Physics AI research that's shaping the industry.

Mistral AI announced the acquisition of Emmi AI on May 27, 2026, emphasizing its commitment to advancing AI research and enterprise solutions for industrial engineering. Emmi's work aims to help engineers build next-generation products faster and achieve continuous performance improvements in operations. Mistral explicitly stated it will double down on foundational physical AI to serve industries that shape the physical world, including aerospace, automotive, semiconductor, and energy.

The Emmi team has released a series of physical AI research results since 2024. In February 2024, UPT (Universal Physics Transformer) was introduced as a framework for efficiently scaling neural operators, supporting mesh and particle simulations. In November 2024, NeuralDEM was released as the first end-to-end deep learning surrogate, enabling real-time simulation of large-scale multiphysics processes such as fluidized bed reactors.

In February 2025, AB-UPT (Anchored Branch Universal Physics Transformer) was released, designed for aerodynamic CFD, capable of handling 9M surface and 140M volume cells on a single GPU without remeshing. In October 2025, AB-UPT added datasets for automotive and aerospace applications. Also in October, GyroSwin was released, targeting plasma turbulence in nuclear fusion with a 5D surrogate model, aiming to address a key bottleneck in fusion energy development.

In November 2025, the team published the 'Fluid Intelligence' paper, envisioning the application of AI foundation models in computational fluid dynamics (CFD) and deconstructing the core components of industrial-grade CFD simulations. In December, the team released a new 3D wing transonic CFD dataset containing approximately 30,000 samples, filling the gap left by existing 2D datasets in 3D phenomena.

The core advantage of AB-UPT lies in processing raw geometry without remeshing, greatly simplifying the simulation workflow. NeuralDEM reduces industrial process simulation from hours to real-time, which is transformative for scenarios like reactor design. GyroSwin targets plasma turbulence in nuclear fusion, a key obstacle in next-generation reactor design. These technologies collectively aim to accelerate physical simulation with AI, reducing engineering trial-and-error costs.