“Machine learning for thin film growth and microelectronics: Exploring process optimization, self-driving reactors, and application-driven codesign”
In this presentation, I will explore how machine learning can help accelerate the development, optimization, and integration of novel materials. This is a critical need in applications such as microelectronics and decarbonization. In particular, I will focus on three different areas: First, at the process development and manufacturing scale, I will explore how machine learning can accelerate the optimization of thin film growth processes based on atomic layer deposition. Here we have pursued two complementary approaches: we have demonstrated the use of deep neural networks to carry out few-shot and one-shot optimization, and we have also experimentally realized a self-driving ALD tool capable of optimizing an ALD process without a human in the loop. This results on a x100 speed up of the process optimization. Second, at the materials design scale, we are developing new machine learning techniques to realize application-driven co-design in microelectronics. In particular, I will introduce a backpropagation through hardware approach that can be used to optimize architectures for AI at the edge for smart sensing applications. As an exemplar, we have explored the optimization of architectures incorporating tunable resistive coatings to enable computing and sensing under extreme environments, focusing on RF signal processing applications. Finally, in the structure-property relation space, I will briefly describe how we are generalizing effective medium approximations to incorporate realistic constraints such as fluctuations in composition of composite phases or uncertainty in the properties of the individual phases or the interfaces between components. We are currently applying these models to understand the intrinsic limitations of advanced conductors based on metal-carbon composites.
