Mistral AI Physics AI Research and Emmi AI Acquisition

Mistral AI has acquired Emmi AI to develop foundational Physics AI for industrial engineering, targeting sectors such as aerospace, automotive, semiconductors, and energy. This initiative aims to accelerate product development and enable continuous performance gains in large-scale operations.

Integration of Emmi AI into Mistral

Mistral AI is integrating Emmi AI's research to build state-of-the-art enterprise solutions for industrial engineering. The goal is to provide engineers with tools to build next-generation products more rapidly and secure operational performance gains at scale.

Computational Fluid Dynamics (CFD) and Aerospace Breakthroughs

Mistral AI is addressing critical gaps in aerospace simulation by moving beyond 2D airfoils to 3D phenomena.

Transonic Regimes and Neural Surrogates

In a December 2025 publication, Mistral presented a new dataset of CFD simulations for 3D wings in the transonic regime. This dataset includes volumetric and surface-level fields for approximately 30,000 samples with unique geometry and inflow conditions, specifically targeting highly-turbulent transonic regimes where existing datasets were previously lacking.

CFD Foundation Models

In November 2025, Mistral published research bridging the machine learning and CFD communities by deconstructing industrial-scale CFD simulations into their core components to explore the future of AI foundation models in fluid intelligence.

Physics Transformers and Neural Surrogates

Mistral is utilizing Transformer-based architectures to handle complex physics simulations without the need for traditional remeshing.

Anchored-Branched Universal Physics Transformer (AB-UPT)

Introduced in February 2025, the AB-UPT is designed for aerodynamics CFD. It can handle raw geometry without remeshing, processing 9 million surface cells and 140 million volume cells on a single GPU. Subsequent reports in October 2025 expanded the empirical evaluation of AB-UPT with new datasets for automotive and aerospace applications.

Universal Physics Transformer (UPT)

Published in February 2024, the UPT framework enables the efficient scaling of neural operators across diverse spatio-temporal problems, supporting both grid and particle simulations.

Multi-Physics and Energy Simulations

Mistral's research extends to high-energy physics and industrial process simulation.

GyroSwin for Nuclear Fusion

Published in October 2025, GyroSwin provides 5D surrogates for gyrokinetic plasma turbulence simulations. This research addresses plasma turbulence, a primary obstacle to viable nuclear fusion power and plasma confinement in next-generation reactor design.

NeuralDEM for Industrial Processes

Introduced in November 2024, NeuralDEM is the first end-to-end deep learning surrogate for large-scale multi-physics processes. It enables the real-time simulation of industrial processes, such as fluidized bed reactors.