Minisymposium

Since its release in 2021, SmartSim has been gaining momentum in several scientific domains, including climate modeling and CFD. SmartSim’s unique set of features allow researchers to perform in-situ data analysis, extend capabilities of well-established numerical software with cutting-edge AI techniques, and orchestrat the launch of applications on HPC systems. This integration not only streamlines modern scientific workflows but also addresses the longstanding challenge of language and paradigm disparity in scientific computing.In our talk, we will delve into the core functionalities of SmartSim, highlighting its latest enhancements, and discuss how they are used in state-of-the-art projects developed at the intersection of HPC and AI.

Relexi is a powerful tool that allows to use existing simulation codes as training environments for reinforcement learning (RL) on high-performance computing (HPC) systems. This framework allows to apply RL to problems typically requiring HPC hardware such as computational fluid dynamics (CFD) or related fields. For this, Relexi applies the SmartSim library, which allows to manage the individual simulation environments on HPC systems and provides an efficient communication channel between itself and the simulation code. In this talk, we demonstrate two specific applications for the use of RL in CFD. First, we apply the framework to a task in active flow control. Here, the RL agent is trained to minimize the drag for the flow around a two-dimensional cylinder using blowing and suction jets at the cylinder’s poles. For this case, the agent is demonstrated to reduce the experienced drag by about 15%. Moreover, Relexi is applied to the task of turbulence modeling in large eddy simulation, where it was found to outperform traditional models, while being robust against changes in resolution, Reynolds number, and also when applied to heavily deformed meshes.

The ORCHIDEE land surface model is one of the IPSL's Earth System Model components. The radiative transfer portion of the land model calculates reflected, absorbed and transmitted light at multiple canopy levels. This calculation is crucial to the climate system, but is also the most time-consuming in ORCHIDEE. To solve this problem, we show the results of our random forest-based emulator to represent the calculation in a fast and accurate way. This emulator closely mimics the original numerics-based model with relative errors of < 10% and correlations > 0.9, while taking ~50% less computational time. The second challenge describes the process of integrating this emulator in an online-way within the context of ORCHIDEE. We describe the process of integrating using the SmartSim open-source, machine-learning tool into the Fortran-based model and will show initial results and performance benchmarks that demonstrate the future viability of hybrid HPC/AI climate modelling.

The effectiveness of AI for a variety of scientific tasks has rapidly improved over the last few years, changing the way that we can perform scientific workflows on HPC. Our recent work deploying a workflow around a Large-Language Model (LLM) for generating protein sequences is a good example of many of the opportunities and challenges. Embedding the LLM within a larger protein screening workflow enabled us to target simulations more effectively and find better sequences faster; but was not easily accomplished with conventional workflow tools. Interleaving AI predictions required expressing dynamic actions with the workflow application, the size of data being transferred around the workflow required adding a secondary data transfer fabric, and the evolving nature of the tasks being deployed on HPC required particular attention to caching elements of workflow tasks. We will discuss how we addressed these and other challenges in this presentation.