AI-Aided Engineering · Group news
Introducing NVIDIA Research’s AI-Aided Engineering Group
Learning methods for physical systems
AI-Aided Engineering (AIE) is now an official research group within the Learning and Perception Research (LPR) organization at NVIDIA Research, led by Jean Kossaifi.
For decades, progress in engineering—from aerospace to materials science—has relied on Computer-Aided Engineering (CAE). Numerical solvers remain essential, but high-fidelity simulations can require days or weeks, limiting how many designs, operating conditions, and hypotheses engineers can evaluate. GPU acceleration has substantially reduced these costs. AIE studies how learning-based methods can extend this progress and further accelerate scientific discovery.
AI-Aided Engineering is our research program for using AI to accelerate analysis, design, and discovery in physical systems, building on NVIDIA's work in AI, accelerated computing, and CAE.
The core principle of AIE is to develop data-driven models that learn solution operators and predictive distributions for physical systems from observational and simulation data. These models can serve as computationally efficient surrogates for expensive simulations and support uncertainty quantification, optimization, and inverse design.
Our research focuses on the algorithms and model architectures required to make these models accurate, reliable, and useful. A central challenge is to move beyond treating a particular discretized mesh as the model's only representation. We develop neural operators and related methods that can represent engineering geometry—and native CAD where appropriate—while conditioning on boundary and initial conditions, forcing, material properties, and other physical parameters. We study accuracy, stability, uncertainty, and generalization across discretizations, resolutions, geometries, operating conditions, and physical regimes.
We are initially pursuing this program in three application areas. In CFD and engineering design, we study surface and volumetric flow prediction on complex automotive and aerospace geometries, generalization across designs and operating conditions, and learned models for optimization and inverse design. In weather, we study multiscale atmospheric dynamics and probabilistic medium-range forecasting. In materials, we study mechanical and thermal response under variations in geometry, material properties, loads, and boundary conditions.
Modern AI research relies heavily on open-source software and the open exchange of results. Whenever release, data, and intellectual-property constraints allow, we contribute papers, implementations, and tools that others can inspect, use, and improve. Real engineering problems motivate our research, and validated methods and implementations flow back into the broader scientific and engineering ecosystem.
Atlas, developed through a broad collaboration across NVIDIA, is a public example of this research direction in multiscale probabilistic medium-range weather forecasting. NeuralOperator is our open-source PyTorch library for developing, training, and evaluating neural operator methods, with documentation and examples for learning solution operators of partial differential equations and other physical systems.
At NVIDIA, AIE focuses on algorithmic research while collaborating with teams across accelerated computing, scientific software, AI frameworks, data, and visualization. These collaborations allow us to evaluate new methods at scale and connect successful research with tools used by scientists and engineers.
Our longer-term objective is to shorten the design-to-analysis cycle. Connecting learned physical models with interactive engineering environments and digital twins could provide low-latency feedback, allowing engineers to explore larger design spaces, test more hypotheses, quantify uncertainty, and optimize designs more efficiently.
The AIE team includes Jean Kossaifi, Nikola Kovachki, and Daniel Leibovici, together with research interns Dawon Ahn, Boyuan Yao, and Hojjat Kaveh. Our research involves collaborators across NVIDIA, and individual publications and project pages will reflect the complete contributor lists. Please get in touch if you would be interested in collaborating with us!