Speaker
Description
Over the past decade, Machine Learning and Artificial Intelligence technologies have evolved at an extraordinary pace, making collaboration among geographically distributed experts and students more critical than ever.
The AI_INFN Platform is designed to play a key role in providing access to hardware accelerators for research communities in both fundamental and applied physics.
The platform integrates cloud-based interactive environments, provisioned via JupyterHub on Kubernetes, with remote access to HTC and HPC resources for batch processing. These resources are connected through Virtual Kubelet technology, using InterLink as a provider.
A distinctive feature of AI_INFN is its portability and scalability: for example, during training events and hackathons, it is regularly deployed across heterogeneous sites and hardware configurations.
In this contribution, we present the current status of the platform and illustrate two representative workflows orchestrated through AI_INFN JupyterLab instances, combining Cloud, HTC, and HPC resources. The first workflow focuses on Deep Neural Network training on CINECA Leonardo, integrated into GitHub Actions pipelines and leveraging Azure resources to package and distribute trained models as artifacts. The second workflow concerns GPU-powered particle detector simulations, integrating Snakemake workflows with INFN Cloud S3 storage to coordinate HTC and HPC sites.
Finally, we outline future plans toward a more modular architecture, aimed at facilitating deployment in managed Kubernetes clusters alongside other applications.