3.7 Global Optimization Methods for SPICE Model Parameter Extraction
Another interesting approach is to substitute a lengthy parameter search for a simulation model by reversing the task. Instead of matching simulation to measurement results by tuning the simulation parameter, extract simulation parameter from measurement results. The goal is to find simulation parameters for a semiconductor (BSIM4 MOSFET model)
Project details:
- Previous method:
- match simulation to measurement results by tuning simulation parameters
- 34 different optimization steps
- tuned by skilled workers (PhD)
- ML method:
- extract simulation parameters from measurement results
- training data can be generated by running simulations
- simulation run time was short
- 10s of thousands of training data can be generated
- validation of predictions by simulation
- run simulation with predicted simulation parameters
- compare simulation to measurement results
The previous method required an educated comparison of simulation and measurement results, based on the comparison the simulation parameter were tuned.
Previous method, figure from https://www.iee.et.tu-dresden.de/iee/eb/forsch/AK-Bipo/2018/3__MachineLearningParameterExtraction_Pieper.pdf
In the ML method measurement data are converted to a set of features which are then used as input of the machine learning model. The simulation parameters can then be calculated in one step. Validation of the result can be achieved by running the simulation with the predicted simulation parameters.
ML method, figure from https://www.iee.et.tu-dresden.de/iee/eb/forsch/AK-Bipo/2018/3__MachineLearningParameterExtraction_Pieper.pdf