Transient current prediction method based on single-particle transient current neural network model

Abstract

The invention discloses a transient current prediction method based on a single-particle transient current neural network model, relates to the technical field of simulation of semiconductor devices, and solves the problem that the transient current of a single particle cannot be accurately calculated in the prior art. Calculating transient current according to the trained single-particle transient current neural network model and the single-particle irradiation parameters; wherein the single-particle transient current neural network model comprises a neural network model and a double-exponential current model; constructing a semiconductor device model to obtain a standard transient current set; training a neural network model according to the standard transient current set to obtain a trained neural network model; combining the neural network model with the double-exponential current model to obtain a single-particle transient current neural network model; according to the method, model support is provided for single-particle transient fault injection according to the single-particle transient current neural network model, so that an accurate transient current value is obtained.

Description

Technical Field

[0001] The invention relates to the technical field of simulation of semiconductor devices, and in particular to a transient current prediction method based on a single particle transient current neural network model. Background Art

[0002] With the development of technology, microelectronic components are increasingly used in the aerospace field. Devices will face severe tests of radiation reliability in the space radiation environment. As the size of integrated circuits continues to shrink, the pursuit of better performance, higher speed and lower cost also increases the impact of single-particle effects on chips. When high-energy particles enter the device, free charges are generated along its incident path. When collected by sensitive nodes, significant transient current pulses are generated at the node. This phenomenon is called single-particle transient effects. Single-particle transient effects pose new challenges to the reliability and radiation-resistant design of integrated circuits. The transient current it generates may cause the circuit logic state to change. Once captured by the storage element, it will cause soft errors in the circuit or system. As the device process enters the nanoscale, the density of integrated circuits continues to increase, and the device size becomes smaller and smaller, resulting in a decrease in critical charge. The device is more susceptible to the influence of high-energy particles, which brings challenges to the radiation-resistant reliability design.

[0003] In recent years, the research on single-particle transient effects has tended to model single-particle irradiation simulation and single-particle transient current of semiconductor devices. Device simulation software is an important research tool that can observe information such as potential, electrostatic field, and carrier concentration during charge generation and collection. The redistribution of carriers and potential inside the device under various incident conditions can be obtained, and then the changes in charge, current, and voltage at each end of the device can be obtained. With the development of artificial intelligence, neural network models have been widely used in many fields such as data analysis, machine control, and signal processing. Thanks to its nonlinear mapping, self-learning, self-regulation, generalization, and fault tolerance functions, through training, a result prediction model for the corresponding data can be quickly established.

[0004] In the current context, there is an urgent need to develop a fast and accurate single-particle transient current pulse model. Summary of the invention

[0005] The present invention solves the problem that the transient current of a single particle cannot be accurately calculated in the prior art by providing a transient current prediction method based on a single-particle transient current neural network model, and provides model support for single-particle transient fault injection based on the single-particle transient current neural network model to obtain an accurate transient current value.

[0006] The present invention provides a transient current prediction method based on a single particle transient current neural network model, the method comprising:

[0007] Obtain single particle irradiation parameters;

[0008] The single-particle irradiation parameters are input into the neural network model in the trained single-particle transient current neural network model to obtain a predicted charge collection amount; wherein, the training of the neural network model includes: constructing and adjusting an initial semiconductor device model to obtain a semiconductor device model, and using a plurality of sets of heavy ion incidence parameters to perform irradiation simulation on the semiconductor device model to obtain a standard transient current set; calculating the charge collection amount corresponding to each standard transient current in the standard transient current set to obtain a charge collection amount set; constructing a neural network model, and training the neural network model according to the plurality of sets of heavy ion incidence parameters and the charge collection amount set to obtain a trained neural network model;

[0009] The double exponential current model in the single-particle transient current neural network model calculates the transient current corresponding to the single-particle irradiation parameter according to the predicted charge collection amount.

[0010] In a possible implementation, the constructing and adjusting the initial semiconductor device model to obtain the semiconductor device model includes:

[0011] Using the SDE tool to establish an initial semiconductor device model, and using the SDEVICE tool to simulate the electrical characteristics of the initial semiconductor device model to obtain a transfer characteristic curve corresponding to the initial semiconductor device model;

[0012] Determine a standard transfer characteristic curve with the same conditions as the initial semiconductor device model, compare the transfer characteristic curve with the standard transfer characteristic curve, and adjust the doping concentration in the initial semiconductor device model according to the comparison result until the curve in the transfer characteristic curve coincides with the curve in the standard transfer characteristic curve to obtain a semiconductor device model.

[0013] In a possible implementation, the semiconductor device model performs radiation simulation to obtain a standard transient current set, including:

[0014] Determine the type of semiconductor device in the semiconductor device model. When the type of the semiconductor device is an NMOS device, in the semiconductor device model, connect the well contact region to the ground, connect the substrate contact region to the ground, connect the drain of the NMOS device structure to a high level, and connect the source and the gate to the ground;

[0015] When the type of the semiconductor device is a PMOS device, in the semiconductor device model, the well contact region is connected to a high level, the substrate contact region is grounded, the drain of the PMOS device structure is grounded, and the source and gate are connected to a high level;

[0016] A plurality of groups of heavy ion incidence parameters are determined, and each group of heavy ion incidence parameters is respectively used to irradiate the sensitive area of ​​the semiconductor device model to obtain a standard transient current set.

[0017] In a possible implementation, the heavy ion incident parameters include: incident distance, incident angle and incident LET value;

[0018] The incident distance is the straight-line distance from the incident point of the heavy ions to the edge of the source region of the gate;

[0019] The incident angle is the angle between the straight line where the heavy ion incident track is located and the normal line of the device plane;

[0020] The incident LET value is the linear energy transfer value of the incident particle itself at the moment it enters the device.

[0021] In a possible implementation manner, the sensitive area is a drain region of the semiconductor device.

[0022] In a possible implementation, the learning rate of the neural network model is set to 0.001.

[0023] In one possible implementation, the calculation of the charge collection amount corresponding to each standard transient current in the standard transient current set to obtain the charge collection amount set includes: using the Inspect tool to write a program to calculate the integral of each standard transient current in the standard transient current set over time, and then obtaining the charge collection amount set collected in the sensitive area of ​​the device.

[0024] In a possible implementation, the neural network model includes: an input layer and an output layer;

[0025] The input layer includes three neurons connected in sequence and a sigmoid activation function;

[0026] The output layer includes a neuron and a tanh activation function connected in sequence.

[0027] In a possible implementation, the training of the neural network model according to the multiple sets of heavy ion incidence parameters and the charge collection amount sets to obtain a trained neural network model includes:

[0028] The weight value and bias value of each layer of the neural network model in training are adjusted layer by layer using the forward transfer method and the backward transfer method until the error of the neural network model is less than a preset threshold, thereby obtaining a trained neural network model.

[0029] In a possible implementation, the transient current is obtained by a transient current calculation formula in the double exponential current model, and the transient current calculation formula is expressed as:

[0030]

[0031] Among them, α represents the first constant; β represents the second constant; Q c Indicates the amount of charge collected; I (t) represents transient current; e represents charge; t represents time.

[0032] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0033] The present invention adopts a transient current prediction method based on a single-particle transient current neural network model. The model uses semiconductor device simulation software to model and calibrate the semiconductor device structure, and changes the incident distance, incident angle and incident LET value in the radiation simulation for simulation. The simulation results are used to train the established neural network model, and the neural network model is combined with the double exponential current pulse model. The charge collection amount in the double exponential current pulse model is provided by the neural network, and finally the single-particle transient current neural network model of the semiconductor device is established. Through this model, the transient current pulse of the semiconductor device under single-particle irradiation can be predicted and analyzed, providing model support for single-particle transient fault injection, thereby providing an important basis and certain guidance for the radiation resistance evaluation of semiconductor devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of the steps of a transient current prediction method based on a single particle transient current neural network model provided by an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of the particle incident angle during single particle irradiation simulation provided by an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of particle incident distance during single particle irradiation simulation provided by an embodiment of the present invention;

[0037] Figure 4 A schematic diagram for comparing prediction results of a neural network training set provided in an embodiment of the present invention;

[0038] Figure 5 A schematic diagram for comparing prediction results of a neural network test set provided by an embodiment of the present invention;

[0039] Figure 6 A schematic diagram of the structure of a single-particle transient current neural network model of a semiconductor device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] The present invention provides a transient current prediction method based on a single particle transient current neural network model, such as Figure 1 As shown, the method includes the following steps S101 to S103.

[0042] S101, obtaining single particle irradiation parameters.

[0043] S102, inputting the single particle irradiation parameters into the neural network model in the trained single particle transient current neural network model to obtain the predicted charge collection amount.

[0044] In step S102, the single particle transient current neural network model includes: a neural network model and a double exponential current model. Figure 6 shown.

[0045] Training a neural network model consists of the following steps:

[0046] S1021, construct and adjust the initial semiconductor device model to obtain the semiconductor device model, and use multiple sets of heavy ion irradiation parameters to perform irradiation simulation on the semiconductor device model to obtain a standard transient current set. Here, the semiconductor device model refers to: NMOS and PMOS three-dimensional MOSFET device models using lightly doped source and drain technology and shallow trench isolation technology.

[0047] Specifically, in step S1021, an initial semiconductor device model is constructed and adjusted to obtain a semiconductor device model, including:

[0048] (1) Using the SDE tool to establish an initial semiconductor device model, and using the SDEVICE tool to simulate the electrical characteristics of the initial semiconductor device model to obtain a transfer characteristic curve corresponding to the initial semiconductor device model;

[0049] (2) Determine a standard transfer characteristic curve with the same conditions as the initial semiconductor device model, compare the transfer characteristic curve with the standard transfer characteristic curve, and adjust the doping concentration in the initial semiconductor device model according to the comparison result until the curve in the transfer characteristic curve coincides with the curve in the standard transfer characteristic curve, thereby obtaining the semiconductor device model.

[0050] Exemplarily, a MOSFET model is established using TCAD software, and an initial model of a MOSFET device is established using the SDE tool, whose functions include editing the device structure, defining contact points, doping concentration, generating a grid, and saving files. Subsequently, the SDEVICE tool is used to simulate the electrical characteristics of the built model, and a curve that can reflect the electrical characteristics of the device is obtained: a transfer characteristic curve. At the same time, the electrical characteristics of the SPICE model file of the same size or the model in the process library are simulated under the same conditions. The obtained curves are compared, and the doping concentration of the MOSFET model is adjusted until the two curves coincide, completing the MOSFET device structure modeling and calibration.

[0051] S1022, calculate the charge collection amount corresponding to each standard transient current in the standard transient current set to obtain a charge collection amount set. Specifically, according to the TCAD simulation results, use the Inspect tool to write a program to calculate the integral of the transient current over time to obtain the charge collection amount collected in the sensitive area of ​​the device.

[0052] S1023, constructing a neural network model, and training the neural network model according to multiple sets of heavy ion incidence parameters and charge collection amount sets to obtain a trained neural network model.

[0053] Here, the neural network model includes: an input layer and an output layer; the input layer includes three neurons connected in sequence and a sigmoid activation function; the output layer includes one neuron connected in sequence and a tanh activation function.

[0054] Specifically, in step S1022, multiple sets of heavy ion incidence parameters are used to perform irradiation simulation on the semiconductor device model to obtain a standard transient current set, including:

[0055] (1) determining the type of semiconductor device in the semiconductor device model. When the type of the semiconductor device is an NMOS device, in the semiconductor device model, the well contact region is connected to the ground, the substrate contact region is grounded, the drain of the NMOS device structure is connected to a high level, and the source and gate are grounded;

[0056] (2) When the type of semiconductor device is a PMOS device, in the semiconductor device model, the well contact region is connected to a high level, the substrate contact region is grounded, the drain of the PMOS device structure is grounded, and the source and gate are connected to a high level;

[0057] (3) The sensitive area of ​​the semiconductor device model is irradiated according to each set of heavy ion irradiation parameters to obtain a standard transient current set.

[0058] Here, the heavy ion incident parameters include: incident distance, incident angle and incident LET value; the incident distance is the straight-line distance from the heavy ion incident point to the edge of the source region of the gate; the incident angle is the angle between the straight line where the heavy ion incident track is located and the normal line of the device plane; the incident LET value is the linear energy transfer value of the incident particle itself at the moment it enters the device. Figure 2 When simulating single particle irradiation, the particle incident distance is as follows: Figure 3 shown.

[0059] Here, the sensitive region is the drain region of the semiconductor device.

[0060] Here, the learning rate of the neural network model is set to 0.001.

[0061] Exemplarily, respectively changing the incident angle, incident distance and incident LET value to perform irradiation simulation means: for example, controlling the incident distance and incident angle to remain unchanged, changing the incident LET to different values ​​for simulation. Then controlling the incident distance to remain unchanged, changing the incident angle, and traversing the simulation for each incident LET value. Finally, changing the incident distance, traversing each incident angle value, and traversing the incident LET value at each angle value for simulation. If the incident distance is set to 12 groups, the incident angle is set to 6 groups, and the incident LET value is set to 8 groups, there will be 576 groups of simulations in the end.

[0062] The sensitive region refers to: the drain region of the off-state MOSFET device. The irradiated heavy ion incidence parameter (simulation model) refers to: the heavy ion incidence model, that is, activating the single particle transient simulation caused by heavy ion collision in the simulation.

[0063] In a specific embodiment provided by the present invention, a neural network model is trained according to multiple sets of heavy ion incidence parameters and charge collection sets to obtain a trained neural network model, which also includes: using the neural network model to predict the predicted charge corresponding to the multiple sets of heavy ion incidence parameters. The specific prediction process is expressed by the formula: Q c =F(Angle, D, LET); calculate the error between the predicted charge and the charge collection under the same heavy ion incident parameters, and adjust the weight and bias value of each layer of the neural network model in training layer by layer according to the error, forward transfer method and reverse transfer method, until the error of the neural network model is less than the preset threshold or the set number of training times is reached, and the trained neural network model is obtained. Among them, Angle represents the incident angle; D represents the incident distance; LET represents the linear energy transfer value of the incident particle itself at the moment it enters the device; Q c Represents the predicted value of the neural network for the amount of charge collection.

[0064] Exemplarily, in a specific embodiment, the neural network model is specifically a BP neural network model; the training of the BP neural network model refers to: taking the incident distance, incident angle and incident LET value as inputs of the BP neural network model, and the charge collection amount as output of the BP neural network model.

[0065] The input layer of the BP neural network model has 1 layer, including 3 neurons. The output layer of the BP neural network model has 1 layer, including 1 neuron. The number of neurons in the hidden layer is estimated using an empirical formula to a starting value, and through repeated experimental adjustments, a configuration that optimizes system performance is found.

[0066] The empirical formula is specifically expressed as: Among them, M represents the number of neurons, n is the number of neurons in the input layer, m is the number of neurons in the output layer, and a is an adjustment factor, whose value is usually between 1 and 10.

[0067] The learning rate affects the update amplitude of weights and thresholds. The learning rate of the BP neural network model is set to 0.001.

[0068] The training of the neural network model means: taking the incident distance, incident angle and LET value as the input of the BP neural network model, and the charge collection amount as the output of the BP neural network model, and training the device model with the data of single particle irradiation simulation.

[0069] The training process of BP neural network is divided into two processes: forward transfer calculation and error back propagation. Forward transfer calculation means that the neurons in each layer summarize the information of the previous layer and pass it to the next layer after processing using the activation function. When the information is finally passed to the output layer, the system will calculate the error between the output result and the actual result. Error back propagation means that the error is passed back from the output layer to the input layer, while adjusting the connection weights and bias values ​​layer by layer until the network error is reduced to an acceptable preset threshold or reaches a specified number of training times.

[0070] Among the 576 sets of simulation data, 403 sets of data were randomly selected as training sets, and the remaining 173 sets of data were used as test sets, that is, the ratio of the number of samples in the training set to the test set was 7:3.

[0071] The coefficient of determination R2 and root mean square error RMSE are selected as model evaluation indicators. R2: reflects the model's ability to explain data variability. The closer its value is to 1, the better the model fits the data. RMSE: measures the degree of difference between the predicted value and the actual value. The smaller its value, the higher the accuracy.

[0072] The comparison of the prediction results of the neural network training set is shown in the figure Figure 4As shown in the figure, the RMSE of the training set is 0.56628 and the R2 is 0.99973. The comparison of the prediction results of the neural network test set is shown in the figure Figure 5 As shown, the RMSE of the test set is 0.76128 and the R2 is 0.99927. In the comparison of the prediction results of the training set and the test set, the horizontal axis is different samples, the vertical axis is the value of the charge collection, the red asterisk is the true value, and the blue circle is the predicted value of the neural network model.

[0073] S103, the double exponential current model in the single-particle transient current neural network model calculates the transient current corresponding to the single-particle irradiation parameter according to the predicted charge collection amount.

[0074] Exemplarily, the neural network model predicts the charge collection amount of the semiconductor device based on the incident heavy ion conditions, substitutes the charge collection amount into the double exponential current pulse model, and establishes the transient current pulse model generated by the semiconductor device under the corresponding heavy ion conditions, that is: given the incident heavy ion conditions, the corresponding single-particle transient current pulse model of the semiconductor device is obtained.

[0075] In step S103, the transient current calculation formula is expressed as:

[0076]

[0077] Among them, α represents the first constant; β represents the second constant; Q c Indicates the amount of charge collected; I (t) represents transient current; e represents charge; t represents time.

[0078] The present invention uses semiconductor device simulation software to model and calibrate the semiconductor device structure, and in the radiation simulation, the incident distance, incident angle and incident LET value are respectively changed for simulation. The simulation results are used to train the established neural network model, and the neural network model is combined with the double exponential current pulse model. The charge collection amount in the double exponential current pulse model is provided by the neural network, and finally a transient current prediction method based on the single-particle transient current neural network model is established. Through this model, the transient current pulse of the semiconductor device under single-particle irradiation can be predicted and analyzed, providing model support for single-particle transient fault injection, thereby providing an important basis and certain guidance for the radiation resistance evaluation of semiconductor devices.

[0079] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. All or part of the present invention can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the present invention.

Claims

1. A transient current prediction method based on a single particle transient current neural network model, characterized in that: include: Obtain single particle irradiation parameters; The single-particle irradiation parameters are input into the neural network model in the trained single-particle transient current neural network model to obtain a predicted charge collection amount; wherein, the training of the neural network model includes: constructing and adjusting an initial semiconductor device model to obtain a semiconductor device model, and using a plurality of sets of heavy ion incidence parameters to perform irradiation simulation on the semiconductor device model to obtain a standard transient current set; calculating the charge collection amount corresponding to each standard transient current in the standard transient current set to obtain a charge collection amount set; constructing a neural network model, and training the neural network model according to the plurality of sets of heavy ion incidence parameters and the charge collection amount set to obtain a trained neural network model; The double exponential current model in the single-particle transient current neural network model calculates the transient current corresponding to the single-particle irradiation parameter according to the predicted charge collection amount.

2. The transient current prediction method based on the single particle transient current neural network model according to claim 1 is characterized in that: The step of constructing and adjusting the initial semiconductor device model to obtain the semiconductor device model includes: Using the SDE tool to establish an initial semiconductor device model, and using the SDEVICE tool to simulate the electrical characteristics of the initial semiconductor device model to obtain a transfer characteristic curve corresponding to the initial semiconductor device model; The transfer characteristic curve is compared with the standard transfer characteristic curve, and according to the comparison result, the doping concentration in the initial semiconductor device model is adjusted until the curve in the transfer characteristic curve coincides with the curve in the standard transfer characteristic curve, thereby obtaining a semiconductor device model.

3. The transient current prediction method based on the single particle transient current neural network model according to claim 1 is characterized in that: The semiconductor device model is irradiated and simulated using multiple sets of heavy ion incidence parameters to obtain a standard transient current set, including: Determine the type of semiconductor device in the semiconductor device model. When the type of the semiconductor device is an NMOS device, in the semiconductor device model, connect the well contact region to the ground, connect the substrate contact region to the ground, connect the drain of the NMOS device structure to a high level, and connect the source and the gate to the ground; When the type of the semiconductor device is a PMOS device, in the semiconductor device model, the well contact region is connected to a high level, the substrate contact region is grounded, the drain of the PMOS device structure is grounded, and the source and gate are connected to a high level; The sensitive area of ​​the semiconductor device model is irradiated according to each group of heavy ion irradiation parameters to obtain a standard transient current set.

4. The transient current prediction method based on the single particle transient current neural network model according to claim 3 is characterized in that: The heavy ion incident parameters include: incident distance, incident angle and incident LET value; The incident distance is the straight-line distance from the incident point of the heavy ions to the edge of the source region of the gate; The incident angle is the angle between the straight line where the heavy ion incident track is located and the normal line of the device plane; The incident LET value is the linear energy transfer value of the incident particle itself at the moment it enters the device.

5. The transient current prediction method based on the single particle transient current neural network model according to claim 3 is characterized in that: The sensitive area is the drain region of the semiconductor device.

6. The transient current prediction method based on the single particle transient current neural network model according to claim 3 is characterized in that: The learning rate of the neural network model is set to 0.

001.

7. The transient current prediction method based on the single particle transient current neural network model according to claim 1 is characterized in that: The method of calculating the charge collection amount corresponding to each standard transient current in the standard transient current set to obtain the charge collection amount set includes: using the Inspect tool to write a program to calculate the integral of each standard transient current in the standard transient current set over time, and then obtaining the charge collection amount set collected in the sensitive area of ​​the device.

8. The transient current prediction method based on the single particle transient current neural network model according to claim 1 is characterized in that: The neural network model comprises: an input layer and an output layer; The input layer includes three neurons connected in sequence and a sigmoid activation function; The output layer includes a neuron and a tanh activation function connected in sequence.

9. The transient current prediction method based on the single particle transient current neural network model according to claim 1, characterized in that: The step of training the neural network model according to the multiple sets of heavy ion incidence parameters and the charge collection amount set to obtain a trained neural network model comprises: The weight value and bias value of each layer of the neural network model in training are adjusted layer by layer using the forward transfer method and the backward transfer method until the error of the neural network model is less than a preset threshold, thereby obtaining a trained neural network model.

10. The transient current prediction method based on the single particle transient current neural network model according to claim 1, characterized in that: The transient current is obtained by the transient current calculation formula in the double exponential current model, and the transient current calculation formula is expressed as: Among them, α represents the first constant; β represents the second constant; Q c Indicates the amount of charge collected; I (t) represents transient current; e represents charge; t represents time.

Images