This blog entails my journey of reproducing the research titled *"SiC MOSFET with Integrated SBD Device Performance Prediction Method Based on Neural Network."*
From MOSFETs to neural networks, we'll explore the Silvaco implementation, progressing from hard coded values to a more generalized approach. We'll obtain graphs and then finally utilize the results in a neural network architecture.
My name is Mohammad Owais. I am a third year student of electronics and communication engineering,
passionate about machine learning?! Yes I am,
passionate about electronics?! Yes I am.
Struggling disastrously in both? Oh dear god T_T - an obvious yes!
I love diving into cool machine learning and deep learning research, and I also love learning about electronics. But that doesn’t stop me from being a complete douche in both. I struggle, I learn, I forget, I feel envious of others being better than me, I relearn that’s how the cycle goes.
I started machine learning during the summer break of my first year, loved it because of the math, and my deep rooted disdain for the computer engineering gizmos grinding their way into data structures so that they become more hire-able, mehh…..
Now I chose electronics or was forced into it, some might say. At first, I was overwhelmed by the sheer amount of physics behind everything, from diodes to transistors, BJTs, MOSFETs, and beyond. But over time i made peace with it and set my sights on merging my knowledge of machine learning and electronics into something meaning full
What follows is simply my perspective. I’m no electronics prodigy, but I’ll do my best to break things down as clearly as I can.
Alright let’s dive in!
The paper explores using neural networks and machine learning to predict both the static and dynamic characteristics of a SiC MOSFET integrated with a Schottky barrier diode.
The SBD-MOSFET devices are modeled and simulated using Sentaurus TCAD software, resulting in the generation of 625 sets of device structures and sample data, which serve as the sample set for the neural network.
Four input parameters are utilized for studying the behavior of the device, namely P-Well Doping Concentration, Width of JFET, Width of SBD, and P-Well Depth, to obtain four kinds of outputs for each simulated device. These outputs include breakdown voltage, threshold voltage, specific on-resistance, and switching power dissipation.
The proposed network architecture, grounded in deep learning principles, contains four different modules: the input layer, feature expansion module, feature extraction module, and output layer.
The goal is to utilize machine learning and deep learning for device parameter prediction instead of using high compute simulation software which consumes a lot of time
Now, I did not fully implement the paper as it was, i had to alter some of the proposed ideas, will state it down for clarity:-