Circuit aging simulation method based on machine learning

Abstract

The invention discloses a circuit aging simulation method based on machine learning, and belongs to the technical field of circuit aging simulation, and the method comprises the steps: selecting model parameters reflecting the aging condition of a to-be-tested circuit based on an aging model of a circuit device; the model parameters are scanned through an SPICE simulator, and simulated electrical parameters are obtained; mapping the simulated electrical parameters to model parameters by using the trained neural network to obtain specific numerical values of the model parameters; and inputting the specific numerical values of the model parameters into an SPICE simulator for aging simulation to obtain a current aging simulation result of the circuit to be tested. According to the method, the device model parameters are updated in real time according to the actual aging condition by using the aging model of the device, a circuit aging simulation tool can be helped to realize high-precision aging simulation prediction based on machine learning multi-parameter mapping, and the obtained simulation result can flexibly reflect the actual degradation condition of the device.

Description

Technical Field

[0001] The present invention belongs to the technical field of circuit aging simulation, and particularly relates to a circuit aging simulation method based on machine learning. Background Art

[0002] As the integrated circuit process nodes enter the nanoscale, the reduction of device size has led to a significant exacerbation of reliability problems. Non-ideal effects such as hot carrier injection (HCI), bias temperature instability (BTI), and self-heating effect (SHE) can cause key parameters such as device threshold voltage, mobility, and leakage current to degrade over time. This degradation can cause the circuit performance (such as timing, power consumption, noise margin) to deviate from the design expectations after long-term operation, and even lead to functional failures. Therefore, it is crucial to perform aging simulation at the circuit design stage to predict the performance changes during the life cycle, so as to take compensation measures in a timely manner (such as redundant design, dynamic voltage regulation).

[0003] Traditional aging simulation uses SPICE (Simulation Program with Integrated Circuit Emphasis) simulation tools to analyze the steady-state, transient, and long-term reliability degradation of circuits. The tool combines the aging models of devices, applies stress, calculates the electrical performance degradation of devices under transient simulation conditions, extrapolates the degradation to the actual aging time based on the calculation results, and updates the circuit netlist for secondary simulation to evaluate the long-term performance. However, in the entire simulation process, the model parameters of the devices are regarded as fixed values, and only the external degradation amount is superimposed on the original parameters, which is equivalent to defaulting that the model parameters of the devices themselves have not changed, which does not conform to the actual situation. Therefore, after extrapolating to a longer time, the results of traditional aging simulation will deviate from the actual situation.

[0004] Currently, the mainstream aging simulation schemes can be divided into two categories. One is the simulation based on static aging models. In this type of method, during the simulation process, the device model parameters are regarded as fixed values, and only the external degradation amounts are superimposed on the original parameters. For example, in the Cadence RelXpert tool, the degradation amounts directly correct the SPICE model parameters through look-up tables or formulas, while the internal non-linear relationships of the model are not dynamically updated. The other is the improvement of dynamic aging models. In this type of method, some research attempts to embed the aging effect into the physical equations of the device model and achieve dynamic parameter updates by introducing time-related degradation variables. However, this type of method still relies on preset aging model parameters and cannot adjust the model parameters in real time according to the actual degradation situation (such as non-linear accumulation under different stress histories).

[0005] In summary, the existing circuit aging simulation methods have great limitations in parameter selection. Secondly, the model parameters of the device itself are fixed values and cannot be adjusted in real time according to the actual degradation situation, resulting in poor simulation accuracy and inability to flexibly reflect the actual degradation situation of the device. Summary of the Invention

[0006] To solve the above problems existing in the prior art, the present invention provides a circuit aging simulation method based on machine learning. The technical problems to be solved by the present invention are achieved through the following technical solutions:

[0007] A circuit aging simulation method based on machine learning, comprising:

[0008] Select model parameters reflecting the aging situation of the circuit to be measured based on the aging model of the circuit device;

[0009] Use a SPICE simulator to scan the model parameters to obtain the electrical parameters after simulation;

[0010] Use the trained neural network to map the electrical parameters after simulation to the model parameters to obtain the specific values of the model parameters;

[0011] Input the specific values of the model parameters into the SPICE simulator for aging simulation to obtain the current aging simulation results of the circuit to be measured.

[0012] Advantages of the present invention:

[0013] A circuit aging simulation method based on machine learning provided by the present invention first selects model parameters reflecting the aging situation of a circuit to be measured based on the aging models of circuit devices, then uses a SPICE simulator to scan the model parameters to obtain electrical parameters after simulation; then uses a trained neural network to map the electrical parameters after simulation to the model parameters to obtain specific values of the model parameters; finally, inputs the specific values of the model parameters into the SPICE simulator for aging simulation to obtain the current aging simulation result of the circuit to be measured. On the one hand, starting from the aging model, this method screens different model parameters that can best reflect the circuit aging situation based on the self-degradation of devices. On the other hand, based on machine learning technology, it uses a pre-trained neural network to map the electrical performance degradation caused by device aging to the model parameters related to device aging, realizing the real-time update of device model parameters according to the actual aging situation using the device aging model, and can help circuit aging simulation tools achieve high-precision aging simulation prediction based on machine learning multi-parameter mapping. By accurately simulating and predicting the degradation of devices, the simulation results can flexibly reflect the actual degradation of devices.

[0014] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0015] Figure 1 is a flowchart of a circuit aging simulation method based on machine learning provided by an embodiment of the present invention;

[0016] Figure 2 is a schematic diagram of the effect of the LASSO algorithm provided by an embodiment of the present invention;

[0017] Figure 3 is a schematic diagram of the relationship between parameter scanning and mapping provided by an embodiment of the present invention.

[0018] Figure 4 is an architecture diagram of a parameter mapping neural network model provided by an embodiment of the present invention;

[0019] Figure 5 is another architecture diagram of a parameter mapping neural network model provided by an embodiment of the present invention;

[0020] Figure 6 is a flowchart of integrated circuit reliability simulation provided by an embodiment of the present invention. Detailed Embodiments

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] The present invention provides a circuit aging simulation method based on machine learning, mainly starting from the aging model of circuit devices, to the selection of aging-related parameters, and then mapping electrical parameters to model parameters, so as to help the circuit aging simulation complete highly accurate circuit aging prediction.

[0023] Specifically, please refer to Figure 1 , Figure 1 which is a schematic flow chart of a circuit aging simulation method based on machine learning provided by an embodiment of the present invention. The method mainly includes the following steps:

[0024] Step 1: Select model parameters reflecting the aging situation of the circuit to be measured based on the aging model of circuit devices.

[0025] 11) Select multiple electrical parameters reflecting the aging situation of the circuit to be measured according to the aging model of circuit devices.

[0026] Specifically, for the current circuit to be measured, that is, the circuit to be simulated and evaluated, the electrical parameters should be selected as those parameter combinations that can best reflect the circuit degradation situation, which is usually determined by the aging model of the device. Different aging models will bring different degradation effects. For example, SHE will cause changes in the device temperature, and the HCI and BTI effects mainly affect the threshold voltage and leakage current of the device, etc. Therefore, in combination with the aging model, the selection of electrical parameters includes but is not limited to temperature T, threshold voltage change ΔV th , leakage current change ΔId l and ΔId s etc.

[0027] 12) Use the SPICE simulator to perform a sweep simulation on the selected electrical parameters, and screen out several aging-sensitive model parameters according to the simulation results.

[0028] First, after determining the electrical parameters reflected by the aging model, use the SPICE simulator to perform a sweep simulation on the selected electrical parameters, and record the changes in the electrical degradation-related indicators corresponding to different parameters to obtain the simulation results.

[0029] Then, based on the simulation results, perform feature selection optimization, and finally screen out a set of key aging-sensitive parameters, including but not limited to temperature T, threshold voltage V th0 , electron mobility U 0, electron saturation rate V sat etc.

[0030] Optionally, as an implementation of parameter screening, in this embodiment, based on the simulation results, the LASSO (Least Absolute Shrinkage and Selection Operator) algorithm can be used for parameter screening to screen out several aging-sensitive model parameters.

[0031] Please refer to Figure 2 , Figure 2 which is a schematic diagram of the effect of the LASSO algorithm provided by the embodiment of the present invention. Figure 2 In [the figure], the abscissa represents different parameters, and the ordinate represents the correlation coefficient between each parameter and the linear region drain current idlin. It can be seen that different parameters have different correlations. The larger the value, the higher the correlation and the more important the parameter. By discarding the parameters with low correlations, parameter screening can be achieved. This method can simply and efficiently sort the parameter importance and select appropriate parameters for subsequent parameter mapping.

[0032] Optionally, as another implementation of parameter screening, in this embodiment, the elastic net regularization algorithm can also be used to establish a sparse regression model of parameter - performance, so as to screen out several aging-sensitive model parameters.

[0033] Specifically, in establishing the sparse regression model of parameter - performance, the selection intensity is controlled by setting the λ parameter, and the Pearson correlation coefficient (|ρ|≥0.7) and variance inflation factor (VIF<5) of each parameter are calculated. Finally, the parameters are screened out. The elastic net algorithm combines the advantages of ridge regression and LASSO and has better effects on high-dimensional data.

[0034] The specific implementation processes of the LASSO algorithm and the elastic net regularization algorithm can refer to the existing related technologies, and this embodiment will not elaborate here. In specific applications, either of the two algorithms can be selected according to the actual situation to achieve parameter screening.

[0035] Step 2: Use the SPICE simulator to scan the model parameters to obtain the electrical parameters after simulation.

[0036] First, the scanning range of the model parameters needs to be determined.

[0037] Optionally, as an implementation, after determining the model parameters in step 1, the scanning range of each parameter can be determined according to the nature of the parameters themselves in the semiconductor and experience.

[0038] In another embodiment of the present invention, the scanning range can also be determined by the 2σ principle.

[0039] Specifically, for each selected parameter \(X_i\), set the scanning range as \([X_{mean}-2\sigma, X_{mean}+2\sigma]\), where \(\sigma\) is obtained through Monte Carlo analysis.

[0040] Then, based on the determined scanning range, use the SPICE simulator to scan the model parameters to obtain the electrical parameters after simulation.

[0041] Exemplarily, taking the BTI effect of a BSIMCMG (Berkeley Short-Channel IGFET Model–Common Multi-Gate) model device as an example, through SPICE, for different specifications of \(V\) th0 、\(U\) 0 、\(V\) sat perform IV curve characteristic simulation, and record the threshold voltage \(V\) th 、saturation drain current \(I_d\) s 、turn-on drain current \(I_d\) l , complete the acquisition of a three-input three-output data for subsequent work.

[0042] Step 3: Use the trained neural network to map the electrical parameters after simulation to the model parameters to obtain the specific values of the model parameters.

[0043] Specifically, if the SPICE simulation is regarded as the calculation of a function, the process of parameter mapping is equivalent to the inverse function of this function, which is a complex and not necessarily achievable proposition. Traditional matrix mapping does not have the complexity of the neural network and cannot achieve accurate mapping for some parameters, resulting in too large errors, which is contrary to the idea of using parameter mapping to improve the prediction accuracy. Therefore, in this embodiment, a neural network model is introduced for parameter mapping.

[0044] It can be understood that before performing the parameter mapping in Step 3, it is necessary to first construct and train a neural network to obtain a trained neural network for parameter mapping.

[0045] In this embodiment, the trained neural network is obtained in the following manner:

[0046] A) Use the SPICE simulator to scan the model parameters of different circuits to obtain the electrical parameters after simulation as the training data set.

[0047] Specifically, first obtain a large number of model parameters of different circuits according to the method in Step 1. Then, process the obtained model parameters according to the method in Step 2 to correspondingly obtain the electrical parameters. Finally, collect all types of electrical parameters obtained as the training data set.

[0048] In the process of constructing the training data set in this embodiment, in step 2, the parameter scanning range determination process combining Monte Carlo analysis and the 2σ principle can improve the effectiveness of the data used for model training, thereby improving the speed and accuracy of subsequent network training.

[0049] It can be understood that the output in step 2 is used as the input of the mapping, and the input in step 2 is used as the output of the mapping. Please refer to Figure 3 , Figure 3 which is a schematic diagram of the relationship between parameter scanning and mapping provided by the embodiment of the present invention.

[0050] B) Construct a fully connected neural network model, and use the training data set to train the fully connected neural network model to obtain a trained neural network.

[0051] Optionally, as an implementation method, this embodiment can construct a complex fully connected neural network model and train all parameters in the same network model. The specific process is as follows in steps S101 - S102.

[0052] S101. Construct a fully connected neural network model.

[0053] Specifically, the fully connected neural network model constructed in this embodiment can adopt a general fully connected neural network structure. The detailed structure can be implemented with reference to the prior art and will not be specifically introduced here.

[0054] S102. Use all types of electrical parameters in the training data set to train the fully connected neural network model to obtain a trained neural network.

[0055] Please refer to Figure 4 , Figure 4 which is an architecture diagram of a parameter mapping neural network model provided by the embodiment of the present invention. This network architecture includes a fully connected neural network model. By training this model with all parameters in the training data set, a neural network that can be used to process threshold voltage V th , leakage current Id l and Id s and other multiple parameter mappings can be obtained.

[0056] Adopting the above network architecture will make the model complexity increase exponentially with the increase of the parameter dimension. And because the influence of different parameters on the result is different, some parameter features with weak effects may be masked, so the model is difficult to train. The advantage is that the coupling relationship between different parameters may be captured.

[0057] Further, in another embodiment of the present invention, fully connected neural networks can be constructed for different model parameters respectively and trained to obtain a parameter mapping model including multiple fully connected neural networks. The specific process is as follows in steps S201 - S202.

[0058] S201. For different types of electrical parameters, construct fully connected neural network models respectively.

[0059] Specifically, for different types of electrical parameters, the same or similar fully connected neural network models can be constructed to process the corresponding type of data respectively.

[0060] S202. Use different types of electrical parameters in the training dataset to train the corresponding fully connected neural network models respectively to obtain trained neural networks.

[0061] Please refer to Figure 5 , Figure 5 which is another architecture diagram of the parameter mapping neural network model provided by the embodiment of the present invention. Exemplarily, this network architecture contains three fully connected neural network models, which are trained with V th data, Id l data, and Id s data in the training dataset respectively, and correspondingly, neural networks for processing the threshold voltage V th , the turn - on leakage current Id l , and the saturation leakage current Id s can be obtained.

[0062] Adopting the above - mentioned network architecture, the complexity of the model will be greatly reduced, the model training is simple, and the prediction accuracy for a single parameter may be relatively high, but the coupling effect between parameters cannot be considered.

[0063] In addition, compared with a single model, the network structure with multi - parameter separate training can locate the advantages and disadvantages of the parameter mapping results more quickly, and new model training optimization can also be carried out for a single parameter.

[0064] In practical applications, the neural network obtained by steps S101 - S102 or steps S201 - S202 can be flexibly selected according to comprehensive considerations of parameter categories, quantities, training difficulties, and training accuracies, so as to achieve a high - precision multi - parameter mapping function.

[0065] After obtaining the trained neural network, the simulated electrical parameters can be mapped to the model parameters to obtain the specific values of the model parameters, thereby obtaining the changes in the device's own model parameters.

[0066] Based on the idea of machine learning, this embodiment introduces a neural network for parameter mapping, which can not only reduce the workload of engineers during training but also obtain high-precision parameter mapping results.

[0067] This embodiment transforms the forward characteristic simulation of the SPICE simulator into an inverse parameter mapping problem, constructs a fully connected neural network to realize the mapping function from electrical parameters degradation to model parameter adjustment, achieves high-precision parameter mapping, and at the same time realizes the real-time update of the device model parameters according to the actual aging situation using the aging model of the device.

[0068] It should be noted that during the training process of the fully connected neural network model, it also includes:

[0069] Segment the electrical parameters of different types in the training dataset according to the degradation amplitude, and use the data of each segment to train the fully connected neural network model respectively to obtain a segmented neural network.

[0070] Specifically, due to the complexity of different devices, a simple single model may not be able to accurately predict the actual degradation caused by large-scale parameter changes. Based on this, this embodiment introduces a segmented model. During the network training stage, the constructed fully connected neural network can be trained using segmented training sets respectively to obtain a neural network model for processing data at different stages. For example, train one model when the threshold voltage degrades slightly and another model when it degrades significantly. In this way, when performing parameter mapping, use one trained model when the threshold voltage degrades slightly and another trained model when the threshold voltage degrades significantly.

[0071] This embodiment realizes a dynamic model switching mechanism based on the degradation degree by introducing a segmented model. Set segmented intervals for key parameters such as the threshold voltage, and use different data to train models at different stages. Combining multiple models can always ensure the accuracy of parameter mapping measurement and improve the flexibility of parameter mapping.

[0072] Step 4: Input the specific values of the model parameters into the SPICE simulator for aging simulation to obtain the current aging simulation results of the circuit under test.

[0073] Specifically, the model parameters can be passed to the simulator SPICE through the OMI tool. At this time, the device model is updated to the new model after aging, and SPICE can accurately simulate the aging situation of the current circuit.

[0074] Furthermore, after obtaining the current aging simulation results of the circuit under test, it also includes:

[0075] Perform aging extrapolation based on the current aging simulation results of the circuit under test until the preset aging time;

[0076] Update the model parameters and pass them to the SPICE simulator for the next round of aging simulation, so as to evaluate the long-term performance of the circuit under test.

[0077] Specifically, please refer to Figure 6 , Figure 6 FIG. is a schematic diagram of the integrated circuit reliability simulation process provided by the embodiments of the present invention. Combining this drawing, the process of performing reliability simulation on a certain integrated circuit using a machine learning-based circuit aging simulation method proposed by the present invention can be described as follows:

[0078] First, for the designed integrated circuit to be simulated, obtain the unaged circuit netlist of the circuit, which contains all the devices of the circuit and their connection information.

[0079] Then, determine the model parameters related to aging according to the aging model of the device.

[0080] Next, map the parameters in the aging model to the actual circuit elements, that is, the model parameters, for use in the subsequent SPICE simulation. The parameter mapping in this step is implemented using the neural network in the present invention.

[0081] After that, combine the SPICE simulator and the aging model to perform aging calculations to obtain the instantaneous degradation amount of the device parameters. Then, extrapolate the transient simulation results to the actual aging time (such as 10 years) through aging extrapolation using a time acceleration factor (such as the power-law model).

[0082] Finally, after parameter update, enter the next round of simulation to achieve the evaluation of the long-term performance of the circuit under test.

[0083] A circuit aging simulation method based on machine learning provided by the present invention first selects model parameters reflecting the aging condition of a circuit to be measured based on the aging models of circuit devices, then uses a SPICE simulator to scan the model parameters to obtain the electrical parameters after simulation; then uses a trained neural network to map the electrical parameters after simulation to the model parameters to obtain the specific values of the model parameters; finally, inputs the specific values of the model parameters into the SPICE simulator for aging simulation to obtain the current aging simulation result of the circuit to be measured. On the one hand, starting from the aging model, this method selects different model parameters that can best reflect the circuit aging condition based on the self-degradation of the devices. On the other hand, based on machine learning technology, it uses a pre-trained neural network to map the electrical performance degradation caused by device aging to the model parameters related to device aging, realizing the real-time update of device model parameters according to the actual aging condition using the device aging model, and can help the circuit aging simulation tool to achieve high-precision aging simulation prediction based on machine learning multi-parameter mapping. By accurately simulating and predicting the degradation of the device, the simulation result can flexibly reflect the actual degradation of the device.

[0084] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A circuit aging simulation method based on machine learning, characterized in that: include: Selecting model parameters that reflect the aging of the circuit to be tested based on the aging model of the circuit device; Scanning the model parameters using a SPICE simulator to obtain simulated electrical parameters; Mapping the simulated electrical parameters to model parameters using a trained neural network to obtain specific values ​​of the model parameters; The specific values ​​of the model parameters are input into the SPICE simulator to perform aging simulation, and the current aging simulation results of the circuit to be tested are obtained.

2. The circuit aging simulation method based on machine learning according to claim 1, characterized in that: The circuit device-based aging model selects model parameters that reflect the aging of the circuit to be tested, including: Selecting multiple electrical parameters reflecting the aging condition of the circuit to be tested according to the aging model of the circuit device; The selected electrical parameters are scanned and simulated using a SPICE simulator, and several aging-sensitive model parameters are screened out based on the simulation results.

3. The circuit aging simulation method based on machine learning according to claim 2, characterized in that: The SPICE simulator is used to scan and simulate the selected electrical parameters, and several aging-sensitive model parameters are further screened out according to the simulation results, including: Use the SPICE simulator to scan and simulate the selected electrical parameters, and record the changes of electrical degradation related indicators corresponding to different parameters to obtain simulation results; Based on the simulation results, the LASSO algorithm is used for parameter screening to screen out several aging-sensitive model parameters; or, the elastic network regularization algorithm is used to establish a parameter-performance sparse regression model to screen out several aging-sensitive model parameters.

4. The circuit aging simulation method based on machine learning according to claim 1, characterized in that: The method of scanning the model parameters using a SPICE simulator to obtain simulated electrical parameters includes: determining a scanning range of the model parameters; Based on the determined scanning range, the model parameters are scanned using a SPICE simulator to obtain simulated electrical parameters.

5. The circuit aging simulation method based on machine learning according to claim 4, characterized in that: The scanning range of the model parameters is determined according to the properties of each parameter in the semiconductor, or determined by the 2σ principle; wherein σ is obtained by Monte Carlo analysis.

6. The circuit aging simulation method based on machine learning according to claim 1, characterized in that: The trained neural network is obtained in the following way: The model parameters of different circuits are scanned using a SPICE simulator to obtain the simulated electrical parameters as a training data set; A fully connected neural network model is constructed, and the fully connected neural network model is trained using the training data set to obtain a trained neural network.

7. The circuit aging simulation method based on machine learning according to claim 6, characterized in that: The step of constructing a fully connected neural network model and training the fully connected neural network model using the training data to obtain a trained neural network includes: Build a fully connected neural network model; The fully connected neural network model is trained using all types of electrical parameters in the training data set to obtain a trained neural network.

8. The circuit aging simulation method based on machine learning according to claim 6, characterized in that: The step of constructing a fully connected neural network model and training the fully connected neural network model using the training data to obtain a trained neural network includes: For different types of electrical parameters, fully connected neural network models are constructed respectively; The corresponding fully connected neural network models are trained using different types of electrical parameters in the training data set to obtain a trained neural network.

9. A circuit aging simulation method based on machine learning according to claim 7 or 8, characterized in that: The training process of the fully connected neural network model also includes: Different types of electrical parameters in the training data set are segmented according to degradation amplitudes, and the fully connected neural network model is trained using the data of each segment to obtain a segmented neural network.

10. The circuit aging simulation method based on machine learning according to claim 1, characterized in that: After obtaining the current aging simulation results of the circuit under test, it also includes: Perform aging extrapolation according to the current aging simulation result of the circuit under test until a preset aging time; The model parameters are updated and passed to the SPICE simulator to perform the next round of aging simulation, thereby achieving the evaluation of the long-term performance of the circuit under test.

Images