A differentiable deep learning approach for inverse optimization of hopper flows in particulate manufacturing

Understanding granular dynamics is essential for many industrial applications, yet significant challenges persist. The discrete element method allows for direct tracking of particle motions, but it suffers from high computational costs, in particular for inverse problems. Recently, machine learning has seen rapid development and brings new possibilities for tackling these challenges. In this work, a differentiable model designed for rapid prediction and inverse optimization of particulate processes is developed. The proposed method is used to improve the maximum discharge rate of hopper flows and automatically optimize the hopper shape based on the target discharge rate. Additionally, controlling the degree of mixing of two particle components is explored and further validated with experiments. The modeling outcomes demonstrate that the differentiable deep learning approach developed in this work can efficiently address inverse optimization challenges in particulate processes, providing a new tool for the design and optimization of particulate manufacturing processes.