Neural network-based model of photoresist reflow

Abstract

When polymers are heated above their glass transition temperature, they enter a viscous rubbery state that allows the polymer to be reshaped in a process called “reflow.” The final shape depends on the material, substrate, the initial dimensions of the structure, the reflow temperature, and time and is mostly governed by energy minimization. Most empirical models so far have used linear regression to predict scalar parameters like the thickness of the reflowed structure but do not account for intermediate shapes. In this work, the authors measure the profiles of photoresist patterns subjected to various reflow conditions, complementing results in the literature. Using shallow neural networks, they develop models to predict the type of shape produced after reflow and its full cross-sectional height profile. These models can serve as an aid for polymer engineering and fabrication and also demonstrate the usefulness of a neural network-based approach to physical optimization problems without analytical solutions.