COMPUTATIONAL FLUID DYNAMICS

Brainchip has demonstrated the use of Regression Analysis using AKIDA technology for regression analysis. In this linked video Brainchip's Todd Vierra presents this application running at 200 fps at less than 1 mw:


The fact that AKIDA technology has this capacity fits nicely with the work of one of Brainchip's research engagements with TATA Consulting Services
which produced the following paper exploring the potential use of Spiking Neural Networks for Computational Fluid Dynamics.

What are Computational Fluid Dynamics in simple terms:

2. Stabilization and Acceleration of CFD Simulation by Controlling Relaxation Factor Based on Residues: An SNN Based Approach

Mithilesh Maurya, Dighanchal Banerjee, Sounak Dey, Dilshad Ahmad

TCS Research, India

{mk.maurya|dighanchal.b|sounak.d|dilshad.ahmad}@tcs.com

Abstract

Computational Fluid Dynamics (CFD) simulation involves the solution of a sparse system of linear equations. Faster convergence to a physically meaningful CFD simulation result of steady-state physics depends largely on the choice of optimum value of the under-relaxation factor (URF) and continuous manual monitoring of simulation residues. In this paper, we present an algorithm for classifying simulation convergence (or divergence) based on the residues using a spiking neural network (SNN) and a control logic. This algorithm maintains optimum URF throughout the simulation process and ensure accelerated convergence of the simulation. The algorithm is also able to stabilize and bring back a diverging simulation to the converging range automatically without manual intervention. To the best of our knowledge, SNN is used for the first time to solve such complex classification problem and it achieves an accuracy of 92.4% to detect the divergent cases. When tested on two steady-state incompressible CFD problems, our solution is able to stabilize every diverging simulation and accelerate the simulation time by at least 10% compared to a constant value of URF.

1 Introduction

System of linear equations governing the model of solid-fluid interactions in Computational Fluid Dynamics (CFD) domain are usually solved using iterative methods [6, 19, 7] to reach to an acceptable numerical solution. This iterative process is time consuming, compute intensive, and the solution does not guarantee convergence to a physically meaningful solution for a given set of simulation parameters. This results into huge waste of man hours and computation resource. Though controlling factors such as under relaxation factor (URF) and convergence indicators such as simulation residue history (SRH) are tuned to ensure convergence, but it requires manual intervention to select suitable values. Thus, an automatic monitoring mechanism of residue history (to interpret convergence or divergence) and a subsequent control logic to auto-tune the URF would help in: (i) stabilising a diverging simulation, (ii) reaching faster convergence by accelerating converging simulation.

Fuzzy logic based approaches [15, 3] and Expression based methods [12, 20] have been applied to calculate & control URF but they tend to increase the computation time and are not applicable across all classes of CFD problems. RL based acceleration methods [13] for CFD simulations are computationally expensive and slow. Looking at the compute-intensive nature of simulation and to cater the requirement of early detection of divergence from as less data as possible, use of Spiking Neural Networks (SNN), a 3rd generation ML framework inspired from functionalities of mammalian brain, can be a good choice to reach to a useful solution [14]. SNNs are comparatively faster to learn from sparse data and are far less compute & power intensive. SNNs achieve this through asynchronous event handling and co-location of memory and computation [1, 4, 22].

Machine Learning and the Physical Sciences workshop, NeurIPS 2022.

In this paper, we propose an improved and robust solution to stabilize diverging CFD simulations and accelerate the converging CFD simulation by keeping the URF to its optimum value. The residual value over time is treated as a continuous time series. Current data of a fixed size window from this time series is passed to the SNN to be classified either as diverging or as converging. Depending on the classification, a control logic is used to change the URF. The classifier and control logic work together without interfering with the CFD simulation i.e. the process is non-invasive and works in parallel to the simulation. We demonstrate the efficacy of our approach via two steady-state incompressible CFD problems namely Backward Facing Step and Flow inside Tundish. We verified our findings by running the algorithm with OpenFOAM [18] and Ansys Fluent [11] simulators. We found that (i) the SNN is able to detect the divergent cases with 92.4% accuracy (with window size = 30), (ii) our solution is able to stabilize each and every diverging simulations and finally, (iii) it is also able to accelerate the simulation time by at least 10% compared to a fixed URF value…

Conclusion and Future Works. Our proposed solution is generic and can be integrated with any CFD tools. This automated process saves a lot of manual effort and time thereby making it beneficial for many industry scale CFD problems that run for days. Moreover, the SNN model can easily be retrained with additional data to improve the accuracy and to cater to different classes of problems. In future, we intend to improve the existing fixed control logic to an adaptive URF controller varying with the SR value. We also intend to test the performance and power consumption of the SNN by running it on a real neuromoprhic hardware such as Intel Loihi [2, 9] or Brainchip Akida [1].