Learning Materials at eXascale

Challange

Solving large and sparse numerical linear systems in materials science on massively parallel supercomputers is a complex endeavour that requires a delicate balance between accuracy and computational efficiency. Challenges include managing the scale and complexity of these systems, optimising scalability on parallel architectures, and addressing real-world material complexities.

Project goal

The LimitX project represents a ground-breaking step in the field of computational Materials Science and aims to develop an innovative recommender system. This system aims to revolutionise the solution of large-scale sparse linear systems by accelerating and scaling the solutions of linear systems so that materials science research can be routinely performed on exascale clusters. At its core, this system relies on a two-pronged approach: first, the development of a spectral predictor system and, second, the use of an extensive database of matrices that encapsulate the essence of surrogate space in the field of materials science. The spectral predictor system is the heart of the recommendation system. It utilises deep learning techniques to predict spectral properties that are crucial for the efficient solution of linear systems. The extensive matrix dataset captures the diversity of spectral patterns that occur in material science calculations.

The application of this recommender system promises to be transformative, as it enables simulations with hundreds of thousands of atoms, a feat previously unrealisable on current pre-exascale clusters. The linear scaling of DFT (density functional theory) codes such as BigDFT will enable researchers to simulate and analyse complex material systems with unprecedented accuracy and computational efficiency, opening up new horizons for scientific exploration in this field. 

Researchers working on the project: Davor Davidović