Research On Modeling Routine For Metal Oxide Thin-Film Transistors Based On Artificial Neural Network

With their advantages of high carrier mobility and good uniformity,the metal oxide thin film transistors(MOTFTs)have a great application potential in the field of digital circuits.Its application range has expanded from display panel circuits to the system on panel(SOP),etc.The MOTFT technology is still in an emerging phase and the modeling method has been an important development direction in this field.Physical models need a long development period and have a quantity of physical parameters that need to be set.The table model needs amounts of data but the accuracy is limited by the measurement noise and interpolation methods.Hence,using dataoriented modeling methods to quickly incorporate the newly generated device data into circuit simulations has important significance.As universal approximators,the artificial neural networks can build a neural network model with high accuracy and good generalization in a very short time.Therefore,in this paper,swarm intelligence optimization algorithm is combined with artificial neural network to rapidly establish the MOTFT model that can be used in circuit simulation.Firstly,IV curves of the IZO TFTs with different aspect ratios are measured,and the neural network modeling for these devices is built.Limited memory BroydenFletcher-Goldfarb-Shanno(L-BFGS)method is applied for ANN to update the weights and biases as an optimization.L-BFGS is sensitive to the initial location of which all the initial weights and biases are randomized.Therefore,a hybrid algorithm consisting of particle swarm optimization(PSO)and L-BFGS based on ANN is proposed in this work.Besides it,a mutation strategy for PSO is derived to enhance the searching ability and accelerate convergence further.The modeling result shows that this hybrid modeling method has the benefits of rapid fitting from the L-BFGS algorithm and universal searching ability from PSO.Subsequently,a simulation program for MOTFT devices is coding by using Verilog-A programming language,and this model is embedded into Cadence Spectre.Various types of inverting circuits are statically simulated.Compared with the measurements,the simulation result shows that this model has a good performance.In addition,the model that can be used for transient simulation is trained by ANN with the capacitance data,and the transient simulations are performed using a basic inverter and a scan driver.The excellent agreement between predictive values and measured ones proves that our hybrid modeling method can be used in MOTFT modeling and circuit simulation.