Multiscale geometry transformers enable accurate drag/lift prediction in CAE simulations

The challenge of accurately simulating complex physical phenomena, such as air flow around a vehicle, requires increasingly sophisticated computational techniques. Corey Adams, Rishikesh Ranade, and Ram Cherukuri, along with Sanjay Choudhry from NVIDIA Corporation, introduce GeoTransolver, a new approach to operator learning that significantly improves the accuracy and efficiency of these simulations. This innovative system employs a multi-scale geometry-aware physics attention transformer, effectively connecting the laws of physics with the complex shapes of real-world objects. By permanently anchoring computations to the domain structure and motion regime, GeoTransolver achieves superior performance on difficult aerodynamic datasets, demonstrating improved robustness and data efficiency compared to existing methods, and paves the way for high-fidelity surrogate modeling across complex and irregular domains.

Geometric transformer learns partial differential equations

In this work, we introduce Geometry-Informed Neural Operator Transformers (GNOTs), a novel architecture that combines neural operators (NOs) and transformers for learning and predicting solutions to partial differential equations (PDEs). The key innovation lies in incorporating geometric information into the transformer architecture by leveraging graph neural networks to encode the domain geometry and employing geometry-aware attention mechanisms. This approach improves prediction accuracy and captures the underlying physical phenomena more effectively. This work builds on existing research on neural operators, such as Fourier neural operators and graph neural operators, and extends the application of transformers to scientific computing.

We also leverage Physics-Informed Machine Learning techniques such as Physics-Informed Neural Networks. Benchmarking on standardized datasets, such as those used in automotive aerodynamics, demonstrates improved performance compared to existing methods. The field of operator learning is rapidly evolving, with increasing emphasis on combining different machine learning techniques and incorporating geometric information into models. Benchmarking on standardized datasets is essential for evaluating algorithm performance, and the combination of neural operators and transformers holds great promise in solving complex partial differential equations and advancing scientific computing.

Geometry-aware transformers for physics simulation

Scientists have developed GeoTransolver, a new transformation architecture for computational analysis that addresses the limitations of handling complex geometries and different physical conditions. The system replaces standard attention mechanisms with Geometry-Aware Latent Embeddings (GALE), combining physics-aware self-attention with mutual attention to a shared geometric context. This context obtained from the multiscale analysis permanently anchors the calculations to the domain structure and operating conditions throughout the simulation. The team designed a way to precompute multiscale features and add them to the local input with a shared context before the first transformer block. This ensures consistent geometric and physical awareness throughout the calculation. Tested on benchmarks such as DrivAerML and Luminary SHIFT models, GeoTransolver shows improved accuracy and robustness compared to existing methods.

Geometry-aware transformers for engineering simulation

Scientists have developed GeoTransolver, a new transformation architecture for computer-aided engineering, making significant advances in surrogate modeling for complex physical simulations. In this study, we introduce Geometry-Aware Latent Embedding (GALE) attention, which combines physics-aware self-attention with mutual attention to multiscale geometric neighborhoods and global context. This innovative approach directly addresses challenges in physical modeling such as irregular geometries and limited data availability. Experiments demonstrate that GeoTransolver achieves improvements in accuracy, data efficiency, and robustness across datasets including DrivAerML, Luminary SHIFT-SUV, and Luminary SHIFT-Wing. The team benchmarked GeoTranssolver against state-of-the-art architectures and consistently demonstrated superior performance. A key element of the design is the geometry context projection strategy, which maps geometry and global features into physical state space and injects them into all transformation blocks. This reduces representational drift and improves stability.

Geometry-aware transformers improve simulation accuracy

GeoTransolver, a new transformation architecture for computational engineering, successfully integrates geometry into the learning process, improving simulation accuracy and robustness. The team achieved this by replacing standard attention mechanisms with Geometry-Aware Latent Embeddings (GALE), which combines physically aware self-attention with mutual attention to a shared geometric context. This context obtained from the multiscale analysis permanently anchors the calculations to the domain structure and operating conditions throughout the simulation. Extensive benchmarks on datasets such as DrivAerML, SHIFT-SUV, and SHIFT-Wing demonstrate that GeoTransolver consistently matches or exceeds the performance of existing state-of-the-art methods. The model is more resilient to changes in geometry and operating conditions while maintaining efficient data usage. The GeoTranssolver architecture and associated software are openly available, facilitating broad access and collaboration within the research community.