Leuven | More than two weeks ago
State-of-the-art GaN transistor models range from black-box models (using machine learning), with limited or no physical interpretation, up to physics-based compact models that give insight into the influence of scaling rules and technology parameters. The latter is of interest as to provides insight into scaling and gives feedback to technology researchers.
Current state-of-the-art compact GaN transistor models (e.g., the ASM-HEMT model) have a large number of physical and tuning parameters that need to be extracted from a large number of measurements, originating from various measurements setups. This includes DC, linear S-parameter, and nonlinear large signal RF measurements, while taking thermal and trapping effects into account. This model extraction can be seen as a machine learning problem as it has a large parameter- and dataset. However, neither the uncertainty on the extracted parameters, nor the correlation between different effects / parameters are quantified. This lack of uncertainty quantification also hampers the design of the test structures and measurements to be used to increase the accuracy of the model.
The PhD envisions to bridge the gap between (big-data) machine learning techniques and (statistics-based) system identification modeling to extract the large parameter set in GaN modeling from the big measurement dataset. This includes:
This
research is a collaboration between
Required background: Electronic Engineering and/or Computer Science (or equivalent)
Type of work: 30% theoretical challenge for merging machine learning and statistics-base system identification for GaN modeling, 30% Python implementation, 20% validation with circuit simulations, 20% experimental validation on GaN designs.
Supervisor: Gerd Vandersteen
Co-supervisor: Bertrand Parvais
Daily advisor: Gerd Vandersteen
The reference code for this position is 2025-037. Mention this reference code on your application form.