WO2025124166A1 - 一种基于深度学习的模型仿真预测方法及系统 - Google Patents

一种基于深度学习的模型仿真预测方法及系统 Download PDF

Info

Publication number
WO2025124166A1
WO2025124166A1 PCT/CN2024/135598 CN2024135598W WO2025124166A1 WO 2025124166 A1 WO2025124166 A1 WO 2025124166A1 CN 2024135598 W CN2024135598 W CN 2024135598W WO 2025124166 A1 WO2025124166 A1 WO 2025124166A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
deep learning
parameters
multiple parameters
prediction
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/CN2024/135598
Other languages
English (en)
French (fr)
Inventor
李永胜
石凯
王可亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Primarius Tehcnologies Co Ltd
Original Assignee
Primarius Tehcnologies Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Primarius Tehcnologies Co Ltd filed Critical Primarius Tehcnologies Co Ltd
Publication of WO2025124166A1 publication Critical patent/WO2025124166A1/zh
Anticipated expiration legal-status Critical
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/32—Circuit design at the digital level
    • G06F30/33—Design verification, e.g. functional simulation or model checking
    • G06F30/3308—Design verification, e.g. functional simulation or model checking using simulation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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

Definitions

  • the present invention belongs to the technical field of chip design, and in particular relates to a model simulation prediction method and system based on deep learning.
  • Model simulation calculation is to input a set of model parameters and calculate the output, such as current and capacitance.
  • the traditional method is to use complex calculation formulas, such as the quasi-Newton method, to calculate.
  • the disadvantages are slow calculation speed, difficulty in obtaining the partial derivatives of output and parameters, difficulty in parallel calculation of multiple outputs corresponding to multiple sets of parameters, and inability to directly obtain the partial derivatives of the simulation to the parameters, so differential differentiation is necessary.
  • the purpose of the present invention is to provide a model simulation prediction method and system based on deep learning, which can accurately predict the resistance and capacitance data of the model, and train and predict the current and capacitance change curves with bias under different parameters through neural network training.
  • the technical solution of the present invention is: a model simulation prediction method based on deep learning, comprising the following steps: determining multiple parameters required for the model simulation, setting an adjustable boundary for each parameter; calling a nanospice simulator to obtain multiple sets of data of the multiple parameters, and normalizing the multiple parameters based on their respective adjustable boundaries; inputting the normalized data of the multiple parameters into a pre-configured deep learning model to train the deep learning model; obtaining the parameters of the current simulation prediction model and inputting them into the trained deep learning model for prediction to obtain the prediction result.
  • the machine learning model in this embodiment can be a deconvolutional neural network, a fully connected network, a GAN, an AE, etc.
  • the core of this technical solution is how to obtain parameters and how to pre-process the parameters on the basis of the above machine learning models, so that the model can be effectively trained in the end, and finally a machine learning model that can accurately predict the resistance and capacitance of the model (moudle) is obtained.
  • determining multiple parameters required for the model simulation and setting adjustable boundaries for each parameter further includes: determining the multiple parameters based on historical real engineering experience, including: k2, u0, ua, eu, uc, vsat, a1, a2, nfactor, eta0, etab, pclm, pdiblc, ndep; adopting standard average distribution random or logarithmic average distribution random to obtain multiple groups of values for the multiple parameters; replacing the corresponding values in the preset model card based on the multiple groups of values of the multiple parameters.
  • the method further includes: writing multiple groups of numerical data of the multiple parameters into the netlist file based on a pre-constructed netlist file, wherein the netlist file is pre-constructed according to actual measurement points of the corresponding model; dividing according to the measurement points, each size contains multiple groups of different curves.
  • normalizing the multiple parameters based on their respective adjustable boundaries further includes: when the standard average distribution random method is adopted to normalize the parameters, the normalization is performed based on the upper and lower limits of the adjustable boundaries of each parameter, and the processing result is between 0 and 1; when the random method of the average distribution after taking the logarithm is adopted to normalize the parameters, the logarithm of the parameter value is first taken, and the logarithm of the upper and lower limits of the adjustable boundaries of each parameter is also taken, and then the normalization is performed based on the logarithm of the value, the logarithm of the upper limit and the logarithm of the lower limit, and the processing result is between 0 and 1.
  • the method further comprises: after randomly selecting multiple groups of values of the multiple parameters using standard average distribution or logarithmically average distribution, performing preliminary screening based on set adjustable boundaries and physical facts to prevent the simulator from making mistakes.
  • the method before inputting the normalized data of multiple parameters into the pre-configured deep learning model to train the deep learning model, the method also includes: translating the simulation results of the nanospice simulator into a tfrecord format for storage through a parsing module to facilitate network training.
  • inputting the normalized data of multiple parameters into a pre-configured simulator to train the deep learning model further includes: using the fit method of the Model class in the tensorflow.keras library, with a ratio of 6:1 between the training set and the validation set, and the number of samples in each batch is 600; using Adam as the optimizer, with a learning rate of 1e-3, using current and capacitance as outputs, and using Mean Squared Logarithmic Error as the loss function.
  • the present invention also provides a model simulation prediction system based on deep learning, including: a parameter determination module, used to determine multiple parameters required for the model simulation, and set an adjustable boundary for each parameter; a preprocessing module, used to call the nanospice simulator to obtain multiple groups of data of the multiple parameters, and normalize the multiple parameters based on their respective adjustable boundaries; a model training module, used to input the data of the normalized multiple parameters into a pre-configured deep learning model to train the deep learning model; a result prediction module, used to obtain the parameters of the current simulation prediction model and input them into the trained deep learning model for prediction to obtain prediction results.
  • a parameter determination module used to determine multiple parameters required for the model simulation, and set an adjustable boundary for each parameter
  • a preprocessing module used to call the nanospice simulator to obtain multiple groups of data of the multiple parameters, and normalize the multiple parameters based on their respective adjustable boundaries
  • a model training module used to input the data of the normalized multiple parameters into a pre-configured deep learning model to train the deep learning model
  • the present invention also provides an electronic device, characterized in that it includes: a memory, the memory is used to store a processing program; and a processor, the processor implements any one of the above-mentioned deep learning-based model simulation prediction methods when executing the processing program.
  • the present invention also provides a readable storage medium, characterized in that a processing program is stored on the readable storage medium, and when the processing program is executed by a processor, any one of the above-mentioned deep learning-based model simulation prediction methods is implemented.
  • the present invention has the following advantages and positive effects compared with the prior art:
  • multiple groups of data of multiple parameters are randomly and automatically generated for preprocessing.
  • the preprocessed data and the resistance and capacitance simulation results obtained by calling the nanospice simulator are input into a deep learning model for training, thereby realizing simulation prediction of resistance and capacitance data based on the deep learning model, and improving the simulation prediction efficiency by training and predicting the current and capacitance change curves with bias under different parameters through neural network.
  • the present invention uses a nanospice simulator to define specific circuit conditions and specific measurement conditions through a netlist.
  • a simulation netlist a corresponding netlist file is constructed based on a set of real measurement points of a foundry, giving the machine learning model the physical meaning of the application scenario.
  • FIG1 is a flow chart of a model simulation prediction method based on deep learning of the present invention.
  • FIG2 is a schematic diagram of a deep learning neural network model in one embodiment of the present invention.
  • FIG3 is a schematic diagram of the prediction results of the model simulation based on deep learning of the present invention.
  • this embodiment provides a model simulation prediction method based on deep learning, comprising the following steps:
  • S100 Determine multiple parameters required for the model simulation, and set an adjustable boundary for each parameter
  • S200 calling the nanospice simulator to obtain multiple sets of data of the multiple parameters, and performing normalization processing on the multiple parameters based on their respective adjustable boundaries;
  • multiple sets of data of multiple parameters are randomly and automatically generated, and preprocessing is performed.
  • the data after preprocessing and the resistance and capacitance simulation results obtained by calling the nanospice simulator are input into the deep learning model for training, thereby realizing simulation prediction of resistance and capacitance data based on the deep learning model, and improving the simulation prediction efficiency.
  • an adjustable boundary is set for each parameter.
  • the parameter size is modified one by one according to a fixed step size, and simulation is performed to obtain the current and capacitance change curve with bias.
  • the parameters and curves are the data required for training.
  • determining multiple parameters required for the model simulation and setting adjustable boundaries for each parameter further includes: determining the multiple parameters based on historical real engineering experience, including: k2, u0, ua, eu, uc, vsat, a1, a2, nfactor, eta0, etab, pclm, pdiblc, ndep; adopting standard average distribution random or logarithmic average distribution random to obtain multiple groups of values for the multiple parameters; replacing the corresponding values in the preset model card based on the multiple groups of values of the multiple parameters.
  • a new model card can be obtained by replacing the corresponding parameter values in the pre-configured model card with the random values generated in the previous process. Since the nanospice simulator is used, a netlist is required to define the specific circuit conditions and specific measurement conditions.
  • the randomly generated parameters and the simulated resistance and capacitance results based on the nanospice simulator are written into a pre-configured netlist file.
  • the preferred embodiment of the present invention is based on tensorflow, so after the simulation corresponding to a model card is completed, a parsing module is used to translate the simulation results into tfrecord format for storage in order to train the network.
  • the CPU main frequency of the simulator is 3.6GHz.
  • normalizing the multiple parameters based on their respective adjustable boundaries further includes: when the standard average distribution random method is adopted to normalize the parameters, the normalization is performed based on the upper and lower limits of the adjustable boundaries of each parameter, and the processing result is between 0 and 1; when the random method of the average distribution after taking the logarithm is adopted to normalize the parameters, the logarithm of the parameter value is first taken, and the logarithm of the upper and lower limits of the adjustable boundaries of each parameter is also taken, and then the normalization is performed based on the logarithm of the value, the logarithm of the upper limit and the logarithm of the lower limit, and the processing result is between 0 and 1.
  • the technical solution of this embodiment normalizes each parameter, which can make the parameter processing of subsequent models more convenient, avoid the model from additionally processing complex data, improve the accuracy of data processing, and reduce the possibility of model failure.
  • the method further comprises: after randomly selecting multiple groups of values of the multiple parameters using standard average distribution or logarithmically average distribution, performing preliminary screening based on set adjustable boundaries and physical facts to prevent the simulator from making mistakes.
  • a netlist is needed to define the specific circuit conditions and specific measurement conditions. Because the combination of random parameters may violate physical facts, such as the effective length cannot be less than 0. Therefore, a preliminary screening step is required to skip the model cards that are bound to have problems based on some criteria to reduce simulator errors and reduce the simulation data generation time.
  • inputting the normalized data of multiple parameters into a pre-configured simulator to train the deep learning model further includes: using the fit method of the Model class in the tensorflow.keras library, with a ratio of 6:1 between the training set and the validation set, and the number of samples in each batch is 600; using Adam as the optimizer, with a learning rate of 1e-3, using current and capacitance as outputs, and using Mean Squared Logarithmic Error as the loss function.
  • a deep learning model Before training the model, a deep learning model must be built first, see Figure 2, which shows a schematic diagram of a deep learning neural network model in an embodiment.
  • the neural network takes 14 model parameters as input; the network adopts a deconvolution network, that is, it is composed of several fully connected layers and several deconvolution layers; then 6 curve data of 61 points are obtained by deconvolution calculation.
  • the training sample is 500,000 simulation curves.
  • the training adopts the fit method of the Model class in the tensorflow.keras library.
  • the number of samples is 450,000, the ratio of training set to verification set is 6:1, and the number of samples in each batch is 600.
  • Adam is used as the optimizer with a learning rate of 1e-3.
  • the loss function uses Mean Squared Logarithmic Error.
  • the average fitting error of the simulation curve is 1%.
  • the present invention also provides a model simulation prediction system based on deep learning, including: a parameter determination module, used to determine multiple parameters required for the model simulation, and set an adjustable boundary for each parameter; a preprocessing module, used to call the nanospice simulator to obtain multiple groups of data of the multiple parameters, and normalize the multiple parameters based on their respective adjustable boundaries; a model training module, used to input the data of the normalized multiple parameters into a pre-configured deep learning model to train the deep learning model; a result prediction module, used to obtain the parameters of the current simulation prediction model and input them into the trained deep learning model for prediction to obtain prediction results.
  • a parameter determination module used to determine multiple parameters required for the model simulation, and set an adjustable boundary for each parameter
  • a preprocessing module used to call the nanospice simulator to obtain multiple groups of data of the multiple parameters, and normalize the multiple parameters based on their respective adjustable boundaries
  • a model training module used to input the data of the normalized multiple parameters into a pre-configured deep learning model to train the deep learning model
  • multiple groups of data of multiple parameters are randomly and automatically generated for preprocessing, and the preprocessed data and the resistance and capacitance simulation results obtained by calling the nanospice simulator are input into the deep learning model for training, thereby realizing simulation prediction of resistance and capacitance data based on the deep learning model, and improving the simulation prediction efficiency by training and predicting the current and capacitance change curves with bias under different parameters through neural network.
  • the present invention also provides an electronic device, characterized in that it includes: a memory, the memory is used to store a processing program; and a processor, the processor implements any one of the above-mentioned deep learning-based model simulation prediction methods when executing the processing program.
  • the present invention also provides a readable storage medium, characterized in that a processing program is stored on the readable storage medium, and when the processing program is executed by a processor, any one of the above-mentioned deep learning-based model simulation prediction methods is implemented.
  • the model simulation prediction method based on deep learning is implemented in the form of program instructions and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the technical solution of this embodiment is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of software, and the computer software is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure.
  • the aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program code.
  • the identification content specifically executed by the above-described system and device can refer to the corresponding process in the aforementioned method embodiment.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Hardware Design (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Health & Medical Sciences (AREA)
  • Geometry (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开一种基于深度学习的模型仿真预测方法，包括以下步骤：确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。能够准确的对模型的电阻、电容数据进行预测，通过神经网络训练和预测不同参数下电流、电容随bias变化曲线，提高仿真预测效率。

Description

一种基于深度学习的模型仿真预测方法及系统 技术领域
本发明属于芯片设计技术领域，尤其涉及一种基于深度学习的模型仿真预测方法及系统。
背景技术
在芯片设计技术领域，针对复杂继承电路的模型参数计算一直都是一项核心工作，模型仿真计算即输入一组模型参数，计算得出如电流、电容值作为输出。传统方法是通过复杂计算公式，用如拟牛顿方法计算，缺点是计算速度慢、输出与参数的偏导数不易得出，多组参数对应的多输出并行计算困难，不能直接获取仿真对参数的偏导数，迫不得已差分求导。
随着科技的发展，集成电路技术得到了长足的进步，集成度不断提高，功耗不断降低。与此各种复杂的应用场景对集成电路的性能和功耗提出了更高的要求。传统的硬件设计方法已经无法满足这些需求，需要借助算法设计来进行优化仿真。
发明内容
为解决上述问题，本发明的目的是提供一种基于深度学习的模型仿真预测方法及系统，能够准确的对模型的电阻、电容数据进行预测，通过神经网络训练和预测不同参数下电流、电容随bias变化曲线。
本发明的技术方案为：一种基于深度学习的模型仿真预测方法，包括以下步骤：确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。优选的，本实施例中的所述机器学习模型可以是反卷积神经网络、全连接、GAN、AE等网络，本技术方案的核心在于在以上机器学习模型的基础上，如何获取参数以及如何对参数进行预处理，使得最终能够对模型进行有效的训练，最终获得一个能够准确对模型(moudle)的电阻、电容进行预测的机器学习模型。
优选的，确定所述模型仿真所需的多个参数，针对每个参数设定可调边界进一步包括：基于历史真实工程经验确定所述多个参数，包括：k2、u0、ua、eu、uc、vsat、a1、a2、nfactor、eta0、etab、pclm、pdiblc、ndep；采取标准平均分布随机或是取对数后的平均分布进行随机以得到所述多个参数的多组数值；基于所述多个参数的多组数值替换预设模型卡中的对应数值。
优选的，在对所述多个参数基于各自的可调边界进行归一化处理之前所述方法还包括：基于预先构建的网表文件将所述多个参数的多组数值数据写入所述网表文件中，其中所述网表文件预先根据对应模型的真实测点构建；依据测量点进行划分，每个尺寸均包含多组不同的曲线。
优选的，对所述多个参数基于各自的可调边界进行归一化处理进一步包括：当采取标准平均分布随机方法对参数进行归一化处理时，基于每个参数的所述可调边界的上限和下限进行归一化处理，处理结果处于0～1之间；当采取取对数后的平均分布进行随机方法对参数进行归一化处理时，先将参数值取对数、每个参数的所述可调边界的上限和下限也取对数，进而基于值的对数、上限的对数以及下限的对数进行归一化处理，处理结果处于0～1之间。
优选的，所述方法还包括：在采取标准平均分布随机或是取对数后的平均分布进行随机以得到所述多个参数的多组数值之后，基于设定的可调边界以及物理事实进行初步筛选以预防所述仿真器出错。
优选的，将归一化处理后的多个参数的数据输入预先配置的所述深度学习模型中以训练所述所述深度学习模型之前还包括：通过解析模块将nanospice仿真器仿真的结果转译成tfrecord格式用于存储，以便于训练网络。
优选的，将归一化处理后的多个参数的数据输入预先配置的仿真器中以训练所述深度学习模型进一步还包括：采用了tensorflow.keras库中的Model类的fit方法，训练集与验证集比例为6:1，每批的样本数量为600；采用Adam作为优化器，学习率为1e-3，以电流、电容作为输出，损失函数选用Mean Squared Logarithmic Error。
基于相同的构思，本发明还提供一种基于深度学习的模型仿真预测系统，包括：参数确定模块，用于确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；预处理模块，用于调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；模型训练模块，用于将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；结果预测模块，用于获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。
基于相同的构思，本发明还提供一种电子设备，其特征在于，包括：存储器，所述存储器用于存储处理程序；处理器，所述处理器执行所述处理程序时实现上述任意一项所述的基于深度学习的模型仿真预测方法。
基于相同的构思，本发明还提供一种可读存储介质，其特征在于，所述可读存储介质上存储有处理程序，所述处理程序被处理器执行时实现上述任意一项所述的基于深度学习的模型仿真预测方法。
本发明由于采用以上技术方案，使其与现有技术相比具有以下的优点和积极效果：
1、本发明的技术方案中，通过随机自动生成多个参数的多组数据，进行预处理，基于预处理之后的数据以及调用nanospice仿真器仿真得到的电阻、电容仿真结果输入深度学习模型对其进行训练，从而实现基于深度学习模型对电阻、电容数据进行仿真预测，通过神经网络训练和预测不同参数下电流、电容随bias变化曲线，提高仿真预测效率。
2、本发明中采用nanospice仿真器，通过网表来定义具体电路情况以及具体的测量情况。在生成仿真网表中时，根据一组代工厂的真实测量点构建相应的网表文件，为机器学习模型赋予了该应用场景之下的物理意义。
附图说明
下面结合附图对本发明的具体实施方式作进一步详细说明，其中：
图1为本发明基于深度学习的模型仿真预测方法流程图；
图2为本发明一个实施例中深度学习神经网络模型的示意图；
图3为本发明基于深度学习的模型仿真预测结果示意图。
具体实施方式
以下结合附图和具体实施例对本发明作进一步详细说明。根据下面说明和权利要求书，本发明的优点和特征将更清楚。需说明的是，附图均采用非常简化的形式且均使用非精准的比率，仅用以方便、明晰地辅助说明本发明实施例的目的。
需要说明，本发明实施例中所有方向性指示(诸如上、下、左、右、前、后……)仅用于解释在某一特定姿态(如附图所示)下各部件之间的相对位置关系、运动情况等，如果该特定姿态发生改变时，则该方向性指示也相应地随之改变。
第一实施例
如图1所示，本实施例提供一种基于深度学习的模型仿真预测方法，包括以下步骤：
S100：确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；
S200：调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；
S300：将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；
S400：获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。
本实施例的技术方案中，通过随机自动生成多个参数的多组数据，进行预处理，基于预处理之后的数据以及调用nanospice仿真器仿真得到的电阻、电容仿真结果输入深度学习模型对其进行训练，从而实现基于深度学习模型对电阻、电容数据进行仿真预测，提高仿真预测效率。在确定多个参数之后，针对每个参数设定可调边界，优选的，在本方案中按照固定步长逐个修改参数大小，仿真，得到电流、电容随bias变化曲线。参数和曲线即训练所需数据。
优选的，确定所述模型仿真所需的多个参数，针对每个参数设定可调边界进一步包括：基于历史真实工程经验确定所述多个参数，包括：k2、u0、ua、eu、uc、vsat、a1、a2、nfactor、eta0、etab、pclm、pdiblc、ndep；采取标准平均分布随机或是取对数后的平均分布进行随机以得到所述多个参数的多组数值；基于所述多个参数的多组数值替换预设模型卡中的对应数值。
结合本发明的应用场景，本实施例中所确定的参数包括k2、u0、ua、eu、uc、vsat、a1、a2、nfactor、eta0、etab、pclm、pdiblc、ndep，在此不对这些参数的含义进行冗余解释，以上参数均是半导体行业中常用的参数，模型(moudle)可以是bsim4、bsim6、bsimcmg等不同版本，以上所述参数来源于bsim4、bsim6、bsimcmg等不同版本。
举个例子，参见表1，为了获取模型仿真训练的参数，先随机获取参数值，为了更加贴近真实情况，采用了一组真实的参数范围，并根据工程经验选择了一组较为全局的参数，共计14个。因为不同参数的分布具有不同特点，简便起见，根据工程经验逐参数选择了采取标准平均分布随机或是取对数后的平均分布进行随机：
表1参数表

用上一过程中产生的随机值替换预先配置的模型卡中的对应参数值即可得到一个新的模型卡。由于采用nanospice仿真器，故而需要一个网表来定义具体电路情况以及具体的测量情况。
优选的，在对所述多个参数基于各自的可调边界进行归一化处理之前所述方法还包括：基于预先构建的网表文件将所述多个参数的多组数值数据写入所述网表文件中，其中所述网表文件预先根据对应模型的真实测点构建；依据测量点进行划分，每个尺寸均包含多组不同的曲线。
由于采用nanospice仿真器，故而需要一个网表来定义具体电路情况以及具体的测量情况。在生成仿真网表中时，根据一组代工厂的真实测量点构建相应的网表文件。选取9个不同尺寸的器件：
表2尺寸表
依据测量点进行划分，每个尺寸均包含7组不同的曲线：
表3测量点情况表

将随机生成的参数以及基于nanospice仿真器的仿真电阻、电容结果写入预先配置的网表文件中，优选的本发明是基于tensorflow完成，故而待一个模型卡对应的仿真全部完成后，借助一个解析模块，将仿真的结果转译成tfrecord格式用于存储，以便训练网络。仿真器运行CPU主频为3.6GHz。
优选的，对所述多个参数基于各自的可调边界进行归一化处理进一步包括：当采取标准平均分布随机方法对参数进行归一化处理时，基于每个参数的所述可调边界的上限和下限进行归一化处理，处理结果处于0～1之间；当采取取对数后的平均分布进行随机方法对参数进行归一化处理时，先将参数值取对数、每个参数的所述可调边界的上限和下限也取对数，进而基于值的对数、上限的对数以及下限的对数进行归一化处理，处理结果处于0～1之间。
本实施例的技术方案对每个参数进行归一化处理，能够使得后续模型的参数处理更加便捷，避免模型额外处理复杂的数据，提高数据处理的准确性，降低模型出现故障的可能性。
优选的，所述方法还包括：在采取标准平均分布随机或是取对数后的平均分布进行随机以得到所述多个参数的多组数值之后，基于设定的可调边界以及物理事实进行初步筛选以预防所述仿真器出错。
由于采用nanospice仿真器，故而需要一个网表来定义具体电路情况以及具体的测量情况。因为随机参数组合起来可能会违反物理事实，比如有效长度不能小于0。故而需要进行一步初筛，依据一些判据，跳过必然有问题的模型卡，以减少仿真器出错，降低仿真数据生成时间。
优选的，将归一化处理后的多个参数的数据输入预先配置的所述深度学习模型中以训练所述深度学习模型之前还包括：通过解析模块将nanospice仿真器仿真的结果转译成tfrecord格式用于存储，以便于训练网络。
优选的，将归一化处理后的多个参数的数据输入预先配置的仿真器中以训练所述深度学习模型进一步还包括：采用了tensorflow.keras库中的Model类的fit方法，训练集与验证集比例为6:1，每批的样本数量为600；采用Adam作为优化器，学习率为1e-3，以电流、电容作为输出，损失函数选用Mean Squared Logarithmic Error。
优选的，在对模型进行训练之前，先要进行深度学习模型的搭建，参见图2，示出了一个实施例中深度学习神经网络模型的示意图。该神经网络以14个模型参数为输入；网络采用反卷积网络，也即由几层全连接层和几层反卷积层构成；然后通过反卷积计算得到6条61个点的曲线数据。训练样本为50万仿真曲线。训练采用了tensorflow.keras库中的Model类的fit方法。样本数量为45万，训练集与验证集比例为6:1，每批的样本数量为600。采用Adam作为优化器，学习率为1e-3。以电流电容作为输出，损失函数选用Mean Squared Logarithmic Error。经过训练，参见图3，对于仿真曲线的拟合平均误差在1％。
第二实施例
基于相同的构思，本发明还提供一种基于深度学习的模型仿真预测系统，包括：参数确定模块，用于确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；预处理模块，用于调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；模型训练模块，用于将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；结果预测模块，用于获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。
本实施例的技术方案中，通过随机自动生成多个参数的多组数据，进行预处理，基于预处理之后的数据以及调用nanospice仿真器仿真得到的电阻、电容仿真结果输入深度学习模型对其进行训练，从而实现基于深度学习模型对电阻、电容数据进行仿真预测，通过神经网络训练和预测不同参数下电流、电容随bias变化曲线，提高仿真预测效率。
第三实施例
基于相同的构思，本发明还提供一种电子设备，其特征在于，包括：存储器，所述存储器用于存储处理程序；处理器，所述处理器执行所述处理程序时实现上述任意一项所述的基于深度学习的模型仿真预测方法。
基于相同的构思，本发明还提供一种可读存储介质，其特征在于，所述可读存储介质上存储有处理程序，所述处理程序被处理器执行时实现上述任意一项所述的基于深度学习的模型仿真预测方法。
基于深度学习的模型仿真预测方法如果以程序指令的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本实施例的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件的形式体现出来，该计算机软件存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本公开各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-only memory，ROM)、随机存取存储器(Random access memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
所属领域的技术人员可以清楚地了解到，为描述得方便和简洁，上述描述的系统及设备的具体执行的识别内容，可以参考前述方法实施例中的对应过程。
上面结合附图对本发明的实施方式作了详细说明，但是本发明并不限于上述实施方式。即使对本发明作出各种变化，倘若这些变化属于本发明权利要求及其等同技术的范围之内，则仍落入在本发明的保护范围之中。

Claims (10)

  1. 一种基于深度学习的模型仿真预测方法，其特征在于，包括以下步骤：
    确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；
    调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；
    将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；
    获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。
  2. 根据权利要求1所述的基于深度学习的模型仿真预测方法，其特征在于，确定所述模型仿真所需的多个参数，针对每个参数设定可调边界进一步包括：
    基于历史真实工程经验确定所述多个参数，包括：k2、u0、ua、eu、uc、vsat、a1、a2、nfactor、eta0、etab、pclm、pdiblc、ndep；
    采取标准平均分布随机或是取对数后的平均分布进行随机以得到所述多个参数的多组数值；
    基于所述多个参数的多组数值替换预设模型卡中的对应数值。
  3. 根据权利要求2所述的基于深度学习的模型仿真预测方法，其特征在于，在对所述多个参数基于各自的可调边界进行归一化处理之前所述方法还包括：
    基于预先构建的网表文件将所述多个参数的多组数值数据写入所述网表文件中，其中所述网表文件预先根据对应模型的真实测点构建；
    依据测量点进行划分，每个尺寸均包含多组不同的曲线。
  4. 根据权利要求1所述的基于深度学习的模型仿真预测方法，其特征在于，对所述多个参数基于各自的可调边界进行归一化处理进一步包括：
    当采取标准平均分布随机方法对参数进行归一化处理时，基于每个参数的所述可调边界的上限和下限进行归一化处理，处理结果处于0～1之间；
    当采取取对数后的平均分布进行随机方法对参数进行归一化处理时，先将参数值取对数、每个参数的所述可调边界的上限和下限也取对数，进而基于值的对数、上限的对数以及下限的对数进行归一化处理，处理结果处于0～1之间。
  5. 根据权利要求2所述的基于深度学习的模型仿真预测方法，其特征在于，所述方法还包括：在采取标准平均分布随机或是取对数后的平均分布进行随机以得到所述多个参数的多组数值之后，基于设定的可调边界以及物理事实进行初步筛选以预防所述仿真器出错。
  6. 根据权利要求5所述的基于深度学习的模型仿真预测方法，其特征在于，将归一化处理后的多个参数的数据输入预先配置的所述深度学习模型中以训练所述所述深度学习模型之前还包括：
    通过解析模块将nanospice仿真器仿真的结果转译成tfrecord格式用于存储，以便于训练网络。
  7. 根据权利要求6所述的基于深度学习的模型仿真预测方法，其特征在于，将归一化处理后的多个参数的数据输入预先配置的仿真器中以训练所述深度学习模型进一步还包括：
    采用了tensorflow.keras库中的Model类的fit方法，训练集与验证集比例为6:1，每批的样本数量为600；
    采用Adam作为优化器，学习率为1e-3，以电流、电容作为输出，损失函数选用Mean Squared Logarithmic Error。
  8. 一种基于深度学习的模型仿真预测系统，其特征在于，包括：
    参数确定模块，用于确定所述模型仿真所需的多个参数，针对每个参数设定可调边界；
    预处理模块，用于调用nanospice仿真器获取所述多个参数的多组数据，对所述多个参数基于各自的可调边界进行归一化处理；
    模型训练模块，用于将归一化处理后的多个参数的数据输入预先配置的深度学习模型中以训练所述深度学习模型；
    结果预测模块，用于获取当前需要仿真预测模型的参数输入训练之后的所述深度学习模型进行预测以获取预测结果。
  9. 一种电子设备，其特征在于，包括：
    存储器，所述存储器用于存储处理程序；
    处理器，所述处理器执行所述处理程序时实现如权利要求1至7中任意一项所述的基于深度学习的模型仿真预测方法。
  10. 一种可读存储介质，其特征在于，所述可读存储介质上存储有处理程序，所述处理程序被处理器执行时实现如权利要求1至7中任意一项所述的基于深度学习的模型仿真预测方法。
PCT/CN2024/135598 2023-12-13 2024-11-29 一种基于深度学习的模型仿真预测方法及系统 Pending WO2025124166A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202311710768.3A CN117807929A (zh) 2023-12-13 2023-12-13 一种基于深度学习的模型仿真预测方法及系统
CN202311710768.3 2023-12-13

Publications (1)

Publication Number Publication Date
WO2025124166A1 true WO2025124166A1 (zh) 2025-06-19

Family

ID=90428833

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/135598 Pending WO2025124166A1 (zh) 2023-12-13 2024-11-29 一种基于深度学习的模型仿真预测方法及系统

Country Status (2)

Country Link
CN (1) CN117807929A (zh)
WO (1) WO2025124166A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120850061A (zh) * 2025-09-25 2025-10-28 易事特(江苏)储能科技有限公司 一种软包锂电池的铝塑膜壳成型加工检测方法及装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117807929A (zh) * 2023-12-13 2024-04-02 上海概伦电子股份有限公司 一种基于深度学习的模型仿真预测方法及系统
CN118886323B (zh) * 2024-07-26 2025-11-21 杭州法动科技有限公司 基于电磁耦合深度学习模型的射频器件建模方法及装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114997092A (zh) * 2022-06-16 2022-09-02 南京邮电大学 基于机器学习和模型的半导体器件电学特性仿真方法
CN115983166A (zh) * 2021-10-13 2023-04-18 三星电子株式会社 预测半导体器件的特性的方法和执行该方法的计算装置
US20230125401A1 (en) * 2021-10-22 2023-04-27 Samsung Electronics Co., Ltd. Method of predicting characteristic of semiconductor device and computing device performing the same
CN116151174A (zh) * 2023-04-14 2023-05-23 四川省华盾防务科技股份有限公司 一种通用器件模型优化方法及系统
CN117093847A (zh) * 2023-06-07 2023-11-21 上海概伦电子股份有限公司 基于高斯过程回归的统计延时特性计算方法及系统
CN117807929A (zh) * 2023-12-13 2024-04-02 上海概伦电子股份有限公司 一种基于深度学习的模型仿真预测方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115983166A (zh) * 2021-10-13 2023-04-18 三星电子株式会社 预测半导体器件的特性的方法和执行该方法的计算装置
US20230125401A1 (en) * 2021-10-22 2023-04-27 Samsung Electronics Co., Ltd. Method of predicting characteristic of semiconductor device and computing device performing the same
CN114997092A (zh) * 2022-06-16 2022-09-02 南京邮电大学 基于机器学习和模型的半导体器件电学特性仿真方法
CN116151174A (zh) * 2023-04-14 2023-05-23 四川省华盾防务科技股份有限公司 一种通用器件模型优化方法及系统
CN117093847A (zh) * 2023-06-07 2023-11-21 上海概伦电子股份有限公司 基于高斯过程回归的统计延时特性计算方法及系统
CN117807929A (zh) * 2023-12-13 2024-04-02 上海概伦电子股份有限公司 一种基于深度学习的模型仿真预测方法及系统

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120850061A (zh) * 2025-09-25 2025-10-28 易事特(江苏)储能科技有限公司 一种软包锂电池的铝塑膜壳成型加工检测方法及装置

Also Published As

Publication number Publication date
CN117807929A (zh) 2024-04-02

Similar Documents

Publication Publication Date Title
WO2025124166A1 (zh) 一种基于深度学习的模型仿真预测方法及系统
KR20180041174A (ko) 위험 평가 방법 및 시스템
CN109710948A (zh) 机器翻译引擎推荐方法及装置
CN108681750A (zh) Gbdt模型的特征解释方法和装置
CN112100374B (zh) 文本聚类方法、装置、电子设备及存储介质
CN116561542A (zh) 模型的优化训练系统、方法以及相关装置
WO2023159756A1 (zh) 价格数据的处理方法和装置、电子设备、存储介质
CN111506798A (zh) 用户筛选方法、装置、设备及存储介质
CN118917323B (zh) 一种结合大模型的金融文档检测方法及系统
CN114168318B (zh) 存储释放模型的训练方法、存储释放方法及设备
CN115526498A (zh) 一种服务评估的方法、装置及介质
CN113836899A (zh) 一种网页识别方法、装置及电子设备和存储介质
CN114492742A (zh) 神经网络结构搜索、模型发布方法、电子设备和存储介质
CN112419098A (zh) 基于安全信息熵的电网安全稳定仿真样本筛选扩充方法
CN114676820A (zh) 网络模型的调优方法、装置、电子设备及存储介质
CN120104173A (zh) 代码注释信息处理方法、装置、计算机设备和存储介质
CN111966978B (zh) 异常用户的确定方法、电子设备和存储介质
CN115237355B (zh) 一种基于闪存数据类识别的定向交换方法及系统
CN119337704A (zh) 基于线下事件激励模型的线上观点演化仿真方法和装置
WO2025035626A1 (zh) 版图质量检测工具的参数确定方法、装置、设备及介质
CN117786062A (zh) 一种基于改进信息检索算法的工艺知识规范化方法及系统
CN119559029A (zh) 图像处理方法、装置、设备及存储介质
CN115203556A (zh) 一种评分预测模型训练方法、装置、电子设备及存储介质
CN114444761A (zh) 一种风险预测方法、装置及计算机设备
CN118898259B (zh) 一种基于预训练模型的汉藏语翻译装置及藏语翻译方法

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: 24902588

Country of ref document: EP

Kind code of ref document: A1