Emmi AI releases AB-UPT: Scaling Neural Surrogates to 100M+ Mesh Cells

Emmi AI has released Anchored-Branched Universal Physics Transformers (AB-UPT), a new generation of neural surrogate models that redefine the scalability and fidelity of AI based computational fluid dynamics (CFD) simulations. AB-UPT demonstrates state of the art predictive accuracy for automotive aerodynamics on volumetric meshes exceeding 100 million cells, trained and executed on a single GPU.

Scaling Neural CFD Surrogates

Recent advances in neural surrogate modeling have shown promise for transforming industrial simulations - but applying them to full-scale automotive aerodynamics has remained infeasible due to the enormous resolution of high-fidelity CFD meshes. Industrial grade datasets such as DrivAerML and AhmedML feature Hybrid RANS–LES simulations with up to 140 million volumetric cells, creating scalability and consistency challenges for existing neural operators.

AB-UPT addresses these barriers through two key innovations:‍

1. Anchored attention, which reduces the quadratic complexity of self-attention and allows efficient inference on hundreds of millions of mesh cells.‍

2. A multi-branch Transformer architecture, which separates geometry, surface, and volume representations while maintaining rich cross-attention between them.

This architecture allows AB-UPT to simulate complex flow phenomena with linear inference complexity, while preserving the expressivity and accuracy of full self-attention.

Performance at Industrial Scale

Across multiple high-fidelity datasets ShapeNet-Car, AhmedML, and DrivAerML -AB-UPT achieves state of the art accuracy on surface pressure, volume velocity, and vorticity.

It is the first neural surrogate capable of end to end inference on 9 million surface and 140 million volumetric cells using a single NVIDIA GPU, two orders of magnitude larger than any previous method.

AB-UPT further enables accurate estimation of aerodynamic drag and lift coefficients, reaching near-perfect correlation with ground-truth CFD values.
Unlike previous models, it supports direct inference from CAD geometry, removing the need for a meshed simulation input.

Physics Consistent Neural Simulation

Beyond accuracy and scalability, AB-UPT introduces a divergence-free vorticity formulation that enforces physical consistency as a hard architectural constraint.
This guarantees that predicted vorticity fields are free of non-physical sources or sinks - preserving the fundamental conservation laws that govern real world fluid dynamics.

By treating its architecture as a conditional neural field, AB-UPT can also apply local differential operators and enforce conservation properties during prediction, offering a path toward more interpretable and trustworthy neural simulations.

Enabling High-Fidelity Design Workflows

AB-UPT extends Emmi AI’s mission to empower engineering teams, simulation experts, and designers to perform high fidelity aerodynamic simulations orders of magnitude faster - directly on production-scale geometries.
By bridging numerical rigor with scalable machine learning, it represents a foundational step toward AI-native simulation tools that operate at full industrial fidelity.

Resources

• Paper: https://arxiv.org/abs/2502.09692v4
• OpenReview: https://openreview.net/forum?id=nwQ8nitlTZ
• Hugging Face: https://lnkd.in/dsPGc8CQ
• Code: https://lnkd.in/dppN9pQW
• Demo: https://demo.emmi.ai

‍