Method with Deep Reinforcement Learning in Analog Design Space Exploration" at the 58th Design Automation Conference in 2021, which recorded the completion of highly robust size optimization by continuously conquering PVT corners that currently do not meet circuit index requirements. However, this technology requires training an independent network model for each PVT corner, which brings additional computational overhead.

[0005] 3) Optimize the size of each PVT corner through multi-task reinforcement learning; for example, Wei Shi et al. from the University of Texas at Austin published the article "RobustAnalog: Fast Variation-Aware Analog Circuit Design Via Multi-task RL" and the related optimization tool RobustAnalog at the 2022 Machine Learning for CAD (MLCAD) conference. This tool deploys a multi-task reinforcement learning algorithm to handle the size optimization process at each PVT corner, but the tool is sensitive to the quality of the initial sampling points. If the performance indicators corresponding to the initial sampling points are poor, the convergence speed of the size optimization process is slow and the success rate is low. The shortcomings of this method are more obvious when automatically designing circuits such as bandgap reference sources that are extremely sensitive to size changes. In terms of processing the PVT angle, this technology first uses the mean clustering algorithm to cluster the PVT angles, and then selects the PVT angles with the worst performance indicators in each category to form a simplified PVT set. During sampling, only the size parameters are sampled under the PVT angles included in the simplified PVT set, thereby completing PVT angle pruning and improving sampling efficiency. However, the disadvantage of this technology is that as long as the size design process under the current simplified PVT set has not been completed, the PVT angles included in the simplified PVT set will not change, and it is difficult to take into account the changes in performance indicators under other PVT angles, thereby limiting the convergence speed of the size design process. Summary of the invention

[0006] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a method for automatic design of analog integrated circuit size with high PVT robustness. The present invention first divides the high PVT robustness size design process into 1) initial sampling based on size design under a single PVT angle and 2) size design under multiple PVT angles. The present invention first uses a Bayesian optimization (BO: Bayesian Optimization) algorithm to perform size optimization under a single PVT angle to establish an initial data set, and then uses a multi-task reinforcement learning algorithm to perform size optimization under multiple PVT angles. This size design method that integrates Bayesian optimization and multi-task reinforcement learning can solve technical problems existing in the automatic design of analog circuit size with high PVT robustness, including sensitivity to initial sampling points, the need for multiple independent network models, and low sampling efficiency.

[0007] In order to further improve the sampling efficiency of the size design process, the present invention proposes a dynamic PVT angle pruning algorithm based on the latest sampling. This pruning algorithm not only reduces the number of sampling times required to evaluate a single size parameter by constructing a simplified PVT set, but also dynamically adjusts the PVT angle in the simplified PVT set in each iteration according to the latest sampling under each PVT angle, so as to take into account the conflict of size optimization directions between different PVT angles, and further improve the convergence speed of the size design process.

[0008] The present invention adopts BO algorithm and multi-task reinforcement learning algorithm. The focus of the present invention is 1) proposing a sizing method that integrates Bayesian optimization and multi-task reinforcement learning and 2) proposing a method that uses the latest sampling cache to evaluate the sizing progress under each current PVT angle, and based on this, a dynamic PVT angle pruning method based on the latest sampling.

[0009] For convenience, the definitions of terms in this invention include:

[0010] Sampling: refers to simulating a set of dimensional parameters under a certain PVT angle and obtaining circuit performance indicators;

[0011] Sampling efficiency: used to describe the number of sampling times required by the automatic dimension design tool to complete the dimension design. The fewer the number of sampling times required, the higher the sampling efficiency of the automatic design tool.

[0012] PVT corner: refers to the combination of process conditions (Process), power conditions (Voltage) and temperature conditions (Temperature) set when performing analog integrated circuit simulation. For the following PVT corner, {Process=FF, Voltage=0.9Vref, Temperature=27℃}, it means that the process conditions are Fast NMOS&FastPMOS, the power conditions are 90% of the reference voltage, and the operating temperature is 27℃.

[0013] Reduced PVT set: A set of several PVT angles that is a subset of the set consisting of all PVT angles.

[0014] PVT angle pruning: when evaluating size parameters, sampling is not performed at all PVT angles. Instead, some PVT angles are selected for sampling through certain rules (such as building a reduced PVT set) to reduce sampling overhead.

[0015] The technical solution provided by the present invention is:

[0016] A method for automatically designing the size of an analog integrated circuit with high PVT robustness comprises the following steps:

[0017] A. Determine the integrated circuit size design task;

[0018] A1. Determine the circuit topology and the size parameters to be adjusted;

[0019] A2. Determine the performance indicators that the circuit needs to achieve;

[0020] A3. Build a test circuit;

[0021] B. Perform initial sampling based on the Bayesian optimization algorithm, and use the size optimization under a single PVT angle based on the Bayesian algorithm as a way to construct the initial data set to improve the efficiency and success rate of size optimization sampling; the implementation method is;

[0022] B1. Under a single PVT angle (e.g. {Process = TT, Voltage = Vref, Temperature = 27°C}), a set of dimensional parameters that meet the performance index constraints required for all circuits under this PVT angle is obtained using the trust region Bayesian optimization algorithm;

[0023] B2. Select the best n sets of size parameters (those with the best quality factor (FoM)) found during the Bayesian optimization algorithm; n is generally set to 4 to 8 to balance sampling efficiency and convergence speed;

[0024] B3. Sample the n groups of size parameters selected in B2 at all PVT angles to generate an initial data set;

[0025] The total PVT corner includes all combinations of process, voltage and temperature variations. For example, if the possible process deviations are {TT, SS, FF, SF, FS}, the voltage conditions considered are {0.9Vref, Vref, 1.1Vref}, and the temperature deviations considered are {-40℃,

[0026] 27℃,80℃}, then all PVT angles refer to all possible PVT angles contained in all 5*3*3 combinations. The data set includes n sets of dimensional parameters with the best quality factor generated in B2, as well as the performance indicators and quality factors of these dimensional parameters under all PVT angles. Each data in the data set is a typical data of one word, and each data has four key values, including "dimensional parameters", "PVT angle", "performance indicators" and "quality factor".

[0027] C. Perform PVT robust circuit size optimization based on multi-task reinforcement learning; the implementation method is:

[0028] C1. Build a data cache. Build a data cache for each PVT angle. Each data cache consists of two parts: 1) historical data cache, i.e., data obtained from all simulations performed under this PVT angle, and 2) latest sampling cache, i.e., data obtained from the most recent simulation performed under this PVT angle;

[0029] C2. Load initial data. Load the initial data set into the historical data cache built in C1, and randomly select a set of data from it and load it into the latest sampling cache;

[0030] Random selection is based on the construction method of the initial data set. In the initial data set, there are n data under each PVT angle. Because the latest sampling cache cannot be empty, one of the n data is randomly selected and loaded into the latest sampling cache.

[0031] Construct a failed PVT set (or other PVT corners of interest) and initialize it to empty. In the subsequent process, when the performance index under a certain PVT corner does not meet the requirements, the PVT corner is added to the failed PVT set. The failed PVT set will not be initialized again during the entire size optimization process, that is, once a PVT corner appears in the failed PVT set, it will not be removed.

[0032] C3. Perform dynamic PVT corner pruning based on the latest sampling;

[0033] Construct a streamlined PVT set to reduce the PVT angle required for sampling in each iteration and improve sampling efficiency.

[0034] Read the data in the latest sampling cache, perform K-means clustering on the PVT corners according to the performance indicators of the most recent sampling recorded under each PVT corner, and select the PVT corner with the worst performance indicator in each category to be included in the simplified PVT set.

[0035] Adjust the simplified PVT set according to the situation. Add the failed PVT set to the simplified PVT set. For specific operations, see step D3;

[0036] C4. Use a multi-task reinforcement learning agent to give the size parameter to be sampled in each iteration, and sample the size parameter under the PVT angle contained in the simplified PVT set;

[0037] In the specific implementation, the structure of the multi-task reinforcement learning agent used is composed of an Actor network and a Critic network. The agent regards the size optimization under a PVT angle as a task. In the agent, the input of the Actor network is the PVT angle, each PVT angle is represented by a 7-dimensional vector, the first five dimensions of the vector are the process angles represented by the one-hot code (such as [1, 0, 0, 0, 0] represents the TT process angle), and the last two dimensions of the vector represent the normalized power supply voltage and operating temperature. The output of the Actor network is the size parameter to be sampled. The input of the Critic network is the size parameter to be sampled, and the output is the predicted circuit performance index of the size parameter under each PVT angle.

[0038] C5. Update the sampling results in C4 to the historical data cache and the latest sampling cache of the corresponding PVT angle;

[0039] C6. Read the latest sampling cache corresponding to all PVT angles. If the performance indicators of the latest sampling under each PVT angle meet the requirements, jump to step D. PVT pruning is performed on the sampling stage in each round, and the latest sampling of each PVT angle does not necessarily correspond to the same size parameter.

[0040] C7. Update the network weights of the reinforcement learning agent according to the data in the data cache (the update method is described in the implementation method).

[0041] ), and jump to step C3;

[0042] D. Conduct full PVT angle test, the specific implementation plan is:

[0043] D1. Sample the size parameters to be sampled given in the most recent step C4 at all PVT angles.

[0044] D2. If the performance indicators of each PVT angle sampling in D1 meet the requirements, the full PVT angle test is passed, and the sampling size parameters in the most recent D1 are the final output of the size automatic design tool, and the process ends;

[0045] By performing full PVT testing, the size parameters are sampled at all PVT angles without PVT pruning, and the PVT robustness of the size parameters is finally verified. When the performance indicators at a certain PVT angle in D1 do not meet the requirements, the simplified PVT set is updated and sampling and testing continues.

[0046] D3. If the performance index under a certain PVT angle in D1 does not meet the requirements, then add the PVT angle to the failed PVT set;

[0047] D4. Update the sampling results of the full PVT angle test to the latest sampling cache, and jump to step C3.

[0048] Through the above steps, the automatic design of the size of the analog integrated circuit with high PVT robustness is achieved.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1) The design framework provided by the present invention, which first uses Bayesian optimization for initial sampling and then uses multi-task reinforcement learning for PVT robust size optimization, can improve the success rate and convergence speed of the PVT robust size optimization process by improving the data quality of the initial data set;

[0051] 2) Through the PVT angle pruning algorithm based on the latest sampling, the most recent sampling under each PVT angle is updated in batches, and then the PVT robustness of the size parameters is verified through full PVT angle testing. This approach can greatly reduce the number of simulations required for size optimization and improve sampling efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of the method for automatically designing the size of an analog integrated circuit with high PVT robustness provided by the present invention.

[0053] Figure 2 It is the structure of the multi-task reinforcement learning agent used in the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described below by way of embodiments in conjunction with the accompanying drawings, but the scope of the present invention is not limited in any way.

[0055] The high PVT robustness size automation design provided in the present invention aims to design a set of optimization algorithms so that, given the circuit topology netlist, expected performance constraints, size design space and PVT variation space, the optimization algorithm can give a set of reasonable circuit sizes so that the circuit performance under all PVT angles can reach the expected performance.

[0056] The present invention adopts Bayesian optimization algorithm and multi-task reinforcement learning algorithm. The focus of the present invention is 1) proposing a sizing method that integrates Bayesian optimization and multi-task reinforcement learning and 2) proposing a method that uses the latest sampling cache to evaluate the sizing progress under each current PVT angle, and based on this, a dynamic PVT angle pruning method based on the latest sampling.

[0057] The present invention adopts the Bayesian optimization algorithm for initial sampling to provide initial sampling points and initial data sets for the high PVT robustness dimensional optimization process. In the initial sampling, the PVT environment is the default environment, that is, the Typical-Typical processing technology, 100% power supply voltage and 27°C ambient temperature. The objective function in the Bayesian optimization algorithm is the mapping from the dimensional design scheme X to the dimensional quality factor FoM. When the FoM reaches the expected value, the Bayesian optimization algorithm ends and the optimal initial sampling point X' for the current FoM is obtained, and N groups of dimensional design schemes including X' are randomly generated with X' as the center. At this time, the generated N groups of dimensional design schemes will be sampled under all k PVT angles to generate an initial data set.

[0058] When considering the impact of PVT changes on circuit performance, the circuit size design problem can be expressed as a constraint satisfaction problem. The constraint satisfaction problem does not seek the optimal value of a certain objective function, but is committed to finding a combination of input variables that can achieve all specified constraints. In the circuit size design problem, the input variables are the dimensions of each device, such as the channel length, channel width, resistance, capacitance value, etc. of the transistor, and the constraints are the various performance indicators that the circuit needs to achieve, including bandwidth, gain, noise, etc. When all performance indicators reach the target value under a certain size design scheme, the size design scheme is the input variable combination that satisfies the constraint satisfaction problem. The mathematical form of the size design problem is expressed as follows:

[0059] minimize 0,subject to F i (X|T j ) <C i ,i=1,2,…,m,j=1,2,…,k

[0060] where X∈D; X=X1,X2,…,X n ; D=D1,D2,…,D n ; T=T1,T2,…,T k ; C=C1,C2,…,C m . n is the number of independent size variables, m is the number of performance indicators, and k is the number of all PVT angles that need to be considered. X is an n-dimensional size vector, D is a total size design space with n elements, and each element represents the design space of an independent size variable. T represents the PVT angle, and C represents the m-dimensional performance indicator constraint vector. F is a nonlinear mapping from the size design scheme to the circuit performance indicator. For example, F i (X|T j ) means that when the dimension design dimension X is at T j The value of the i-th performance indicator after sampling at the PVT angle represented.

[0061] In order to describe the quality of the size design solution under a certain PVT angle, the present invention defines the quality factor FoM (Figure of Merit) to normalize multiple circuit performance indicators into a single indicator. The FoM definition is as follows:

[0062]

[0063]

[0064] Where FoM is the quality factor; r represents the distance between the normalized performance indicator and the normalized constraint; F i is the i-th performance index; C i is the target value of the i-th performance indicator; i is the serial number of the performance indicator; m is the number of types of performance indicators. For example, for an operational amplifier that only cares about power consumption, gain and bandwidth, the number of types of performance indicators is 3.

[0065] In the calculation process of FoM, the constraints are divided into upper bound constraints and lower bound constraints according to the nature of the constraints. The upper bound constraint refers to the constraint of a certain performance indicator being less than the target value, such as power consumption less than 1mW; the lower bound constraint refers to the constraint of a certain performance indicator being greater than the target value, such as gain greater than 100dB. When the constraint condition is not met, FoM is assigned a value of r, which expresses the distance between the normalized performance indicator and the normalized constraint. When there is a performance indicator that does not reach the expected value, r is a negative value; when all performance indicators reach the expected value, the calculation result of r is 0, and FoM is set to a positive value close to zero (here 0.2). If the FoM values ​​of a certain size design scheme X′ are positive under all PVT angles, then the scheme X′ is the desired size design scheme with high PVT robustness.

[0066] In summary, the core technology to be solved by the present invention is how to realize an efficient automatic size design method so as to find a size design solution X′ with positive FoM values ​​at all PVT angles with as few sampling times and time overhead as possible.

[0067] Figure 1 The figure shows the process of the automatic design method of the size of the high PVT robust analog integrated circuit provided by the present invention. Taking the size design of the folded cascode operational amplifier as an example, the various steps of the present invention are further explained.

[0068] A. Determine the integrated circuit size design task;

[0069] A1. Determine that the topology of the circuit is a folded cascode operational amplifier. There are 20 dimensional parameters to be adjusted, including 7 transistor channel lengths, 7 transistor channel widths, 4 transistor replication ratios, and two capacitor values. The PVT changes that need to be considered are: process deviations are {TT, SS, FF, SF, FS}, voltage conditions are {0.9Vref, Vref, 1.1Vref}, and temperature deviations are {-40℃, 27℃, 80℃}. After combining with each other, there are 45 PVT angles in total, and these 45 PVT angles are numbered from 1 to 45;

[0070] A2. Determine the performance indicators that the circuit needs to achieve, which are: 1) gain greater than 100dB, 2) unity gain bandwidth greater than 30MHz, 3) phase margin greater than 60 degrees, 4) power consumption less than 1mW, 5) output reference noise less than 30mV, 6) common mode rejection ratio greater than 100dB, 7) power supply rejection ratio greater than 100dB;

[0071] A3. Build the test circuit and obtain the circuit netlist file;

[0072] B. Perform initial sampling based on the trust region Bayesian optimization algorithm, and the implementation method is as follows;

[0073] B1. Under a single PVT angle (usually the PVT angle corresponding to a typical process, an ideal power supply and room temperature {Process = TT, Voltage = Vref, Temperature = 27°C}) a set of dimensional parameters that meet the performance index constraints required for all circuits under this PVT angle is obtained using the trust region Bayesian optimization algorithm;

[0074] B2. Select n sets of size parameters with the best quality factor under a single PVT angle found during the operation of the Bayesian optimization algorithm. n is generally set to 4 to 8 to take into account both sampling efficiency and convergence speed. Here, n=8 is selected;

[0075] B3. The 8 sets of dimensional parameters selected in B2 are sampled at all 45 PVT angles to generate an initial data set. At this time, the initial data set contains 45*8=360 data items, 8 data items at each PVT angle, and each data item is stored in a dictionary. The key values ​​of the dictionary are "dimensional parameters", "PVT angle", "performance index" and "quality factor".

[0076] C. Perform PVT robust circuit size optimization based on multi-task reinforcement learning; the implementation method is:

[0077] C1. Build data cache. Build a data cache for each PVT angle, and here a total of 45 data caches are built. Each data cache consists of two parts, namely 1) historical data cache, that is, the data of all simulations performed under this PVT angle, and 2) latest sampling cache, that is, the data of the most recent simulation performed under this PVT angle.

[0078] C2. Load initial data. Load the initial data set into the data cache constructed in C1. At this time, each data cache corresponds to a PVT angle and contains 8 data. Then randomly select one of the 8 data and load it into the latest sampling cache as the initialization of the latest sampling cache; at the same time, build the failed PVT set and initialize it to empty;

[0079] C3 builds a simplified PVT set. Read the data in the latest sampling cache, perform K-means clustering on the PVT angles according to the performance indicators of the most recent sampling recorded under each PVT angle, and select the PVT angle with the worst performance indicator in each category to be included in the simplified PVT set. Here, the clustering category is set to 6, and the clustering algorithm will divide the 45 PVT angles into 6 categories, and the simplified PVT set finally constructed will also contain 6 PVT angles.

[0080] C4. Adjust and streamline the PVT set.

[0081] Add the PVT angles in the failed PVT set to the simplified PVT set. If the number of PVT angles in the failed PVT set is m, then the number of PVT angles in the simplified PVT set is not less than 6 and not more than 6+m. The specific operation of constructing the failed PVT set is shown in step D3;

[0082] C5. Use a multi-task reinforcement learning agent to give the size parameter to be sampled in each iteration, and sample the size parameter under the PVT angle contained in the simplified PVT set;

[0083] In the specific implementation, the structure of the multi-task reinforcement learning agent used is composed of an Actor network and a Critic network, such as Figure 2As shown. The agent regards the size optimization under a PVT angle as a task. In the agent, the input of the Actor network is the PVT angle. Each PVT angle is represented by a 7-dimensional vector. The first five dimensions of the vector are the process angles represented by the one-hot code (such as [1, 0, 0, 0, 0] represents the TT process angle), and the last two dimensions of the vector represent the normalized power supply voltage and operating temperature. For example, 1.1 times Vref and 80℃ are the maximum power supply voltage and maximum operating temperature, respectively, which will be normalized to 1.0; 0.9 times Vref and -40℃ are the minimum power supply voltage and minimum operating temperature, which will be normalized to -1.0; and 1.0 times Vref and 27℃ operating temperature will be linearly mapped to the (-1, 1) interval according to the maximum and minimum values ​​of the power supply voltage and operating temperature. The output of the Actor network is the size parameter to be sampled. The input of the Critic network is the size parameter, and the output is the predicted circuit performance index of the size parameter under each PVT angle.

[0084] C6. Update the sampling results in C5 to the historical data cache and the latest sampling cache of the corresponding PVT angle;

[0085] C7. Read the latest sample cache corresponding to all PVT angles. If the performance indicators of the latest samples under each PVT angle meet the requirements, jump to step D1;

[0086] C8. Update the network weights of the reinforcement learning agent according to the data in the data cache, and jump to step C3. The pseudo code of the weight update process of the multi-task reinforcement learning agent used in the present invention is as follows:

[0087]

[0088] In the weight update process of the multi-task reinforcement learning agent, each round of weight update is divided into two stages. The first stage is to update the Critic network first, and the second stage is to update the Actor network using the updated Critic network. When updating the network, the update gradient direction of the network weight under each task is first calculated, and then the conflicting gradients are eliminated using the PCGrad (projecting conflicting gradients) algorithm. Among them, the agent regards the size optimization under a PVT angle as a task. The agent consists of an Actor network and a Critic network. Among them, the input of the Actor network is the PVT angle, each PVT angle is represented by a 7-dimensional vector, the first five dimensions of the vector are the process angles represented by the one-hot code (such as [1, 0, 0, 0, 0] represents the TT process angle), and the last two dimensions of the vector represent the normalized power supply voltage and operating temperature. The output of the Actor network is the size parameter to be sampled. The input of the Critic network is the size parameter, and the output is the quality factor of the circuit predicted by the size parameter under each PVT angle.

[0089] D. Conduct full PVT angle test, the specific implementation plan is:

[0090] D1. Sample the size parameters to be sampled given in the most recent step C5 at all PVT angles.

[0091] D2. If the performance indicators of each PVT angle sampling in D1 meet the requirements, the full PVT angle test is passed, and the sampling size parameters in the most recent D1 are the final output of the size automatic design tool, and the process ends;

[0092] D3. If the performance index under a certain PVT angle in D2 does not meet the requirements, then the PVT angle is added to the failed PVT set. The failed PVT set will not be initialized again during the entire size optimization process. That is, once a PVT angle appears in the failed PVT set, it will not be removed;

[0093] D4. Update the sampling results of the full PVT angle test to the latest sampling cache, and jump to step C3.

[0094] It should be noted that the purpose of publishing the embodiments is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A method for automatically designing the size of an analog integrated circuit with high PVT robustness, characterized in that: The Bayesian optimization algorithm is used to construct the initial data set, and the size optimization under a single PVT angle is performed. Then, the size optimization under multiple PVT angles is performed using a multi-task reinforcement learning algorithm. The steps are as follows: A. Determine the integrated circuit size design task; B. Performing initial sampling based on a Bayesian optimization algorithm, and generating an initial data set based on a size optimization method at a single PVT angle using the Bayesian algorithm; the initial data set includes multiple sets of size parameters with the best quality factors, and performance indicators and quality factors of the size parameters at all PVT angles; C. Perform PVT robust circuit size optimization based on multi-task reinforcement learning; including the following operations: C1. Build data cache; A data cache is constructed for each PVT angle, including: historical data cache and latest sampling cache; the data in the historical data cache is the data obtained from all simulations performed under the PVT angle; the data in the latest sampling cache is the data obtained from the most recent simulation performed under the PVT angle; C2. Load initial data, build failed PVT set and initialize it to empty; The initial data set is loaded into the historical data cache, and a set of data is randomly selected and loaded into the latest sampling cache; the elements in the constructed failed PVT set are the PVT corners that fail to meet the performance indicator requirements in the subsequent process; C3. Perform dynamic PVT corner pruning based on the latest sampling; including: Construct a simplified PVT set to reduce the PVT angles required for sampling in each iteration; Read the data in the latest sampling cache, cluster the PVT corners according to the performance indicators of the most recent sampling recorded under each PVT corner, and select the PVT corner with the worst performance indicator in each category to add to the simplified PVT set; then add the failed PVT set to the simplified PVT set; C4. Use a multi-task reinforcement learning agent to perform multiple rounds of iterations: sample the size parameters to be sampled under the PVT angles contained in the reduced PVT set; The input of the Actor network of the multi-task reinforcement learning agent is the PVT angle, and the output is the size parameter to be sampled; the input of the Critic network of the agent is the size parameter, and the output is the circuit performance index predicted by the size parameter at each PVT angle; C5. Update the sampling results in step C4 to the historical data cache and the latest sampling cache corresponding to the PVT angle; C6. Read the latest sample cache corresponding to all PVT angles. If the performance indicators of the latest samples under each PVT angle meet the requirements, jump to step D to perform full PVT angle test; C7. Update the network weights of the reinforcement learning agent according to the data in the data cache and jump to step C3; D. Perform full PVT corner testing, including: D1. Sampling the size parameters to be sampled given in step C4 at all PVT angles; D2. If the performance indicators of each PVT angle sampling in step D1 meet the requirements, the full PVT angle test is passed. The sampling size parameters in D1 are the final output of the automatic size design tool, and the operation ends; D3. If the performance index under a certain PVT angle in D1 does not meet the requirements, then add the PVT angle to the failed PVT set; D4. Update the sampling results of the full PVT angle test to the latest sampling cache and jump to step C3; Through the above steps, the automatic design of the size of the analog integrated circuit with high PVT robustness is achieved.

2. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 1, characterized in that: Step B performs initial sampling based on the Bayesian optimization algorithm, including: B1. Under a single PVT angle, a set of size parameters that meet the performance index constraints required for all circuits under this PVT angle is obtained using the trust region Bayesian optimization algorithm; B2. Select n sets of size parameters with the best quality factor FoM found during the Bayesian optimization algorithm operation; B3. Sample the n groups of size parameters selected in B2 at all PVT angles to generate an initial data set.

3. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 2, characterized in that: In step B2, the number of groups of size parameters having the optimal quality factor FoM is set to 4 to 8; The quality factor FoM normalizes multiple circuit performance indicators into a single indicator. FoM is defined as follows: Where FoM is the quality factor; r represents the distance between the normalized performance indicator and the normalized constraint; F i is the i-th performance index; C i is the target value of the i-th performance indicator; i is the serial number of the performance indicator; m is the number of types of performance indicators.

4. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 3, characterized in that: In step B3, each piece of data in the generated initial data set includes four key values, namely, size parameter, PVT angle, performance index and quality factor.

5. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 1, characterized in that: In step C4, the multi-task reinforcement learning agent takes size optimization under the PVT angle as a task; the agent structure includes an Actor network and a Critic network; each PVT angle of the Actor network input is represented by a 7-dimensional vector, the first five dimensions of the vector are the process angles represented by the one-hot code, and the last two dimensions of the vector represent the normalized power supply voltage and operating temperature.

6. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 1, characterized in that: During the size optimization process, the failed PVT set will not be initialized again, that is, the PVT corners appearing in the failed PVT set will not be removed.

7. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 1, characterized in that: In step C3, specifically, K-means clustering is performed on the PVT angle.

8. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 1, characterized in that: Step A determines the integrated circuit size design task including the following process: A1. Determine the circuit topology as a folded cascode operational amplifier, adjust the size parameters, and obtain all PVT corners; A2. Determine the performance indicators that the circuit needs to achieve, including: gain, unity gain bandwidth, phase margin, power consumption, output reference noise, common mode rejection ratio and power supply rejection ratio; A3. Build a test circuit and obtain the circuit netlist file.

9. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 8, characterized in that: In step A1, there are 20 dimensional parameters to be adjusted, namely, 7 transistor channel lengths, 7 transistor channel widths, 4 transistor replication ratios and two capacitor values; PVT changes include: process deviation is {TT, SS, FF, SF, FS}, voltage condition is {0.9Vref, Vref, 1.1Vref}, temperature deviation is {-40℃, 27℃, 80℃}, and after mutual combination, a total of 45 PVT angles are included, and the PVT angles are numbered; in step A2, the performance indicators that the circuit needs to achieve are determined as follows: gain greater than 100dB; unit gain bandwidth greater than 30MHz; phase margin greater than 60 degrees; power consumption less than 1mW; output reference noise less than 30mV; common mode rejection ratio greater than 100dB; power supply rejection ratio greater than 100dB.

10. The method for automatically designing the size of an analog integrated circuit with high PVT robustness as claimed in claim 9, characterized in that: Step C7 updates the network weights of the reinforcement learning agent according to the data in the data cache. Specifically, each round of weight update includes two stages. The first stage updates the Critic network first, and the second stage uses the updated Critic network to update the Actor network. When updating the network, first calculate the update gradient direction of the network weight under each task, and then eliminate the conflicting gradients.