A method for predicting the characteristics of amorphous oxide semiconductor devices based on mixed datasets

Abstract

This invention discloses a method for predicting the characteristics of amorphous oxide semiconductor devices based on a hybrid dataset. The method involves constructing a hybrid dataset of drain-source currents for amorphous oxide semiconductor devices and using it to train an artificial neural network model. Device size parameters and voltage parameters are input into the trained artificial neural network model, which outputs predicted drain-source current values. Based on the device's voltage parameters and the corresponding predicted drain-source current values, the electrical characteristics of the amorphous oxide semiconductor device under different device size parameters are predicted. This invention rationally integrates data from different sources within a unified framework, fundamentally resolving the contradiction between insufficient measured data and significant deviations in simulation data in traditional modeling methods. It achieves a data fusion approach of "measured data as the primary source, supplemented by simulation data," making it easier for the neural network to learn electrical laws consistent with real devices, thereby further improving the stability, accuracy, and physical consistency of the prediction model.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device modeling and artificial intelligence-assisted manufacturing technology, specifically relating to a method for predicting the characteristics of amorphous oxide semiconductor devices based on a hybrid dataset. Background Technology

[0002] Amorphous oxide semiconductors, as a new generation of semiconductor materials, have become a key material platform in cutting-edge fields such as flexible displays, transparent electronics, and large-scale sensor arrays due to their unique advantages, including high electron mobility, visible light transparency, low fabrication temperature, and good uniformity. With the continuous miniaturization of device structures, the increasing functional integration, and the expansion of applications to wearable and biocompatible applications, accurately characterizing the electrical properties of amorphous oxide semiconductor devices under different operating conditions (including bias voltage and geometry) is of paramount importance for optimizing material formulations, designing device structures, determining process windows, and assessing long-term reliability.

[0003] In existing technologies, two types of data are typically relied upon to construct predictive models of device characteristics. One type is experimental measurement data from actual devices, which can accurately reflect the current response of real devices in different operating regions (subthreshold region, linear region, saturation region) and has high reliability. However, the acquisition of measured data depends on expensive testing equipment and long experimental cycles, and it is difficult to cover a wide range of gate length, gate width, and voltage bias conditions in a short period of time, especially for certain boundary conditions or extreme-sized devices, where it is difficult to obtain sufficient data support.

[0004] Another type is physical simulation data based on TCAD (Technology Computer-Aided Design). By establishing device structure models and setting material parameters and physical models, a large amount of device characteristic data with broad coverage and high controllability can be generated, making it particularly suitable for exploring possible outputs under unmanufactured devices or different process conditions. Although TCAD simulation has advantages in supplementing data volume and expanding parameter space, due to the complexity of physical models such as material models, interface defects, trapped state distribution, and charge transport mechanisms in oxide semiconductors, there is usually a significant deviation between simulated and measured currents. Especially in areas such as subthreshold current, carrier mobility models, and contact resistance models, simulation results are prone to deviating from actual device behavior. In addition, simulation data lacks local defects, random variations, and non-ideal effects in real manufacturing processes, making it difficult for models trained solely on simulation data to achieve high-accuracy predictive performance on real devices.

[0005] With the development of artificial intelligence, using machine learning methods such as neural networks to predict the characteristics of semiconductor devices has become an important direction. However, existing models often face the problem of an imbalance between data quantity and data quality: if training is mainly based on measured data, the model's coverage and generalization ability are limited; if relying mainly on simulation data, the model will be affected by simulation bias and perform poorly on actual devices. Traditional mixed data training methods lack effective adjustment mechanisms for simulation bias, resulting in the inability to reasonably handle the differences in statistical distribution between measured and simulation data, thus affecting the model's training stability and physical consistency. Especially in terms of the monotonicity, smoothness, and derivative continuity of the I-V curve, neural networks lacking physical constraints are prone to non-physical behaviors, affecting the model's credibility in actual process development.

[0006] Therefore, how to fully utilize the high reliability of measured data, introduce broad-coverage TCAD simulation data, effectively reduce simulation bias, and integrate the two types of data within a unified framework has become a critical issue that urgently needs to be addressed in the current field of amorphous oxide semiconductor device modeling. Simultaneously, it is necessary to construct hybrid data-driven models that can guarantee physical consistency, prediction stability, and high generalization ability, to be applicable to device characteristic prediction under different size and voltage bias conditions. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the characteristics of amorphous oxide semiconductor devices based on a mixed dataset.

[0008] This invention is implemented by providing a method for predicting the characteristics of amorphous oxide semiconductor devices based on a mixed dataset, the method comprising the following steps:

[0009] Construct a mixed dataset of drain-source currents for amorphous oxide semiconductor devices and preprocess it;

[0010] An artificial neural network model was constructed and trained using a preprocessed mixed dataset of drain-source currents.

[0011] The device size parameters and voltage parameters of the amorphous oxide semiconductor device are input into the trained artificial neural network model, and the predicted drain-source current value is output.

[0012] Based on the voltage parameters of amorphous oxide semiconductor devices and the corresponding predicted drain-source current, the electrical characteristics of amorphous oxide semiconductor devices under different device size parameters are predicted.

[0013] The drain-source current hybrid dataset consists of measured drain-source current values, corrected drain-source current simulation values, and extended drain-source current simulation values. The measured drain-source current values ​​and corrected drain-source current simulation values ​​are drain-source current data of amorphous oxide semiconductor devices with the same device size parameters and voltage parameters. The extended drain-source current simulation values ​​are obtained by extending the corrected drain-source current simulation values.

[0014] Preferably, the process of obtaining the measured drain-source current is as follows: obtaining a set of measured data for amorphous oxide semiconductor devices under different device sizes and different voltage combinations; wherein, the set of measured data includes device size parameters, voltage parameters, and measured drain-source current values; the device size parameters include gate length and gate width, and the voltage parameters include gate-source voltage and drain-source voltage.

[0015] Preferably, the process of obtaining the drain-source current correction simulation value is as follows: construct an amorphous oxide semiconductor device structure model in TCAD simulation software, define the device size, material region, electrode position and mesh division, and obtain the corresponding drain-source current simulation value based on the device size parameters and voltage parameters in the measured data set; use a domain adaptive mapping network to map the drain-source current simulation value to a drain-source current correction simulation value that is closer to the distribution of the measured drain-source current value.

[0016] Preferably, the process of obtaining the drain-source current spread simulation value is as follows: fitting the simulation value of each drain-source current correction to obtain a fitting curve, and using the fitting curve to perform interpolation or extrapolation to obtain a series of drain-source current spread simulation values.

[0017] Preferably, the domain adaptive mapping network adopts a multi-layer feedforward fully connected network.

[0018] Preferably, when the domain adaptive mapping network maps the simulated values ​​of drain-source current correction, it is also necessary to perform sample screening and resampling based on physical consistency on the simulated values ​​of drain-source current correction. The implementation process is as follows:

[0019] (1) Physically unreasonable values ​​are eliminated from the sample of the drain-source current correction simulation values. The judgment of physical unreasonableness is based on the physical constraints of the device. The physical constraints of the device are whether the drain-source current changes with the gate-source voltage or the drain-source voltage and satisfies the monotonicity, continuity and non-negativity. If it does not satisfy these conditions, it is physically unreasonable; otherwise, it is reasonable.

[0020] (2) Perform logarithmic transformation on the simulated values ​​of drain source current after removing physically unreasonable values.

[0021] Preferably, during the training process, the loss function of the artificial neural network model is the weighted mean square error of the i-th predicted drain-source current output by the artificial neural network model and the i-th drain-source current data in the mixed dataset; wherein when the i-th drain-source current data in the mixed dataset is the measured drain-source current value, the corresponding weight value is... , `max` represents the maximum value function; when the i-th drain-source current data in the mixed dataset is the drain-source current correction simulation value, the corresponding weight value is... , Indicates the curve-level physical error weights. This represents the error weight of the i-th sample level; when the i-th drain-source current data in the mixed dataset is the extended simulation value of the drain-source current, the corresponding weight value is the confidence coefficient c.

[0022] Preferably, the calculation process for the curve-level physical error weights is as follows:

[0023] The simulated value of the drain-source current correction and the simulated value of the drain-source current extension are combined and denoted as the simulated value of the drain-source current. A simulated drain-source current curve is constructed based on the simulated value of the drain-source current, with the vertical axis representing the simulated value of the drain-source current and the horizontal axis representing the voltage parameter.

[0024] Based on the measured values ​​of the drain-source current, a measured drain-source current curve is constructed, with the vertical axis representing the measured drain-source current and the horizontal axis representing the voltage parameter.

[0025] For the measured and simulated drain-source current curves under the same device size and voltage parameters, the subthreshold slope, effective mobility in the linear region, and channel modulation factor in the saturation region are extracted, and the corresponding physical deviations are calculated. Based on the three physical deviations, curve-level physical error weights are constructed.

[0026] Preferably, the calculation process for the sample-level error weights is as follows:

[0027] The simulated value of the drain-source current correction and the simulated value of the drain-source current extension are combined and denoted as the simulated value of the drain-source current. A simulated drain-source current curve is constructed based on the simulated value of the drain-source current, with the vertical axis representing the simulated value of the drain-source current and the horizontal axis representing the voltage parameter.

[0028] For the voltage parameters corresponding to the simulated drain-source current curve, the point-by-point deviation between the simulated drain-source current value and the measured drain-source current value is calculated, and sample-level error weights are further constructed.

[0029] Preferably, the confidence coefficient It is half the positive difference between the curve-level physical error weight and the sample-level error weight.

[0030] The beneficial effects of this invention are at least as follows:

[0031] First, by constructing a hybrid dataset that integrates measured data and TCAD simulation data, and introducing a bias-based adaptive hierarchical weighting mechanism, data from different sources are rationally integrated within a unified framework, fundamentally resolving the contradiction between insufficient measured data and significant deviations in simulation data in traditional modeling methods. Through domain adaptive mapping networks to correct the simulated current, this invention significantly reduces systematic errors in TCAD simulations regarding material parameters, interface state models, and carrier transport models, making the statistical distribution of simulation data closer to measured data, thereby improving the realism and effectiveness of the hybrid dataset.

[0032] Second, the designed weighting strategy based on the measured-simulation bias achieves a data fusion method of "measured data as the main component and simulation data as an auxiliary supplement," effectively suppressing the cumulative impact of simulation errors in model training and enabling the network to rely more on real device information during the learning process. For samples lacking measured coverage, this invention assigns appropriate weights through confidence coefficients, allowing simulation data to expand the model's parameter coverage range while ensuring stability, thereby significantly enhancing the model's generalization ability.

[0033] Third, this invention further introduces curve-level weighting factors based on device physical parameter deviations. It uses the measured-simulation differences of key physical quantities such as subthreshold slope, effective mobility, and channel modulation factor as physical consistency indicators, expanding the weighting from the original "data-level deviation" to "physical behavior-level deviation." This physical-level weighting reflects the modeling accuracy of TCAD under different physical mechanisms, such as the subthreshold region, electric field saturation region, and linear region. It reduces the overall weight of simulation curves with significant physical deviations, preventing them from systematically misleading model training. Compared to traditional weighting methods based solely on current errors, physical weighting significantly improves the physical credibility and interpretability of the data fusion process, making it easier for neural networks to learn electrical laws consistent with real devices, thereby further improving the stability, accuracy, and physical consistency of the prediction model. Attached Figure Description

[0034] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of the amorphous oxide semiconductor device characteristic prediction method provided in the embodiments of the present invention.

[0036] Figure 2 This is a schematic diagram of the Artificial Neural Network (ANN) model of this invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0038] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for predicting the characteristics of amorphous oxide semiconductor devices based on a hybrid dataset, comprising the following steps:

[0039] Step S1: Construct a mixed dataset of drain-source currents for amorphous oxide semiconductor devices; specifically:

[0040] S11 obtains the drain-source current data of amorphous oxide semiconductor devices with the same device size and voltage parameters; specifically:

[0041] S111 acquires a set of measured data for amorphous oxide semiconductor devices at different device sizes and voltage combinations. The measured data reflects the physical behavior of the devices during the actual manufacturing process, including device size parameters, voltage parameters, and measured drain-source current values. The device size parameters include gate length L and gate width W, and the voltage parameters include gate-source voltage Vgs and drain-source voltage Vds.

[0042] S112 In order to overcome the problems of insufficient data and incomplete parameter coverage caused by relying solely on measured data, this invention further constructs a device simulation model of amorphous oxide semiconductor on the TCAD platform and corrects it based on measured data.

[0043] First, an amorphous oxide semiconductor device structure model is constructed in the TCAD simulation software, defining the device dimensions, material regions, electrode positions, and mesh generation. Based on the device dimension parameters and voltage parameters obtained from the measured data in step S111, the corresponding simulated drain-source current values ​​are obtained. .

[0044] Furthermore, factors such as interface states, traps, and non-ideal charge transport behavior present in different device fabrication processes are often difficult to fully reproduce in TCAD, leading to deviations between simulated and actual currents. To reduce this deviation, this invention employs a domain adaptive mapping network, enabling the simulated data to approximate the distribution of measured data as closely as possible after mapping. Specifically, this embodiment utilizes a domain adaptive mapping network to map the simulated drain-source current values... Mapped to the measured value of drain-source current Drain-source current correction simulation values ​​with closer distribution .

[0045] For example, the domain adaptive network employs a multi-layer feedforward fully connected network, with its input being the simulated drain-source current value. The output is the simulated value of drain-source current correction. The loss function is and The mean square error.

[0046] In one embodiment, the simulated value of the drain-source current is also corrected. The process of performing sample screening and resampling based on physical consistency is as follows:

[0047] (1) Simulation values ​​of drain-source current correction The samples are filtered to remove physically unreasonable values, and the remaining samples in the ideal voltage parameter range with fewer simulated drain-source current correction values ​​are resampled or oversampled. The determination of physical unreasonableness is based on device physical constraints, which are whether the drain-source current changes with the gate-source voltage or drain-source voltage satisfy monotonicity, continuity, and non-negativity. If these constraints are not met, the device is physically unreasonable; otherwise, it is reasonable. The monotonicity condition is based on the measured drain-source current value of an amorphous oxide semiconductor device under a fixed drain-source voltage Vds. Monotonicity as a function of gate-source voltage Vgs;

[0048] (2) The simulated value of the drain-source current after the above steps are corrected. Logarithmic transformation is performed to reduce the impact of differences in current magnitude between different operating regions and different devices.

[0049] S12 obtains the extended simulation value of the drain-source current:

[0050] By fitting the simulated values ​​of each drain-source current correction, a simulation fitting curve is obtained. Interpolation or extrapolation is then performed on the simulation fitting curve to obtain a series of extended simulated values ​​of the drain-source current. .

[0051] S13. Correct the simulation value of drain-source current. Measured values ​​of drain-source current Drain-source current extended simulation values Process the data in a uniform format to construct a hybrid dataset.

[0052] Step S2: Preprocess and divide the mixed dataset to obtain a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to adjust the model's hyperparameters, and the test set is used to evaluate the model's prediction performance.

[0053] Step S3, construct as follows Figure 2 The artificial neural network model ANN is shown, and it is trained, validated, and tested using training, validation, and test sets, respectively.

[0054] In this embodiment, the Sigmoid or tanh function with nonlinear characteristics is selected as the activation function to increase the expressive power of the neural network model. The Sigmoid function has a value range between 0 and 1 (the tanh function has a value range between -1 and 1) to map the input data to a suitable range, which facilitates the learning of the neural network model.

[0055] The Adam optimizer is selected, which combines the advantages of adaptive learning rate and momentum to automatically adjust the learning rate during training and accelerate the convergence speed.

[0056] The Xavier initialization strategy is used to initialize the weights and generate a randomly initialized neural network model.

[0057] The initial learning rate is set to 0.001 and is automatically adjusted by the Adam optimizer during training.

[0058] Configure the number of samples per batch during training to perform batch processing, thereby improving training efficiency and reducing memory usage.

[0059] Based on the ANN model obtained from the above steps, the gate length, gate width, gate-source voltage, and drain-source voltage of the amorphous oxide semiconductor device are selected as the inputs to the constructed ANN model, and the drain-source current is used as the output of the ANN, thus forming the ANN model:

[0060] (1)

[0061] in, This is an ANN mapping function;

[0062] Specifically, the Artificial Neural Network (ANN) model uses the weighted mean squared error (MSE) as the loss function during training. The MSE measures the difference between the model's predicted values ​​and the true values, improving the model's fitting accuracy to the measured data. At the same time, it suppresses the influence of TCAD simulation bias, enhances the model's physical consistency and generalization ability, and significantly reduces the risk of model misleading due to TCAD bias. By minimizing the MSE, the model continuously adjusts its parameters to improve the accuracy of predictions.

[0063] The loss function is as follows:

[0064] (2)

[0065] in This represents the i-th weight. This represents the predicted drain-source current value of the i-th artificial neural network model (ANN). This represents the i-th drain-source current data in the mixed dataset;

[0066] when When the measured value of the drain-source current is given, , ;when When correcting the simulated value for drain-source current, , Indicates the curve-level physical error weights. Represents sample-level error weights; when When expanding the simulation value for drain-source current, , Represents the confidence coefficient. .

[0067] In this invention, in order to make the simulation data in the mixed dataset have a credibility distribution that conforms to the physical characteristics of amorphous oxide semiconductor devices, a curve-level weight based on the device physical mechanism is constructed for each simulation curve with corresponding measured data, and a sample-level weight is further constructed by combining point-by-point current error, thereby forming a hierarchical weighted data fusion strategy.

[0068] Through this weighted fusion mechanism, this invention establishes a hybrid dataset that can cover a wide range of voltage and size parameters globally while maintaining sufficient physical realism, providing high-quality input for stable training of ANN models.

[0069] One embodiment uses curve-level physical error weighting. The calculation process:

[0070] The simulated value of the drain-source current correction and the simulated value of the drain-source current extension are combined and denoted as the simulated value of the drain-source current. A simulated drain-source current curve is constructed based on the simulated value of the drain-source current, with the vertical axis representing the simulated value of the drain-source current and the horizontal axis representing the voltage parameter.

[0071] Based on the measured values ​​of the drain-source current, a measured drain-source current curve is constructed, with the vertical axis representing the measured drain-source current and the horizontal axis representing the voltage parameter.

[0072] For the measured and simulated drain-source current curves under the same device size and voltage parameters, the subthreshold slope values ​​are extracted respectively. Effective mobility in the linear region and saturation region channel modulation factor The corresponding physical deviation is calculated; this physical deviation reflects the systematic difference between TCAD simulation and the behavior of real amorphous oxide semiconductor devices in different working ranges.

[0073] Curve-level physical error weights are constructed based on three physical deviations. For example, exponentially decaying weights:

[0074] (3)

[0075] in, This is the physical sensitivity positive coefficient corresponding to the physical deviation, used to adjust the impact of different physical deviations on the mixed dataset; It is an exponential function;

[0076] In addition, common Gaussian weights, inverse proportional weights, Softmax weights, piecewise weights, and logistic weights can also be used as curve-level physical error weights based on three physical deviations.

[0077] For example, the physical deviation of the subthreshold slope value is calculated as follows: ,in This represents the subthreshold slope value of the measured value. This represents the subthreshold slope value of the simulated values ​​after sample selection and resampling correction.

[0078] For example, the physical deviation of the effective mobility in the linear region is calculated as follows: ,in This represents the effective mobility of the linear region as measured. This represents the effective mobility of the linear region after sample selection and resampling to correct the simulated values;

[0079] For example, the physical deviation of the channel modulation factor in the saturation region is calculated as follows: ,in This represents the measured channel modulation factor in the saturation region. This represents the saturation region channel modulation factor of the simulated values ​​after sample selection and resampling correction.

[0080] One embodiment uses sample-level error weights. The calculation process:

[0081] The simulated value of the drain-source current correction and the simulated value of the drain-source current extension are combined and denoted as the simulated value of the drain-source current. A simulated drain-source current curve is constructed based on the simulated value of the drain-source current, with the vertical axis representing the simulated value of the drain-source current and the horizontal axis representing the voltage parameter.

[0082] For the voltage parameters corresponding to the simulated drain-source current curve, the point-by-point deviation between the simulated and measured drain-source current values ​​is calculated, and sample-level error weights are further constructed. ;

[0083] The point-by-point deviation between the corrected simulated value and the measured value of the drain-source current is calculated as follows:

[0084] (4)

[0085] in This represents the measured value of the i-th drain-source current. This represents the corrected simulated value of the i-th drain-source current;

[0086] Calculation of sample-level error weights, such as exponentially decaying weights:

[0087] (5)

[0088] in, This is a preset positive coefficient.

[0089] In addition, common Gaussian weights, inverse proportional weights, Softmax weights, piecewise weights, and logistic weights can also be used as methods for calculating sample-level error weights.

[0090] In one embodiment, the confidence coefficient The calculation uses the interval average method:

[0091] (6)

[0092] in, express and Multiply by the largest weight, express and The weight that is minimized when multiplied.

[0093] In addition, common methods such as arithmetic mean, weighted mean, grouped mean, median, quantile confidence coefficient, variance-driven confidence system, and entropy confidence coefficient are also used as methods for selecting confidence coefficients.

[0094] After the model has completed iterative training, test whether the characteristic curves and derivative curves of the output model meet the expectations, and whether overfitting or underfitting has occurred.

[0095] Underfitting refers to a model having large errors on both the training and validation sets. This manifests as the predicted I-V curve failing to accurately follow the changing trend of the measured data, and the slope or shape of the characteristic curve deviating significantly from the expected physical behavior. It is usually caused by insufficient model complexity or insufficient training rounds. Underfitting can be improved by increasing the number of neural network layers or neurons, extending the training rounds, or adjusting the learning rate.

[0096] Overfitting refers to a model having a small error on the training set but a significant increase in error on the validation or test set. This manifests as oscillations in the model's output characteristic curve or its derivative curve, excessive noise amplification, or oversensitivity to local data points, disrupting the natural smoothness and monotonicity of the I-V curve in the subthreshold, linear, and saturation regions. Overfitting can be addressed by introducing weight decay, adding regularization terms, increasing the weight smoothing coefficient of the simulation data, using dropout, early stopping, or augmenting the training data.

[0097] Step S4: Input the device size parameters and voltage parameters of the amorphous oxide semiconductor device into the trained artificial neural network model ANN, and output the drain-source current prediction value; based on the voltage parameters of the amorphous oxide semiconductor device and the corresponding drain-source current prediction value, predict the electrical characteristics of the amorphous oxide semiconductor device under different device size parameters.

[0098] The electrical characteristics of amorphous oxide semiconductor thin-film transistors include, but are not limited to, the current-voltage relationship characterized by the transfer characteristic curve (Id–Vg) and the output characteristic curve (Id–Vd), as well as electrical parameters such as subthreshold slope, effective mobility, and channel modulation factor extracted from the above electrical characteristic curves.

[0099] In this embodiment, following the method described above, a B1500A parameter analyzer was used to test the electrical characteristics of the actually fabricated amorphous indium gallium zinc tin oxide thin-film transistor (IGZTO-TFT) device. The gate length L of the tested device ranged from several micrometers to tens of micrometers, and the gate width W ranged from tens of micrometers to hundreds of micrometers. Different gate-source voltages Vgs and drain-source voltages Vds were set, and the corresponding drain-source currents were measured. Complete I-V curves of the device in the subthreshold, linear, and saturation regions were obtained. All test samples were compiled into a format including Vgs, Vds, L, W, and Vds. The measured dataset provides a foundation for subsequent simulation calibration and hybrid data construction.

[0100] To expand the data range and improve the model's adaptability to different operating conditions, two-dimensional or three-dimensional device structure models of IGZTO-TFTs were constructed in Synopsys SentaurusTCAD software, including the gate insulating layer, semiconductor layer, source and drain metals, and related material parameters. Initial simulation currents were generated by setting trap state models, mobility models, and contact models. .

[0101] like Figure 2 As shown, due to the discrepancy between the simulation model and the actual device, this embodiment corrects the simulation data: firstly, it extracts data that simultaneously contains the measured current. With simulated current The dataset was used, and a domain adaptive mapping network was employed to learn the simulated current. To the measured current The mapping relationship (this network is independent of the ANN master model) makes the corrected simulated current The distribution is closer to the measured data. .

[0102] Subsequently, the monotonicity of Ids with respect to Vgs was checked by fixing Vds, and samples that violated physical laws were removed. For regions with few samples or sparse parameters, a resampling strategy was adopted to enhance data coverage. In addition, a logarithmic transformation was applied to the simulated current to reduce the differences between different current magnitudes, making the simulation data more balanced on a numerical scale.

[0103] Next, simulation curves are obtained by fitting the corrected drain-source current simulation values. Interpolation or extrapolation is then performed on these curves to obtain a series of extended drain-source current simulation values. Both the corrected drain-source current simulation values ​​and the extended drain-source current simulation values ​​are considered as the actual drain-source current simulation values.

[0104] The L, W, Vgs, and Vds of the mixed dataset were normalized, and all input parameters were mapped to the [0,1] interval using interval scaling to avoid training instability caused by differences in units. The dataset was then randomly shuffled and divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model training, the validation set was used to monitor the training process and adjust hyperparameters, and the test set was used to test the model's generalization ability on unseen data.

[0105] The model input consists of size parameters L and W and bias parameters Vgs and Vds, which are mapped to the output current Ids through a multi-layer fully connected neural network. The model uses either sigmoid or tanh activation functions, employs the Xavier initialization method, and introduces the Adam optimizer to improve training efficiency.

[0106] To emphasize the dominant role of measured data in the training process, the loss function adopts a weighted mean squared error form. In this embodiment, the weight of the measured samples in the loss function is... The weight of the calibration simulation sample is the product of the curve-level physical error weight and the sample-level error weight, and the weight of the drain-source current extended simulation value is the confidence coefficient c, so that the model pays more attention to the real device data during training.

[0107] During model training, not only are the current values ​​themselves used for supervision, but the output I-V curve and its derivative characteristics are also checked to ensure that the model output conforms to the physical laws of the device. If the model exhibits underfitting, manifested as large errors on both the training and validation sets, it can be improved by increasing network depth, the number of neurons, or extending the training epochs. If the model exhibits overfitting, manifested as a significantly higher error on the validation set than on the training set, or noise amplification and unreasonable oscillations in the output curve, it can be suppressed through regularization, dropout, weight decay, or early stopping mechanisms. Through this training mechanism, it is ensured that the model achieves a reliable level in both numerical accuracy and physical consistency.

[0108] By using a trained neural network model to predict arbitrary input size parameters and voltage bias, the drain-source current characteristics under different L, W, Vgs, and Vds can be quickly obtained. This prediction model can be used for device structure optimization, process parameter scanning, sensitivity analysis, and multi-objective design, thereby significantly improving device development efficiency and reducing experimental and simulation costs.

[0109] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for predicting the characteristics of amorphous oxide semiconductor devices based on a hybrid dataset, characterized in that, The method includes the following steps: Construct a mixed dataset of drain-source currents for amorphous oxide semiconductor devices and preprocess it; An artificial neural network model was constructed and trained using a preprocessed mixed dataset of drain-source currents. The device size parameters and voltage parameters of the amorphous oxide semiconductor device are input into the trained artificial neural network model, and the predicted drain-source current value is output. Based on the voltage parameters of amorphous oxide semiconductor devices and the corresponding predicted drain-source current, the electrical characteristics of amorphous oxide semiconductor devices under different device size parameters are predicted. The drain-source current hybrid dataset consists of measured drain-source current values, corrected drain-source current simulation values, and extended drain-source current simulation values. The measured and corrected drain-source current values ​​are drain-source current data for amorphous oxide semiconductor devices with the same device size and voltage parameters. The extended drain-source current simulation values ​​are obtained by extending the corrected drain-source current simulation values. During the training process, the loss function of the artificial neural network model is the weighted mean square error of the i-th predicted drain-source current output by the artificial neural network model and the i-th drain-source current data in the mixed dataset; where the weight value is set according to the measured drain-source current value in the mixed dataset. , `max` represents the maximum value function; when the i-th drain-source current data in the mixed dataset is the drain-source current correction simulation value, the corresponding weight value is... , Indicates the curve-level physical error weights. This represents the error weight of the i-th sample level; when the i-th drain-source current data in the mixed dataset is the extended simulation value of the drain-source current, the corresponding weight value is the confidence coefficient c.

2. The method according to claim 1, characterized in that, The process of obtaining the measured drain-source current is as follows: obtain a set of measured data for amorphous oxide semiconductor devices under different device sizes and different voltage combinations; wherein, the set of measured data includes device size parameters, voltage parameters, and measured drain-source current values; the device size parameters include gate length and gate width, and the voltage parameters include gate-source voltage and drain-source voltage.

3. The method according to claim 2, characterized in that, The process of obtaining the drain-source current correction simulation value is as follows: In the TCAD simulation software, an amorphous oxide semiconductor device structure model is constructed, the device size, material region, electrode position and mesh division are defined, and the corresponding drain-source current simulation value is obtained based on the device size parameters and voltage parameters in the measured data set; The drain-source current simulation value is mapped to a drain-source current correction simulation value that is closer to the distribution of the measured drain-source current value using a domain adaptive mapping network.

4. The method according to claim 1, characterized in that, The process of obtaining the simulated values ​​of the drain-source current spread is as follows: fitting the simulated values ​​of each drain-source current correction to obtain a fitting curve, and using the fitting curve for interpolation or extrapolation to obtain a series of simulated values ​​of the drain-source current spread.

5. The method according to claim 3, characterized in that, The domain adaptive mapping network adopts a multi-layer feedforward fully connected network.

6. The method according to claim 3, characterized in that, When the domain adaptive mapping network obtains the simulated values ​​of drain-source current correction, it is also necessary to perform sample screening and resampling based on physical consistency on the simulated values ​​of drain-source current correction. The implementation process is as follows: (1) Physically unreasonable values ​​are eliminated from the sample of the drain-source current correction simulation values. The judgment of physical unreasonableness is based on the physical constraints of the device. The physical constraints of the device are whether the drain-source current changes with the gate-source voltage or the drain-source voltage and satisfies the monotonicity, continuity and non-negativity. If it does not satisfy these conditions, it is physically unreasonable; otherwise, it is reasonable. (2) Perform logarithmic transformation on the simulated values ​​of drain source current after removing physically unreasonable values.

7. The method according to claim 1, characterized in that, The calculation process for the curve-level physical error weights is as follows: The simulated value of the drain-source current correction and the simulated value of the drain-source current extension are combined and denoted as the simulated value of the drain-source current. A simulated drain-source current curve is constructed based on the simulated value of the drain-source current, with the vertical axis representing the simulated value of the drain-source current and the horizontal axis representing the voltage parameter. Based on the measured values ​​of the drain-source current, a measured drain-source current curve is constructed, with the vertical axis representing the measured drain-source current and the horizontal axis representing the voltage parameter. For the measured and simulated drain-source current curves under the same device size and voltage parameters, the subthreshold slope, effective mobility in the linear region, and channel modulation factor in the saturation region are extracted, and the corresponding physical deviations are calculated. Based on the three physical deviations, curve-level physical error weights are constructed.

8. The method according to claim 1, characterized in that, The calculation process for the sample-level error weights is as follows: The simulated value of the drain-source current correction and the simulated value of the drain-source current extension are combined and denoted as the simulated value of the drain-source current. A simulated drain-source current curve is constructed based on the simulated value of the drain-source current, with the vertical axis representing the simulated value of the drain-source current and the horizontal axis representing the voltage parameter. For the voltage parameters corresponding to the simulated drain-source current curve, the point-by-point deviation between the simulated drain-source current value and the measured drain-source current value is calculated, and sample-level error weights are further constructed.

9. The method according to claim 1, characterized in that, The confidence coefficient It is half the positive difference between the curve-level physical error weight and the sample-level error weight.

Images