A deep-learning approach to realizing functionality in nanoelectronic devices.

Area of Science:

• Nanoscience and Nanotechnology
• Computational Physics
• Machine Learning

Background:

• Optimizing nanoscale devices is challenging due to increasing complexity and device variations.
• Traditional physics-based models struggle with imperfections and large numbers of parameters.
• Deep neural networks (DNNs) are powerful predictive tools but underutilized for device optimization.

Purpose of the Study:

• To propose a generic deep-learning approach for efficient optimization of complex, multi-terminal nanoelectronic devices.
• To demonstrate the application of this approach for realizing desired functionality in disordered silicon dopant atom networks.
• To enable fast, in situ optimization of (quantum) nanoelectronic devices.

Main Methods:

• Modeling device input-output characteristics using a deep neural network (DNN).
• Optimizing control parameters within the DNN model via gradient descent.
• Applying optimized control settings to the physical device to achieve predicted functionality.

Main Results:

• The DNN accurately modeled the behavior of a disordered network of dopant atoms in silicon.
• Gradient descent optimization of the DNN successfully identified control parameters for specific tasks.
• The physical device, when tuned using DNN-predicted settings, exhibited the desired functionality.

Conclusions:

• The proposed deep-learning approach enables efficient optimization of complex nanoelectronic devices.
• This method overcomes limitations of physics-based models in the presence of device imperfections.
• The approach facilitates fast, in situ tuning for desired device functionality, applicable to quantum devices.