Reuter, Maximilian Jürgen (2024)
Data Driven Compact Modeling of a Reconfigurable FET.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00028629
Ph.D. Thesis, Primary publication, Publisher's Version
Text
diss_maxr_rfet_tuprints_final_submission.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (14MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Data Driven Compact Modeling of a Reconfigurable FET | ||||
Language: | German | ||||
Referees: | Hofmann, Prof. Dr. Klaus ; Becker, Prof. Dr. Jürgen | ||||
Date: | 16 December 2024 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xx, 153 Seiten | ||||
Date of oral examination: | 28 October 2024 | ||||
DOI: | 10.26083/tuprints-00028629 | ||||
Abstract: | This work contributes to predictive circuit simulation of emerging semiconductor devices by proposing two approaches to obtain characteristic device data from technology simulation, resulting in table models. From these intermediate table models, two different machine learning approaches provide candidates of machine learning based compact models. The methods are demonstrated for an ambipolar transistor, called the planar RFET, which allows dynamic reconfiguration of channel polarity at runtime through an additional gate electrode. As a first approach, a cluster simulation tool PyTaurus is proposed and allows efficient setup, simulation and refinement of table models from a factorial setup of DC parameter sweeps. To reduce model building time, PyTaurus provides a cluster simulation functionality and distributes the simulation deck to available computation nodes. A second approach, pseudo transient simulation, covers the bias space of the device under test within a single slowly proceeding transient simulation. Nesting of input voltages is achieved by a systematic setup of frequencies, that drive the respective harmonic electrode voltage signals. The resulting data sets satisfy structural constraints for table models in Verilog-A. Beyond the use as plain table models, however, the obtained data sets are transformed to predictive compact models using machine learning. The high dynamic range, that the drive current of the planar RFET features throughout the operating regions, leads to a solution with an ensemble model. A linear model focuses on high drive currents, while a second model is trained on logarithmically transformed current samples to provide accuracy into the regime of leakage current. For the respective models, two approaches are evaluated: The deep learning approach leads to multilayer perceptrons, which are then sequentially implemented in Verilog-A. The predictions are obtained through inference of the compact model voltages in each simulator step. As a second approach, symbolic regression is employed to optimize an analytical model without structural constraints. The obtained closed form expressions can directly be implemented in Verilog-A. In addition to the DC drive current model, a transient model is formed by symbolic regression, exclusively, as the dynamic range and the expected complexity of the charge model are lower than with drive current. The resulting neural network-based models show improvement over table models when it comes to DC simulation of digital cells. Transient simulation for timing characterization leads to similar accuracy than the table models, with a maximum deviation of 5.1% from the TCAD reference. Neural network-based models accelerate the simulation with a factor of up to 17x. Symbolic regression-based drive current models further improve computational efficiency, but show insufficient accuracy for predictive circuit simulation. The concluding result of this work is that it is recommended to transform a table model into a neural network-based compact model using the proposed data driven approach, which requires minimal domain knowledge. Symbolic regression is successfully employed for modeling of electrode charges, but accurate drive current modeling requires further work. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-286299 | ||||
Classification DDC: | 600 Technology, medicine, applied sciences > 621.3 Electrical engineering, electronics | ||||
Divisions: | 18 Department of Electrical Engineering and Information Technology > Institute of Computer Engineering > Integrated Electronic Systems (IES) | ||||
Date Deposited: | 16 Dec 2024 13:04 | ||||
Last Modified: | 18 Dec 2024 09:35 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/28629 | ||||
PPN: | 524685290 | ||||
Export: |
View Item |