AB-UPT Sets New Benchmarks Across Major CFD Datasets
Emmi AI now holds the top performance across all major publicly available computational fluid dynamics (CFD) datasets, including DrivAerML, DrivAerNet++, AhmedML, and the newly released SHIFT-SUV and SHIFT-Wing from Luminary Cloud.
A new Technical Report, AB-UPT for Automotive and Aerospace Applications, extends the results from the AB-UPT paper published at Transactions on Machine Learning Research (TMLR), demonstrating that Anchored-Branched Universal Physics Transformers (AB-UPT) achieve state of the art accuracy and efficiency on complex real-world automotive and aerospace simulations.
Results at a Glance
SHIFT-SUV (automotive):
AB-UPT attains mean relative errors below 1 % for drag force prediction and maintains near-perfect correlations for lift, with full 3D fields accurately reproduced.
Training on a single NVIDIA H100 GPU converges within 13.5 hours, with inference completed in under 34 seconds for full-vehicle meshes of 45 million cells.
SHIFT-Wing (aerospace):
Across Mach 0.5 and 0.85 regimes, AB-UPT achieves R² = 1.00 for both drag and lift forces — matching ground-truth CFD forces within 2 % error.
Models can infer aerodynamic forces directly from isotropic CAD geometry, without needing access to the simulation mesh, at sub-second inference times.
These findings reinforce AB-UPT as the current state-of-the-art neural surrogate for external aerodynamics, capable of replacing traditional CFD solvers in design and optimization workflows.
Toward Trustworthy AI in Engineering
From DrivAerML to SHIFT-Wing, AB-UPT’s results establish a clear record of reproducible performance across independent datasets and simulation platforms.
As Emmi AI notes in the report, “Trust is built on proof, not promises.”
Each result is backed by transparent evaluation protocols, reproducible baselines, and open access to forward-pass outputs for community verification.
Resources
- Technical Report: https://arxiv.org/abs/2510.15808
- Forward Pass (DrivAerML): https://github.com/Emmi-AI/ab-upt
- AB-UPT Paper: https://arxiv.org/abs/2502.09692
Ready to accelerate your development?
Explore how AI-powered models are transforming physics simulations and learn how to build or integrate yours.
Latest updates
.webp)
.webp)
.webp)
Emmi AI raises EUR 15M to bring AI to the heart of industrial engineering
Funding from 3VC, Speedinvest, Serena, and PUSH marks the largest seed round ever raised by an Austrian startup and accelerates Emmi AI’s vision for AI-enabled engineering.
Emmi AI releases AB-UPT: Scaling Neural Surrogates to 100M+ Mesh Cells
Emmi AI presents AB-UPT, a novel architecture that scales neural surrogates for computational fluid dynamics (CFD) to industrial-scale problems with over 100 million mesh cells - achieving state of the art accuracy, mesh free inference, and physics consistent predictions.
NeuralDEM, Now Open Source: Real-Time Deep Learning for Industrial Particulate Flows
NeuralDEM introduced in November 2024 as the first end to end deep learning alternative to CFD–DEM multiphysics simulations, is now open source. The dataset and model enable real-time, physically consistent simulation of industrial particulate flows at production scale.
.svg)
