Relevance to WCCM–ECCOMAS

Provide computational scientists, engineers, and PhD students with a practical, end-to-end introduction to physics-ML surrogate modelling using NVIDIA PhysicsNeMo, covering data curation, model training, and inference evaluation on real industrial aerodynamics data.

Course description

This short course provides a practical, code-first introduction to physics-informed and data-driven surrogate modelling using NVIDIA PhysicsNeMo, an open-source framework for building physics-constrained machine learning models. The session is centred on two state-of-the-art surrogate modelling architectures:

Transolver – a transformer-based neural PDE solver that exploits physics-aware attention to achieve mesh-independent, generalizable solutions across diverse physical systems. GeoTransolver – an extension of Transolver that incorporates geometric representations of computational domains, enabling accurate surrogate modelling on complex, unstructured, and irregular geometries common in industrial computational mechanics.

By the end of the session, participants will be able to: Set up the PhysicsNeMo and PhysicsNeMo-Curator environments for an end-to-end external aerodynamics ML workflow. Build a data curation pipeline to download, process, and validate the DrivAerML dataset from raw CFD outputs (STL, VTP, VTU) to Zarr-format tensors ready for model training. Configure and train Transolver on external aerodynamics data, including normalization statistics computation and Hydra-based hyperparameter management. Extend the workflow to GeoTransolver, understanding how GALE attention encodes multiscale geometry and boundary conditions to improve predictions across varying vehicle shapes. Run inference, evaluate drag and lift coefficients against held-out geometries, and interpret L2 and R2 metrics in the context of industrial surrogate modelling requirements.

Objectives and target groups

Target Groups

• Computational scientists and engineers who wish to augment or accelerate their simulation workflows with machine-learning surrogates.
• Applied mathematicians and numerical analysts interested in transformer-based neural operators and their theoretical foundations.
• PhD students and early-career researchers seeking hands-on experience with modern physics-ML frameworks used in both academia and industry.
• Practitioners with a background in CFD, structural mechanics, or multi-physics simulation who are new to physics-informed ML.

Prerequisites. Basic familiarity with Python and PyTorch (or any deep-learning framework) is expected. No prior experience with PhysicsNeMo or neural operators is required. Participants are encouraged to bring a laptop with internet access; cloud-based GPU notebooks will be provided.

Scientific and technical areas covered

• Neural operators
• Transformer-based surrogate models (Transolver, GeoTransolver)
• Physics-informed machine learning
• CFD data curation
• External aerodynamics
• Geometry-aware deep learning
• Model order reduction and generative modeling