Machine learning-based approach for automatically tuned feedback-controlled electromigration

Abstract

Feedback-controlled electromigration (FCE) has been employed to control atomic junctions with quantized conductance. An FCE scheme is controlled by many parameters, such as the threshold differential conductance GTH, feedback voltage VFB, and voltage step VSTEP. It is considered possible to achieve a precise and stable control of the quantized conductance by automatically optimizing the FCE parameters. This motivated us to develop an approach based on machine learning (ML) to tune the feedback parameters of FCE. The ML system is composed of three kinds of engines, namely, learning, evaluation, and inference. The learning engine performs the FCE procedure with random parameters, collects various experimental data, and updates the database. Subsequently, four variables and a cost function are defined to evaluate the controllability of the quantized conductance. The evaluation engine scores the experimental data by using the defined cost function. Then, the control quality is evaluated in real time during the FCE procedure. The inference engine selects the new FCE parameter according to the evaluated data. These engines determine the optimal parameters without human intervention and according to the situation. Finally, we actually applied this system to the FCE procedure. The parameter is selected from sample data in the database according to the variation in controllability. As a result, the controllability gradually improves during the FCE procedure that uses the ML system. The results indicate that the proposed ML system can evaluate the controllability of the FCE procedure and change the VFB parameter in real time according to the situation.