Artificial Neural Network for Response Inference of a Nonvolatile Resistance-Switch Array.

Area of Science:

• Materials Science
• Computer Science
• Electrical Engineering

Background:

• Nonvolatile binary resistance switches are key components in emerging electronic systems.
• Crossbar arrays composed of these switches exhibit complex, nonlinear behavior.
• Accurate modeling of these arrays is crucial for device design and application.

Purpose of the Study:

• To develop and validate an artificial neural network for inferring the behavior of random crossbar arrays.
• To assess the performance of a multilayer perceptron (MLP) in capturing the input-output relationship of these arrays.
• To demonstrate the MLP's capability in handling nonlinearities inherent in resistance-switch arrays.

Main Methods:

• Utilized a multilayer perceptron (MLP) with leaky rectified linear units for behavior inference.
• Trained the MLP using a large dataset (500,000 or 1,000,000 examples) of resistance states and applied voltages.
• Employed supervised learning, using calculated current arrays as labels for training.
• Input vectors comprised resistance state distributions and applied voltage arrays for crossbar arrays of varying sizes (e.g., 28 × 27).

Main Results:

• Achieved a high correlation coefficient of 0.9995 between inferred and correct currents for a larger crossbar array.
• Demonstrated the MLP's ability to accurately predict the current flow based on resistance states and applied voltages.
• Validated the effectiveness of the MLP in modeling the nonlinear characteristics of the crossbar arrays.

Conclusions:

• The multilayer perceptron (MLP) is a versatile and effective tool for behavior inference in random crossbar arrays.
• The MLP successfully captures the quantitative linkage between input (resistance states, voltages) and output (current) in these complex systems.
• This approach offers a promising method for analyzing and designing advanced electronic devices utilizing nonvolatile resistance switches.