SC19 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Poster 71: AI-Solver: Uncertainty in Prediction and Error Estimation for AI in Engineering


Authors: Ahmed Al-Jarro (Fujitsu Laboratories of Europe Ltd), Loic Beheshti (Fujitsu Laboratories of Europe Ltd), Serban Georgescu (Fujitsu Laboratories of Europe Ltd), Koichi Shirahata (Fujitsu Laboratories Ltd), Yasumoto Tomita (Fujitsu Laboratories Ltd), Nakashima Kouta (Fujitsu Laboratories Ltd)

Abstract: The AI-Solver is a deep learning platform that learns from simulation data to extract general behavior based on physical parameters. The AI-Solver can handle a wide variety of classes of problems including those commonly identified in FEA, CFD and CEM, to name a few, with speedups of up to 250,000X and extremely low error rate of 2-3%. In this work, we build on this recent effort. We first integrate uncertainty quantification, via exploiting the approximation of Bayesian Deep Learning. Second, we develop bespoke error estimation mechanisms capable of processing this uncertainty to provide instant feedback on the confidence in predictions without relying on the availability of ground truth data. To our knowledge, the ability to estimate the discrepancy in predictions without labels is a first in the field of AI for Engineering.

Best Poster Finalist (BP): no

Poster: PDF
Poster summary: PDF


Back to Poster Archive Listing