NVIDIA Unveils ‘Modulus’: A Framework For Developing Physics-Machine Learning (ML) Models for Digital Twins

NVIDIA unveils ‘Modulus’, a new framework for constructing Physics-Machine Learning (ML) Models for Digital Twins. NVIDIA’s Modulus, formerly known as SimNet, is a platform that allows engineers, scientists, researchers, and students to train neural networks utilizing governing physics equations combined with observed or simulated data.

NVIDIA’s Modulus is a neural network architecture that combines the capabilities of physics and partial differential equations (PDEs) with artificial intelligence (AI) to create more robust models for better analysis. Modulus can help you get started with AI-driven physics challenges or create digital twin models for complicated non-linear, multiphysics systems’.

Modulus uses an AI-based technique to combine the advantages of physics with machine learning. Modulus trains a neural network that encapsulates the physics of the system into a high-fidelity model that can be employed in numerous applications using training data and the governing physics equations. 

The data preparation module of Modulus accepts observed or simulated data and the geometry of the system we’re trying to model in a variety of standard forms, including point cloud format. The brilliance of Modulus is that it takes not just the system’s unique geometry but also the input geometry’s explicit parameterized space. This enables the trained model to search for the best parameters in the design space.

Modulus provides Python-based APIs for constructing physics-informed neural networks using symbolic governing PDEs. It offers curated layers and an ever-expanding variety of network topologies that have been proved to work for physics-based issues.

The training engine module accepts all inputs and trains the final model using PyTorch and TensorFlow, cuDNN for GPU acceleration, and Magnum IO for multi-GPU/multi-node scalability.

Neural Networks Based on Physics

Modulus teaches neural networks how to describe the behavior of complex systems in a variety of domains using fundamental physics equations. The surrogate model may then be employed in various digital twin applications, ranging from industrial applications to climate science.

Modulus has a data preparation module, which, like most AI-based systems, aids in the management of observed or simulated data. It also considers the geometry of the methods it models, as well as the explicit characteristics of the space that the input geometry represents.

Modulus’ primary process and components are as follows:

  • A sampling planner allows the user to choose a technique to enhance the trained model’s convergence and accuracies, such as quasi-random sampling or importance sampling.
  • Python-based APIs for constructing physics-based neural networks using symbolic governing partial differential equations.
  • Layers and network topologies have been demonstrated to work for physics challenges.
  • The Physics-ML engine uses PyTorch and TensorFlow to train the model and cuDNN for GPU acceleration and NVIDIA Magnum IO for multi-GPU and multi-node scalability.

Quick Response Time

The GPU-accelerated toolbox enables faster insights by augmenting traditional analysis with a quick turnaround. Modulus allows users to test the impact of modifying a system’s settings on various setups and scenarios.

Modulus is a high-performance modulus. The TensorFlow-based version improves performance by utilizing XLA, a domain-specific linear algebra compiler that accelerates TensorFlow models. For multi-GPU scalability, it employs the Horovod distributed deep learning training framework.

Simple Adoption

Modulus is adaptable and straightforward to use. It provides APIs for adding additional physics and geometry to games. It’s meant to help people who are just getting started with AI-driven digital-twin applications get up and running quickly.

Step-by-step instructions for getting started with computational fluid dynamics, heat transport, and other topics are included in the framework. It also consists of an increasing number of implementations for turbulence modeling, short wave equations, Navier-Stokes equations, Maxwell’s equation for electromagnetics, inverse difficulties, and other multiphysics issues.

Benefits and characteristics of NVIDIA Modulus

  • Accelerated linear algebra (XLA), automated mixed precision (AMP) capabilities, and multi-GPU/multi-node implementation solve more significant problems quicker.
  • Models several physics types with precision and convergence in forward and inverse simulations.
  • Provides a parameterized system representation that simultaneously solves various situations, allowing you to train once and solve several problems.
  • Provides thorough user guide examples and application programming interfaces (APIs) for building new physics and geometry.

NVIDIA has offered Modulus free to download.


Prathamesh Ingle is a Mechanical Engineer and works as a Data Analyst. He is also an AI practitioner and certified Data Scientist with an interest in applications of AI. He is enthusiastic about exploring new technologies and advancements with their real-life applications

[Announcing Gretel Navigator] Create, edit, and augment tabular data with the first compound AI system trusted by EY, Databricks, Google, and Microsoft