UCLA

Carbon Nanotube has long been seen as a promising candidate for high-performance electronic material, yet its unique 1D structure leads to challenges in device fabrication. Many processing approaches have been proposed to produce better performing CNTFETs and this explosion of data needs an efficient way to explore. In this thesis, I explored the use of several machine learning techniques, including neural networks, simulation-based inference, and generative flow networks, on predicting CNTFETs performance, probing the conductivity properties of CNT network, and generating CNTFETs processing information for target performance. In the beginning, I built up a neural network model for CNTFETs. I begin my work with simple cases where only certain continuous parameters like gate length are considered and developed a data cleaning method. It was shown that neural networks can work as a model for CNTFETs and reasonably perform as a device predictor for symmetric field effect transistors. I’ve also developed a neural network model that can incorporate processing information using encoding technique. The model can predict the performance of CNTFETs with various choices of processing methods and material combinations. At the same time, I explored the conduction properties of non-aligned CNT networks. I built up a compact model for CNTFETs built on non-aligned CNT networks and used simulation-based inference to extract key parameters to fit the model to the experimentally observed data since extraction is impossible through traditional methods. The model with extracted parameters can fit well with the observed data. We show that simulation-based inference can be a powerful tool for building models in cases where a distribution, rather than a certain value, will be the result. In the last step, I developed a generative model to generate device performance with target current performance. I first built a model to generate three key parameters and built the research on a compact model. The results show that this model can successfully generate multiple solutions that meet the goal. I’ve further developed a generative model that can generate device processing information at the same time. Though further improvement will be needed, some of the targets are met. I hope my work can show the ability of machine learning to solve some of the material science problems. Neural network can be a good function approximator for experimental observations, though it doesn’t provide understanding of the phenomenon. If probing of mechanism will be needed, simulation-based inference can be a good way to test human-created models and automatically generate parameters that humans can compare with experimental observations later. This is especially useful when the experiment input or result is a random variable described through the probability mass function or the probability density function. Generative models might be a way for experimental optimization, especially for engineering works like device fabrication, which usually requires testing out different combinations of parameters.