Collaboration between interTwin and AI4EOSC for AI Inference Offloading from the Cloud to HPC

Integrating cloud-native AI inference workflows with large-scale HPC infrastructures poses a number of nontrivial challenges, both technical and architectural. 

One fundamental challenge was enabling transparent and secure offloading of containerised workloads from elastic, cloud-based Kubernetes environments to HPC systems like VEGA, which operate under different execution and scheduling paradigms.

Equally critical was the need for automated, end-to-end service execution that abstracts infrastructure complexity from users, preserving usability without compromising performance or flexibility.

The collaboration between interTwin and AI4EOSC confronted these head-on to demonstrate how Digital Twin applications can execute AI-driven scientific workflows across the Computing Continuum.

To overcome these obstacles, the collaboration leveraged two core modules within the interTwin software stack:

OSCAR, the shared AI inference platform of both interTwin and AI4EOSC, served as the cloud-native interface for managing and executing AI models. It enabled modular deployment of scientific inference services in a flexible, scalable manner.

interLink handled the seamless offloading of Kubernetes-based workloads to VEGA’s HPC infrastructure. It acted as a transparent abstraction layer, allowing workloads to migrate across computing domains while remaining fully compatible.

The integration between OSCAR and interLink formed the core of the automated execution pipeline, allowing AI models to be deployed, triggered, and executed on HPC resources without any manual intervention, an essential step toward user-driven scientific workflows in Digital Twin scenarios.

We demonstrated the feasibility of this solution by executing one of the models of the AI4EOSC Automated Thermography use case: “Thermal Bridges on Building Rooftops Detection”.