AI powder mixing simulation developed by OMU: 350 times faster than conventional methods with hopes of homogenizing pharmaceuticals and batteries

A research group led by Graduate Student Naoki Kishida, Associate Professor Hideya Nakamura, Associate Professor Shuji Ohsaki, and Professor Satoru Watano of the Graduate School of Engineering at Osaka Metropolitan University (OMU) has developed a new artificial intelligence (AI)-based simulation method that can calculate the mixing of powders used as raw materials for pharmaceuticals and batteries 350 times faster than conventional methods. The method is expected to realize the long-time simulation required at manufacturing sites, which has not been achieved to date. The results were published in the online preliminary version of Chemical Engineering Journal.

Powder mixing, in which two or more types of powders are mixed for homogenization, is used in various applications from pharmaceuticals to food, batteries, electronic components, and ceramics. However, the uniform mixing of powders relies on trial and error and the experience of engineers. The discrete element method (DEM) was developed as a simulation method to accurately predict powder mixing. However, because this method requires long calculation times, it cannot predict long-time powder mixing, which is used in actual manufacturing processes.

The research group has developed a new method called recurrent neural network with stochastically calculated random motion (RNNSR) to calculate particle motion with high accuracy even over long time spans. RNNSR is an AI-based model built by learning extremely short-term particle motion patterns calculated using DEM. It can predict particle motion behaviors over time spans that are tens of thousands of times longer than that achieved by conventional methods. The macroscopic movement of powders is predicted using a regression neural network, and the random movement of microscopic particles is predicted using a probability distribution model so as to accurately learn and predict the movement behavior unique to powders. The calculation speed of RNNSR is approximately 350 times faster than that of DEM. Moreover, RNNSR maintains the same accuracy as DEM.

In the future, the research group expects to achieve advancements enabling the prediction of more complex powder-mixing behaviors and handling of heterogeneity (variations in the physical properties of individual particles), which is a characteristic of real powders.

Journal Information
Publication: Chemical Engineering Journal
Title: Development of ultra-fast computing method for powder mixing process
DOI: 10.1016/j.cej.2023.146166