Nvidia launches Modulus, a framework for developing ‘physics-grounded’ AI models

Nvidia today launched Modulus, a framework for developing “physics-grounded” machine learning models in industries that require a high level of physical accuracy. Modulus trains AI systems to use the laws of physics to model the behavior of systems in a range of fields, according to Nvidia, including climate science and protein engineering.

“Digital twin” approaches to simulation have gained currency in many domains. For instance, London-based SenSat helps clients in construction, mining, energy, and other industries create models of locations for projects they’re working on. GE offers technology that allows companies to model digital twins of actual machines and closely track performance. And Microsoft provides Azure Digital Twins and Project Bonsai, which model the relationships and interactions between people, places, and devices in simulated environments.

Gartner predicted that 50% of large manufacturers would have had at least one digital twin initiative launched by 2020, and that the number of organizations using digital twins would triple from 2018 to 2022. Markets and Markets estimates that the global market for digital twin technologies will reach $48.2 billion by 2026, up from $3.1 billion in 2020.

Above: A physics simulation running with Nvidia Modulus.

Image Credit: Nvidia

“Digital twins have emerged as powerful tools for tackling problems ranging from the molecular level like drug discovery up to global challenges like climate change,” Nvidia senior product marketing manager Jay Gould said in a blog post. “Modulus gives scientists a [toolkit] to build highly accurate digital reproductions of complex and dynamic systems that will enable the next generation of breakthroughs across a vast range of industries.”

Physics framework

Nvidia describes Modulus — which was announced during the company’s fall 2021 GPU Technology Conference (GTC) — as a framework to provide engineers, scientists, and researchers tools to build AI models of digital twins. As in most AI-based approaches, Modulus includes a data prep module that helps manage observed or simulated data, accounting for the geometry of the systems it models and the explicit parameters of the space represented by the input geometry.

Modulus includes a sampling planner that enables users to select an approach, such as quasi-random sampling or importance sampling, to improve the model’s accuracy. The framework also ships with APIs to take symbolic governing partial differential equations and build physics models, as well as curated layers and network architectures tailored for physics-based problems.

In addition, Modulus offers a “physics-machine learning” engine that takes inputs to train models using machine learning frameworks including Facebook’s PyTorch and Google’s TensorFlow. The toolkit’s TensorFlow-based implementation optimizes performance by taking advantage of XLA, a domain-specific compiler for linear algebra that accelerates TensorFlow models. Leveraging the Horovod distributed deep learning training framework for multi-GPU scaling, Modulus can perform inference in near-real-time or interactively once a model is trained.

Modulus includes tutorials for getting started with computational fluid dynamics, heat transfer, modeling turbulence, transient wave equations, and other multiphysics problems. It’s available now as a free download through the Nvidia Developer Zone.

“The GPU-accelerated toolkit offers rapid turnaround complementing traditional analysis, enabling faster insights. Modulus allows users to explore different configurations and scenarios of a system by assessing the impact of changing its parameters,” Gould wrote. “Modulus is customizable and easy to adopt. It offers APIs for implementing new physics and geometry. It’s designed so those just starting with AI-driven digital twin applications can put it to work fast.”

Companies including Alphabet’s DeepMind have investigated applying AI systems to physics simulations. Last April, DeepMind described a model that predicts the movement of glass molecules as they transition between liquid and solid states. Beyond glass, the researchers asserted that the work could lead to advances in industries like manufacturing and medicine, including dissolvable glass structures for drug delivery and renewable polymers.