Optimizing the Design of Nanofluidic Chips with Graph Reinforcement Learning A General Graph Attentional Framework Applied on Nanoscale Catalytic Reactor Systems

Abstract

Nanofluidic chips are devices where fluids are controlled at nanoscale. Langhammer Lab at Chalmers University of Technology researches nanofluidic chips for cataly sis reactions. On these nanoscale reactors, catalysts are placed to maximize some property such as the reactant conversion rate to product. However, to this point, no framework exists for optimizing the design of the group’s nanofludic chips. Chips are currently designed through laborious trial and error. In this master’s thesis, a framework based on reinforcement learning with graph attentional neural networks is presented and applied for optimizing the design of nanofludic chips. A reinforcement learning agent is trained on a reward system based on computational fluid dynam ics (CFD) and consistently outperforms simulated manual designs. Additionally, a considerably lighter reward system based on ant colony optimization (ACO) is developed for placing catalysts and forming channels. The ACO reward system is shown to be highly correlated with the CFD reward system, but requires some fur ther developement in order to achieve the same performance as the CFD reward system.