(84b) Large Scale Discrete Element Modeling (DEM) Using Rocky Multi-GPU Solver for Various Process Equipment | AIChE

(84b) Large Scale Discrete Element Modeling (DEM) Using Rocky Multi-GPU Solver for Various Process Equipment


Bharadwaj, R. - Presenter, ESSS North America
Almeida, L., ESSS
Remy, B., Bristol-Myers Squibb Co.
Pandey, P., West Virginia University
Chemical and Pharmaceutical industries have large manufacturing unit operations that involve complex physical processes that are not well understood. For example, there are different particle-level phenomenon at play simultaneously in processes such as granulation or tablet mixing/coating. Scale-up can thus be quite challenging during process development and is often done using large scale design of experiments, which can be difficult and expensive to accomplish. Predictive first-principle models such as the Discrete Element Method (DEM) not only provides insights into the manufacturing process but also reduce the number of trial-error experiments by predicting material behavior in these unit operations.

The DEM is the modeling of particle-level interactions on the computer, but in the past, this method was limited by hardware memory limitations and large simulation run times. In addition, particle shapes were also limited to be either single spheres or a cluster of spheres rigidly bonded together (‘glued’ sphere) as the contact detection algorithms were easier for sphere-sphere interactions. These limitations were overcome, for the first time, in the Rocky DEM software, which uses state of-the-art polyhedral shapes and distributed GPU solver technology. This presentation focuses on highlighting Rocky DEM’s unique shape and multi-GPU features to predict flows in processes such as a commercial scale tablet coater, high shear wet granulator, material crushers, and more.

Rocky DEM was used to model realistic tablet shapes using its polyhedral shape representation abilities in the simulation of a commercial scale (BFC-400) tablet coater (359 kg batch containing 250,000 tablets with 222 vertices each). In addition, fast multi-GPU solvers allowed for the prediction of tablet movements in the coater in reasonable compute times (approx. 3 days using three NVIDIA Tesla P100 GPU cards: 86x faster than an 8-core CPU).

High-shear wet granulation equipment, which is used in the manufacture of solids oral-dosage process to improve the bulk density and flow characteristics of a formulation, was also modeled in Rocky. The challenge of this work was to establish correct scale-up rules by achieving a process level understanding. Granulators at various scales (1 L, 10 L and 150 L) using both dry and wet placebo were simulated and the predictions of particle velocities, stress profiles and residence time distributions were compared with experimental data. The ability to combine the power of several GPU cards allowed the processing of large scale particle count simulations using particle sizes and distributions closer to reality, especially at the 150-L scale, where over 10 Million particles were used. Extensive post-processing analyses were also performed in order to enable direct comparison of particle volume fractions and velocity profiles. In addition, prediction of the stress profiles and collisional behaviors, which are impossible to obtain experimentally, increased process understanding and reduced the number of experiments during scale up studies.