☕️ Cloud Chronicles #6: Building a Cloud-Native HPC Bioreactor Simulation
One of my friends complains about the long time it takes to train neural network models: “the data set is huge, even this supercomputer takes over 6 hours to run a new training cycle” — a statement also containing a tiny sliver of excitement. The training process is often started late in the evening, so that the new model can be analyzed when arriving at work in the morning.
The use of machine learning (ML) is also becoming more prevalent in the field of chemical engineering because it can speed up existing simulation calculations. There are several benefits and downsides to it. Below is a brief summary of the key factors playing a role:
It is too computationally intensive to optimize processes conditions for multiple inlet streams. As a result the calculations are simplified and modeled as a single representative inlet stream.
The problem is that such simplification is inaccurate because it leads to under- or over-estimations of the optimal processes performance. It is possible that much of world’s industrial process output can be made 15%-30% more efficient.
Faster models and algorithms, or more powerful computing capabilities such as HPC, are needed to more accurately optimize processes with multiple inlet streams in the context of the whole manufacturing process.
Machine learning is successfully used to reduce computing time from hundreds of hours to few minutes, and can be used for complex optimization tasks that otherwise would be computationally too expensive.
For more complex situations, the output of these models can become quite sensitive to error, especially when using inputs closer to boundary conditions.
To understand these problems a little bit better, I’m going to do a couple experiments and document my learnings.
The ultimate outcome of this exploration is to take a complex process simulation and reduce the simulation time from 100 hours to close to real-time (i.e. seconds). This would result in much faster engineering cycles and more optimized processes.
Image: CFD is commonly used to understand how fluids flow, such as to optimize the design of a bioreactor:
