Welcome and Introduction

Coupling HPC Plasma Simulations and AI at Exascale

• Richard Pausch (Helmholtz-Zentrum Dresden-Rossendorf (HZDR))

We present recent results on coupling plasma simulations and large-scale AI models on exascale HPC systems and argue that the workflows we are implementing are the future of I/O.

Q&A

• Richard Pausch (Helmholtz-Zentrum Dresden-Rossendorf (HZDR))

PhysicsNeMo

• Abouzar Ghasemi (NVIDIA)

NVIDIA PhysicsNeMo is an open-source Physics-Informed Machine Learning (Physics-ML) platform. It enables the creation of high-precision, physically based digital twins. - [Predict the time evolution of electric field magnitude inside shielded enclosures][1] - [AI-Driven Prediction of Epoxy Underfill Flow: Capturing Interface Dynamics in a Two-Phase System][2] [1]: https://github.com/ghasemiAb/physicsnemo/tree/EM_wave/examples/cfd/enclosure_shielding [2]: https://github.com/ghasemiAb/physicsnemo/tree/underfill-ag/examples/cfd/underfill_dispensing

Q&A

• Abouzar Ghasemi (NVIDIA)

Machine Learning Across Space and Time in Microscopy

• Martin Weigert (TU Dresden)

Modern microscopes generate large, information-dense images across space and time, challenging current vision models. I will present our work on machine learning for quantitative microscopy, including segmentation of complex biological images, self-supervised learning from videos, and multiscale vision transformers that integrate fine detail with broader biological context.

Q&A

• Martin Weigert (TU Dresden)

NVIDIA Cosmos und NVIDIA Omniverse

• Pallavi Mohan (NVIDIA)

NVIDIA Cosmos and NVIDIA Omniverse work hand-in-hand to form the cohesive ecosystem for Physical AI, Omniverse (The Infrastructure and Simulation Environment): Omniverse serves as the base and virtual "data space". Here, physically correct digital twins are created, physical laws are simulated, and environments (e.g., factories, road networks) are mapped using the OpenUSD standard. Cosmos (The World Foundation Models): Cosmos is the AI model platform that interacts directly with the 3D worlds from Omniverse. Cosmos consists of powerful models such as Cosmos Predict (for simulating world states), Cosmos Transfer (for photorealism), and Cosmos Reason (for physical understanding and decision-making).

Q&A

• Pallavi Mohan (NVIDIA)

Can We Trust AI-Generated Knowledge? From Machine-Scale Outputs to Grounded Scientific Workflows

• Michael Färber (TU Dresden)

Scientific knowledge is increasingly produced at machine scale — by large language models, but also as predicted structures, simulations, and data streams. Yet fluent answers are not necessarily correct: LLMs hallucinate and rarely reveal their sources. I will present our work on making AI-generated knowledge trustworthy by grounding LLMs in structured, verifiable knowledge — scholarly knowledge graphs (e.g., SemOpenAlex), multi-agent retrieval-augmented generation with traceable sources (our deployed SQuAI system), and neurosymbolic methods that pair LLM reasoning with knowledge-graph structure.

Q&A

Shaping Your GenAI: An Overview of Compression and Inference Serving

• Nael Fasfous (NVIDIA)

LLM finetuning adapts a pre-trained foundation model, such as Llama, Mistral, or NVIDIA NeMo models, to domain-, company-, or task-specific needs. After fine-tuning, optimization and compression techniques can improve efficiency, reduce serving costs, and prepare the model for deployment. Choosing the right inference serving approach is then critical to delivering a scalable, production-ready GenAI solution. NVIDIA provides an end-to-end software and hardware stack for this workflow, including NVIDIA NeMo for LLM development and highly optimized GPU systems for training and inference.

Q&A

• Nael Fasfous (NVIDIA)

Conclusion & Outlook - Wrap up