DeepSim-HiPAC: Deep Learning High Performance Approximate Calculation for Interactive Design and Prototyping

Authors: Ahmed Al-Jarro (Fujitsu Laboratories Ltd), Serban Georgescu (Fujitsu Laboratories Ltd), Yasumoto Tomita (Fujitsu Laboratories Ltd), Kouta Nakashima (Fujitsu Laboratories Ltd)

Abstract: We present a data-driven technique that can learn from physical-based simulations for the instant prediction of field distribution for 3D objects. Such techniques are extremely useful when considering, for example, computer aided engineering (CAE), where computationally expensive simulations are often required. To accelerate this process, we propose a deep learning framework that can predict the principal field distribution given a 3D object. This work allows us to learn a system's response using simulation data of arbitrarily shaped objects and an auto-encoder inspired deep neural network that maps the input of the 3D object shape to its principal 3D field distribution. We show that our engine, DeepSim-HiPAC, can estimate field distribution for two distinctive applications: micro-magnetics design in computational electromagnetics (CEM) and interactive cooling systems design in computational fluid dynamics (CFD), several orders of magnitude faster, up to 250000X, than the native calculations and at a cost of low error rate.

Best Poster Finalist (BP): no

Poster: pdf
Poster summary: PDF


Back to Poster Archive Listing