New AI Project AlphaNumerics Zero

The Helmholtz Artificial Intelligence Cooperation Unit (Helmholtz AI) is one of five platforms initiated by the Helmholtz Information and Data Science Incubator. Its main goal is to become a driver for applied artificial intelligence (AI) through the development and distribution of AI methods across all Helmholtz centres, effectively combining AI-based analytics with the Helmholtz Association’s unique research questions and datasets.

In the first annual call for proposals, the proposal “AlphaNumerics Zero” (ɑN0), led by JSC together with researchers from the Steinbuch Centre for Computing at KIT, was selected for funding. The objective of this project is to rethink numerical methods on high-performance computers. Traditionally, a lot of effort goes into the design, implementation, and optimization of solvers for differential equations, numerical libraries, etc. Furthermore, performance engineering is necessary to scale simulation codes onto supercomputers. New ideas are urgently needed as developing methods for the upcoming extreme-scale supercomputers is becoming increasingly challenging. Our goal is to use reinforcement learning techniques so that the computer learns the on-average optimal numerical solution method for a given simulation problem by itself. The deliverables for the project will be a working framework and a demonstration in one application case using a time-dependent partial differential equation. This is a moonshot project that – if successful – will be the first step to changing the paradigm of how numerical simulations are designed and performed on extreme-scale computers. It would have an impact on all research fields that rely on numerical simulations.

The project is funded by the Helmholtz Association’s Initiative and Networking Fund (INF) with nearly € 200,000 and runs from August 2020 until January 2023.

Contact: Dr. Robert Speck, r.speck@fz-juelich.de

from JSC News No. 274, 30 July 2020