dataFlow: A Data-driven Fluid Flow Solving Platform

ended

With the recent breakthrough of deep learning methods, we currenty see the advent of employing this methodology in the context of physical simulations. Such simulations are widely used in numerous industrial fields, starting from car and airplane manufacturers, over computer graphics and animations to medical blood flow simulations. The market for computer simulations is currently exceeding 15 billion USD world wide, with rising trends, and 3 billion spent in Europe alone. A significant fraction of these simulations focuses purely on solving various forms of the Navier-Stokes equations. While right now virtually all of these simulations use traditional solvers, we estimate than only a few years from now there will be a significant fraction of deep learning powered solvers. Thus, we are at the right point in time to lay the foundations for commercializing the technology of deep learning for fluid simulations. The goal of this PoC project is to develop a first commercial flow solver based on deep learning that can predict fluid flow solutions almost instantly using a pre-trained model. This project will enable the team of Prof. Thuerey to mature the algorithms developed as part of the ERC Starting Grant \realflow, and turn them into the basis of a marketable product. The initial models will be thoroughly tested and validated, in order to satisfy industrial requirements for reliability and accuracy. In addition, this PoC aims for establishing a platform for flow data collection, interface standards, and trained models. This platform will be developed in conjunction to the deep-learning powered flow solving application, and provide research connections and publicity in parallel to it.