Atomistic modelling plays an important role in representing the properties and behaviour of materials at the atomic scale. This innovation study aims to optimise atomistic materials modelling by adding physics-based functionality to Machine Learning Potentials (MLPs). Traditional interatomic potentials (IPs) have limitations in accurately capturing the complex dynamics of chemical bond formation and breakage, limiting their applicability in various materials science and biomolecular design tasks. MLPs provide a solution to overcome these challenges. Machine Learning Potentials (MLPs) offer the possibility of extending advanced algorithms to learn complex energy landscapes from quantum mechanical simulations, thereby enabling more accurate and versatile atomistic modelling. However, realising the full potential of MLPs requires adaptation to emerging hardware architectures, in particular the parallelism offered by Graphics Processing Units (GPUs). Through an interdisciplinary collaboration between atomistic materials modelling experts from Aalto University and high performance computing (HPC) specialists from CSC, this study aims to maximise the use of supercomputing resources such as LUMI within the EuroHPC Joint Undertaking.
Therefore, this study has three objectives aimed at advancing atomistic modelling capabilities:
- The development of novel algorithms: XCALE aims to develop 3-4 novel algorithms that incorporate physics-based functionality into MLPs. These algorithms will optimise short-range “bonded” interactions, using functions to calculate energy and forces at the atomic level. In addition, long-range “non-bonded” interactions, such as electrostatics and dispersion forces, are modelled using physics-based equations, improving the accuracy and portability of MLPs.
- Optimisation for GPU architectures: Capitalising on the transformative potential of GPU-based supercomputing architectures, XCALE aims to optimise MLP algorithms for efficient parallelization on GPUs. This will involve the adaptation of existing CPU-based codes and the development of new algorithms designed to exploit the parallel computing capabilities of GPUs, resulting in computational efficiency and scalability.
Integration with HPC resources: Collaboration with CSC’s HPC experts and access to advanced supercomputing facilities will enable XCALE to maximise the use of available computing resources. By harnessing the computing power of pre-exascale supercomputers, XCALE aims to accelerate the pace of atomistic simulations and enable high-fidelity modelling of complex materials and biomolecular systems.
This effort faces several challenges, including the complexity of physics-based modelling, the need for algorithmic optimisation for GPU architectures, and the integration of different expertise from the atomistic modelling and HPC domains. Through interdisciplinary collaboration and optimisation for modern computing architectures, XCALE aims to bridge the gap between available computing resources and the readiness of the HPC community to use them.