Researchers from Stanford University and SLAC National Accelerator Laboratory have published a deep-learning method for extracting physical parameters from electrical measurements of two-dimensional transistors. The work focuses on turning measurement data into device parameters automatically, rather than relying on conventional curve-fitting workflows.
According to the paper, the approach extracts quantities such as mobility, Schottky contact barrier height, and defect profiles. These are then used for automated parameter extraction and for technology computer-aided design, or TCADtechnology computer-aided designSoftware-related method or tool for development, control or data processing., model fitting. The authors describe the method as a neural-network-based alternative to classic fitting routines that can be slower and more sensitive to the choice of starting conditions.
To make the training practical, the team uses a simple data-augmentation strategy and pretraining. A secondary neural network is trained to approximate a physics-based device simulator, which provides the basis for the learning workflow. This links the model more directly to the underlying transistor physics while keeping the extraction process automated.
The paper shows how deep learning can be used as a fitting tool for 2D transistor characterization, with the goal of obtaining consistent parameter values from electrical measurement data. The publication is titled Deep Learning to Automate Parameter Extraction and Model Fitting of Two-Dimensional Transistors.