Cristina White, Stanford University

Photo of Cristina White

In machine learning, models are unknown and must be constructed from data, while in computational fluid dynamics models are known and well-defined by a set of partial differential equations. However, the solution – the distribution in space and time of fluid state variables such as density, velocity, momentum and energy – is unknown until a simulation of the model has converged, which can take days or weeks on a supercomputer for a single design. Therefore, automatically iterating through new designs for optimization often is infeasible. What if engineers could predict solutions in seconds, given the information they already have from the simulations they have already run? Many approaches have been tried; some involve models of the underlying physics, while others are model-free and make predictions based only on existing simulation data. We present a novel approach: We reformulate the prediction problem to effectively increase the size of the otherwise tiny data sets and we introduce a new neural network architecture called a cluster network. Compared to state-of-the-art model-based approximations, we show that our approach is nearly as accurate, an order of magnitude faster and vastly easier to apply. Moreover, our method outperforms previous model-free approaches.

Abstract Author(s): Cristina White, Daniela Ushizima, Charbel Farhat