The main goal of MaX is to allow the pre-exascale and exascale computers expected in Europe in the 2020’s to meet the demands from a large and growing base of researchers committed to materials discovery and design. This goal will be achieved by:

i) an innovative software development model, based on the concept of separation of concerns, that will enable performance of the community codes on heterogeneous hardware architectures, without disrupting their internal structure, the richness of their simulation capabilities, and their distributed and open development model. In this way, the most important community codes for quantum mechanical materials modelling will be ready for pre-exascale machines by the completion of MaX programme, and prepared to be ported to new architectures as they will become available; ii) an integrated ecosystem enabling the convergence of high performance and high throughput computing, that will

allow steering the millions to hundreds of millions of simulations that are needed to optimise the properties and performances of a material or a device, with robust and reproducible workflows, all contributing to an ever growing repository of curated data; iii) a new approach to scientific computing in which hardware and software are co-designed and co-developed taking into mutual account the constraints and goals; iv) innovative measures for easy access to materials science applications, for engaging academic and industrial communities and fostering a broader and diverse pool of well trained users and developers.

All this is made possible by the coordinated effort of a team involving developers of the leading EU open source community codes in the materials domain, five leading European HPC centres, two technology partners, and training and communication experts.

MaX is fully aligned with the long term European HPC strategy and community, and to the work program call INFRAEDI-02-2018 – Subtopic (a), area 5 (Materials …).

UGENT will perform quality assessment of MAX codes in an automated fashion exploiting AIIDA plugins and workflows according to established community protocols, as well as to provide curated datasets of crystal structures, pseudopotentials and other relevant input that can enable turn-key workflows. A minimal requirement for any simulation is that it must be verified: further improvement of numerical approximations should not alter the result. Nowadays, computing resources allow us to reach that level of precision for electronic structure calculations. This has been exploited in a recent community endeavour involving several MAX members to assess the precision level of total energy predictions for a test set of 71 elemental crystals, using 40 different flavours of electronic structure codes.18 This was the most stringent test ever for these complex codes. Small differences allowed to iron out a few long-lasting bugs, and the final excellent agreement is a strong mutual reinforcement of the numerical correctness of the codes. This paves the way for a more ambitious attempt: verifying electronic structure codes in more diverse and realistic circumstances, and for more complex properties. Rather than examining 71 elemental crystals, we will consider a test set of about 1000 crystals, selected so that elements are represented in a systematic series of environments (structural variety) and of oxidation states (chemical variety). By comparing the results of an all-electron code (like FLEUR) with a plane-wave pseudopotential code (like QUANTUM ESPRESSO), the quality of the pseudopotentials can the thoroughly assessed. The majority of electronic structure predictions published worldwide relies on the performance of pseudopotential libraries, and our planned assessment are the most stringent ever to select the best-performing libraries and to examine in which conditions some pseudopotentials might fail. Beside expanding the size and systematicity of the test set, we will increase the complexity of examined properties, extending to electronic band structures and energy-derivatives (forces). This requires different crystal sets, containing lower-symmetry structures and vacancies or surfaces. In turn, knowing the precision of forces allows one to assess the precision of phonons and hence of all thermal properties that depend on vibrational free energies. Eventually, these verification efforts are a necessary condition to embark on validation studies: examining how far the numerically correct prediction at a given level of theory deviates from experiment. Recent examples of validation studies involve properties that depend on total energies only. We will bring this to the next level by performing a statistically relevant number of accuracy assessments for complex properties, with phonons as the first target. The verification work in this task will contribute to an increased confidence in the numerical correctness of electronic structure codes. The validation work will allow us to determine for the first time statistically meaningful error bars for DFT-predictions of phonons and phonon-dependent properties.