A Prediction Model Of Total Dose Irradiation Effect Based On Artificial Neural Network

As a component of aerospace equipment,metal oxide field effect transistor has very important research significance in the field of aerospace.It will be affected by a variety of irradiation effects such as total dose effect,single particle effect,displacement damage effect in the space environment Among them,when the total dose effect is serious,the function of the device will be completely lost and its cumulative effect on the device is irreversible,which is the most harmful of all irradiation effects.Traditional physical modeling has the disadvantages of complex process,time-consuming,the need to obtain device-related process parameters,and difficulty in establishing the model.Common analysis models such as the BSIM3V3 model also have the disadvantages of complex reference process and high time cost.In order to solve the above Problem,this article uses a behavioral modeling technique based on artificial neural network,the focus is divided into two parts of the normal model and the irradiation model,the specific content is as follows:1.Under normal conditions,a DC bias model is established in the saturation region and the linear region,which is used as the basis of other models,mainly to establish a mapping model between different gate-source voltage and drain-source voltage to drain-source current.In this paper,the optimized multi-layer feedforward neural network model is used to make the modeling process strip off the complex parameter extraction and fitting process,and the nonlinear fitting ability of the neural network in high-dimensional space is used to train the network algorithm to make the prediction results and The actual results are fit,and the optimization of the network has four main points: the initial weight optimization based on genetic algorithm,the optimization of the number of hidden layers of the adaptive network combined with the empirical formula,the dynamic optimization of the learning rate combined with the LM algorithm,and the data set Verification division.2.Under the irradiation state,gradually study from the data behavior layer to the device bottom layer parameters,a total of five models are established,namely the irradiation state bias DC model,threshold voltage model,"two-state model" and parameter model,which are reflected in Under the dose effect,the intrinsic parameter value of the device changes.Firstly,the bias DC model is established.Under the normal state-based bias model,the input parameters are added to one dimension as the dose change parameter.The bias DC training network in the irradiated state is called the first neural network;the threshold voltage The main function of the model is to predict the degree to which the total dose effect of irradiation affects the threshold voltage drift.This paper proposes a new threshold voltage extraction method,which is different from the traditional threshold voltage extraction methods,such as constant current method,linear extrapolation method,And the second-order derivation method,which have the advantages of high accuracy and short training time;after the establishment of the threshold voltage model,the mid-voltage method is used to separate the trap charge generated by the oxide layer and the interface state of the radiation-influencing device to establish the "two state" Model ";Finally,a simplified model is used to establish the equivalent parameter model of the device affected by irradiation.Finally,the accuracy of the model is verified by comparing the experimental measured data.Under normal conditions,the transfer characteristic curves of the two groups of control groups under different bias conditions are predicted separately,and the extrapolation ability of the model is also evaluated.The error of the final model is within one thousandth,the training time is 12s;under irradiation,Respectively,to verify the five major models,and compare the prediction results of the DC bias and threshold voltage models with the actual experimental data.The final results show that the error is within 1%,and the model time is shorter than the traditional model."Two-state model " And the parameter model is compared with the literature,and finally shows that the prediction results are consistent with the actual physical process.In summary,the modeling method in this paper has the following advantages: the modeling process is less dependent on the underlying process parameters of the device,the model input and output parameters are adjustable and scalable,the model accuracy is verified by experimental data,the model fitting ability is strong,and training The time is short,and the actual irradiation measurement is no longer needed after the modeling is completed,thereby shortening the test cycle,which can provide a certain reference significance for the device total dose modeling and radiation resistance reinforcement.