Reduced Order Modeling using TensorFlow

Sunil
6 min readMay 18, 2017

While in grad school, I worked on an unsupervised machine learning (ML) problem with computational fluid dynamics (CFD) data (link to the paper and the journal article). The singular value decomposition (SVD) based learning algorithm was written in C++ and ran on the CPU. I was recently recalling this work and decided to re-implement it using TensorFlow. The goals of this post are to write about my experiences with using TensorFlow for :

  1. training an SVD based unsupervised learning ML model using TensorFlow
  2. deploy the trained model with TensorFlow serving.

CFD Data and reduced order modeling

The concept behind CFD is the numerical solution of Navier-Stokes equations on a discretized grid. The N-S equations are coupled partial differential equations that require boundary and initial conditions to be solved numerically. There are several solution methods that have been developed over the last few decades with researchers and practitioners taking advantage of advances in hardware and software to compute accurate solutions faster.

The outcome of numerically solving these equations are datasets that depict the flowfield, characterized by the x, y and z components of velocity i.e. (u,v,w) and the pressure (things get more complex when there are other physical phenomena involved). The time-to-solution is usually of the order of several minutes to days, depending on the complexity of the problem. Hence the need for reduced

--

--