Minisymposium

SmartSim has always been developed with the goal of overcoming the divide between standard HPC numerical libraries and AI toolkits. SmartSim’s philosophy is based on loose coupling of different applications and processing units involved in a workflow and it is one of the main reasons for its growing success and fast adoption not only in climate and earth science, but also in CFD and other research fields. SmartSim is currently used in several computationally demanding projects and has been run at different scales, from laptops to Frontier, the world’s first exascale system. With the advent of new AI models and innovative ways of including data-driven techniques into simulations, SmartSim is faced with new challenges, which will drive its future development. In this talk, we will show examples of how SmartSim was used in successful projects, and talk about the project’s future steps.

The rise of machine learning (ML) has seen many scientists seeking to incorporate these techniques into numerical models. Doing so presents a number of challenges, however. The Institute of Computing for Climate Science (ICCS) has explored this problem in the context of coupling ML components/parameterisations into climate models. In this talk we will explore a number of challenges in this area, how ICCS has tackled them, and what has been learnt in the process. We will present FTorch, a library developed by ICCS to bridge the gap between Fortran (in which many large physics models are written) and PyTorch (in which much ML is performed) and lower the technical barrier to scientists seeking to leverage ML in their work. Case studies have been performed using both CPU and GPU architectures from laptops up to HPC systems. We will reflect on the design challenges of coupling ML parameterisations to large numerical models and outline a framework guidelines following software design principles to aid in this process. Finally we will discuss ongoing work using the Community Earth System Model (CESM) to re-deploy a neural net trained using a high-resolution model with a different grid and variables to a new setting.

First-principles materials modelling software can accurately predict many materials properties, but requires the numerical solution of complex, non-linear partial differential equations. Solving these equations is computationally intensive, and first-principles simulations consume a significant fraction of HPC usage (e.g. ~40% of the UK ARCHER2 Tier-1 facility). In recent years, machine learning (ML) methods have been applied to some of these property simulations, to reduce the number of numerical evaluations. The vast materials parameter space means that devising a "universal" ML model is challenging. One promising alternative is to couple the ML and direct numerical simulations more tightly, training and using the ML "on-the-fly". We will discuss these challenges and opportunities, along with results from embedding Gaussian Process-based ML models in the popular CASTEP first-principles modelling software, to reproduce and predict atomic forces substantially faster and with controllable uncertainty.

With the recent rise of machine (ML) and deep learning there have been several efforts to incorporate these techniques into numerical models. Doing so presents a number of challenges however, including but not limited to: framework and language interoperation; ensuring physical compatibility, stability, and constraints; portability and generalisation of models outside their training domain; understanding biases and uncertainties; and the efficient use of differentcomputer architectures. The three invited talks in this minisymposium present progress that has been made across a range of scientific domains whilst also discussing challenges faced andtechniques to tackle them. This discussion session is a chance to reflect on the common ground between these talks, and what can be learnt from one-another. We welcome anyone who is using ML components in their work with interesting stories to share, and anyone interested in incorporating ML into their work who wishes to learn more. This session, and the minisymposium as a whole, is an opportunity to meet others in the domain.