Skip to content

PhysicsNeMo

AeroJEPA: From Research to Enterprise Use

AI-enhanced simulation solutions now support high-fidelity computational fluid dynamics (CFD) simulations for aerodynamic design. Such solutions leverage AI surrogates trained on a corpus of simulated data, and several challenges remain before these solutions become more generalizable. Examples include the transferability of pretrained models, from cross-regime generalization to geometry and boundary condition generation, and the question of what constitutes a universal CFD input/output representation for training on diverse corpora of datasets.

One such challenge is an effective latent representation of the geometry and flow physics that can encode the physical variables. A “good” latent representation helps in scaling by compressing the data. It also preserves the structures that transfer across prediction tasks. For foundation models, this means capturing reusable concepts while filtering out incidental details such as resolution, mesh, or data format. A strong latent space improves sample efficiency, supports generalization to new domains, enables multiple input and output modalities, and makes downstream adaptation cheaper. If the latent space discards important information or encodes dataset-specific shortcuts, scaling the model or training data only reinforces those limitations.

AeroJEPA, introduced by Giral et al., takes on the challenge of creating a compact latent representation through predictive latent learning. The model is the result of a broad collaboration across universities in the US and Spain, led by Professor Ricardo Vinuesa’s group at the University of Michigan. Instead of predicting only a large flow field, it learns a compact representation of the flow that can be decoded when needed and used directly for downstream analysis. This post introduces the central idea behind AeroJEPA and highlights what the experiments show. It also describes how the researchers leveraged PhysicsNeMo and its open-source development to turn the research implementation into a scalable and optimized production workflow that other researchers and enterprises can train, evaluate, and customize.

A Probabilistic Surrogate for Kinetic Monte Carlo Simulations

Kinetic Monte Carlo (KMC) is a foundational method for simulating systems that evolve through a sequence of discrete, random events. It appears across a wide range of science and engineering simulations: defect and trap generation in semiconductor devices (device reliability and endurance simulations in TCAD), ion transport and degradation in battery materials, catalytic reactions on surfaces, crystal and thin-film growth, and microstructure evolution in metals and alloys. KMC is a highly effective method, but it is computationally expensive, and it has a few characteristics that set it apart from the conventional time-stepping and integration-based solvers on which most deep-learning surrogates are built. The time between events is a random quantity produced by the solver, and the solver is itself stochastic, so the same starting configuration yields a different trajectory on every run.

The new kinetic Monte Carlo example in PhysicsNeMo is an end-to-end recipe, from training to trajectory generation, for a probabilistic surrogate that emulates a KMC solver one event at a time. From the current state, it predicts a full probability distribution over the next event, and running it forward repeatedly turns a single starting configuration into an ensemble of plausible trajectories with calibrated uncertainty.

Introducing the PhysicsNeMo Diffusion Module: Composable, Extensible Generative Modeling for Physics-AI

In physics-AI, many problems have not a single answer but a distribution of possible ones, which is what diffusion models sample. The PhysicsNeMo diffusion module brings diffusion to scientific data, built on PyTorch and included in the open-source PhysicsNeMo library.

Its composable components range from ready-to-use defaults to fully custom research implementations, serving a CAE engineer, a weather modeler, and a diffusion researcher alike. They cover the whole workflow, from training a model to sampling from it. A single trained model then serves many tasks at inference, drawing large ensembles, solving inverse problems by sampling the posterior, and enforcing physical constraints, without retraining. The module is designed for high throughput on large ensembles and for scaling to the large domains typical of scientific applications. In this post, we'll show what the module unlocks, from large ensembles to data assimilation to physics-constrained generation, and how its abstractions fit together.

Introducing PhysicsNeMo-Mesh: GPU-Accelerated Mesh Processing for Scientific ML

PhysicsNeMo-Mesh is a GPU-accelerated mesh processing module, built on PyTorch and TensorDict, and included in the open-source PhysicsNeMo library. It provides a) a GPU-native, high-performance mesh data structure with design choices that are particularly well-suited for ML workflows, and b) a diverse suite of accelerated mesh operations that can be used to bridge existing gaps between data preprocessing and training/inference.

Its native data format allows loading meshes from disk much more quickly than VTU (9x-88x faster in our testing; hardware-dependent) while preserving the full mesh structure that flat tensor formats like zarr typically discard. This gives you the speed you need for training while retaining the geometric and physical context needed for model- and dataset-agnostic workflows. In this post, we'll show what PhysicsNeMo-Mesh can do for your training pipeline.

Building AI Surrogate Models for Structural and Automotive Crash Analysis with NVIDIA PhysicsNeMo

Simulating automotive crash is one of the most complex computational workloads in computer-aided engineering (CAE). Crash simulation is a highly non-linear, dynamic event that happens in milliseconds. Therefore automotive crash simulations face significant bottlenecks in both human-driven workflows and raw computational demands. These challenges stem from the need to model highly non-linear, split-second events with extreme accuracy to meet strict regulatory standards (like NHTSA or Euro NCAP).

This post explains how to build an AI surrogate model for structural and automotive crash analysis with NVIDIA PhysicsNeMo, from finite element data curation and architecture selection through temporal training and validation.