These past two days I have participated in the Tri-Lab Support Team (TST) meeting of the CESMIX, the newly founded Center for the Exascale Simulation of Material Interfaces in Extreme Environments at the Massachusetts Institute of Technology. Within the next few years the idea of the CESMIX is to develop a multi-level simulation stack all the way up from DFT over MD to flow simulations to be able to discover novel materials suitable for extremely high temperatures under atmospheric conditions. The prototypical application for such materials would be heat shields for example in space crafts returning to earth or supersonic planes.

One novel aspect of the project is to include progress from modern compiler techniques and programming language design when building the software stack. In particular the challenge is that multiple codes will be involved in the project that feature a large variety of programming languages (FORTRAN, C++, Julia, python, ...). On top of that one goal is to keep track of simulation errors using uncertainty quantification (UQ) and use that insight to construct a multi-fidelity workflow. In such an approach the data generation for the simulation does not only employ a single accurate model, but in fact features multiple simulation layers based on cheaper and cruder models as well as more costly and accurate ones. Using the deduced knowledge of the error one can dynamically switch between these models and reach a compromise between accuracy and computational cost, but in a way that the result still has the quality to be comparable to experiments. As the employed fidelity layers the CESMIX project targets classical molecular dynamics or ab initio molecular dynamics with various kinds of density-functional theory (DFT) methods ... and this is the angle of my involvement in the project.

In particular using our density-functional toolkit (DFTK) the idea is to be able to quickly prototype parts of the workflow. Then, building on DFTK's design as a multi-disciplinary platform (see my related blog articles), we want to start incorporating new techniques (GPU platforms, UQ, multi-fidelity) to see how they could fit. I am excited about the opportunity to contribute to a project, which shares a lot in philosophy to DFTK itself. In particular I am looking forward to seeing how DFTK will play out in a real-world research scenario for connecting the needs from the modelling side with the approaches of the computer science folks.

With respect to the TST meeting, I briefly gave an overview about DFTK to show where we are. Slides and a short demo are attached below.

Link
DFTK.jl: A multidisciplinary Julia code for density-functional theory development (Slides)
Benchmarks and DFTK demo (Tarball)