(448b) Cyberinfrastructure Enabled Parallelization of Population Balance Models for Efficient Simulation of Granulation Processes

Cyberinfrastructure Enabled Parallelization of Population Balance
Models for Efficient Simulation of Granulation Processes

Anik Chaturbedi, Franklin Bettencourt, Srinivas Mushnoori, Subhodh
Karkala, Shantenu Jha, Marianthi Ierapetritou, Rohit Ramachandran 

Rutgers, the State University of New Jersey, Piscataway,
NJ, USA 08854

Particulate processes, in general, are
complex to understand and control efficiently. Granulation is one of the most widely
used particulate process and is used to produce commercial products such as
pharmaceuticals, catalysts, fertilizers etc. In granulation, multiple rate
processes, such as aggregation, breakage, liquid addition, nucleation,
consolidation, layering take place simultaneously. The incorporation of these
multi-dimensional physics is a challenge in any model formulation. Previous
research has shown that a population balance model (PBM) can be used to simulate
granulation processes and to predict the evolution of macroscopic properties of
the system such as particle size distribution (PSD), average composition and
porosity. Due to the multitude of dimensions present and the associated
integrals, population balance models can become computationally intensive. Moreover,
for modeling continuous granulation, additionally, spatial variation of
particle properties has to be incorporated in already complex population
balance models [1]. This increases the computational cost even more. To tune
these models for accurately simulating real systems and eventually move towards
real-time prediction and control of these processes we will need to massively speed
up these simulations. Some previous research has been done on the
parallelization of PBMs using either the MATLAB Distributed Computing Toolbox [2]
or Message Passing Interface (MPI) [3].

In this work, a 3D PBM (distributed solid
volume, lumped liquid and gas volumes) and two additional spatial dimensions
has been parallelized using a simpler Message Passing Interface (MPI) approach
and a more complex hybrid MPI+OpenMP approach. The speed up and scale up of
these parallelization techniques have been studied. Also, a pilot job system,
RADICAL-Pilot has been used to efficiently parallelize and distribute
computational work among different computational units in a National Science
Foundation-supported distributed computing resource, Stampede and Rutgers
University School of Engineering High Performance Computer (SOE HPC). Parallel
simulation of PBMs through RADICAL-Pilot will not only aid in more efficient
and user friendly implementation of these parallel PBMs but also enable
coupling of PBMs with micro-physics models such as discrete element models
(DEM) to increase the accuracy of PBMs and accurately capture the multi-scale
dynamics of complex particulate processes. All codes were compiled with the
‘–ofast’ option. Initial runs show a 40 times speedup for the hybrid code
compared to the serial one on 128 cores on 8 nodes. Also the hybrid code
resulted in 80% less memory usage compared to the MPI code, which presents an
advantage in simulating larger systems with same computing resources. This work
will eventually lead to the incorporation of these fast and more mechanistic
population balances in real-time applications such as control and optimization.

  References