As humankind we strive to understand nature and learn from it. Over the years our understanding has grown, from the laws governing the physical world to the inner workings of the human genome. We are closer than ever to making this knowledge accessible and usable by AI, opening a new era for AI systems exceeding today’s success stories in vision and translation and opening the door for countless new industrial applications.

Inspired by nature

Have you ever watched an eagle fly? Gliding over snow-crusted mountain ridges or soaring over the expanses of the tundra? It almost looks as if those impressive birds are being carried by air current.

What seems natural and intuitive to the eagle is a flying style perfected to save energy by adjusting wing beat to air current to control the flight pattern. But how do eagles do this? How are they able to develop an intuitive understanding of the complex physics behind air flow and turbulence and exploit it to their advantage? And most of all, how are they able to learn this within weeks [1]?

Challenges of today’s AI methods

Within the field of artificial intelligence (AI), we try to mimic this natural learning behavior humans and animals are born with to enable learning of complex behavior we cannot explicitly program. Many currently successful AI methods e.g., Deep Neural Networks, rely on data with examples of real-world behavior to learn. For most applications, the amount of data needed for the training process of such networks is huge – much bigger than the few weeks of flying practice a young eagle gets.

For use cases where it is possible to gather these huge amounts of data, today’s AI methods typically work well. We have, for instance, seen them excel at vision problems, like grasping of objects [2] or visual end of line inspection on the shop floor. These methods, however, fail to learn complex behavior that we do not have a lot of data for. Let us look at an example from engineering design: Engineers have been exploiting physics of airflow to design more aerodynamic cars, airplanes, or trains. By explicitly describing the physics, i.e., the forces acting on the system at a given time, engineers can design objects to work optimally for those environments. However, even though we can model the complexity of air flow with simulation tools, standard neural networks miss to extrapolate from a limited number of simulations to predict airflow around an arbitrary object.

Visualization of aerodynamic forces acting on the SimRod car and its passengers during testing in a wind tunnel, rendered using Nvidia Omniverse. In front: car design with windshields; in the back: car design without windshields

Introducing physics to AI

When trying to overcome these barriers in today’s AI methods researchers turn back to an old principle of technology development – understanding nature and learning from it. They ask, how can we build AI systems that can tap into the knowledge we have about the physics of our world?

This question led to the development of a new set of AI methods referred to as physics-aware or physics-informed AI [3]. Instead of providing data with examples of real-world behavior to our AI systems we are now capable of expressing physics-based constraints in the training process of our AI. Just like the eagle does not steer against the forces in the air, but uses them to accelerate, we can use what we know about the physics of our world to guide our AI to an optimal solution.

Example architecture of a Physics-Informed Neural Network

Physics-informed AI approaches open up the realm of possible industrial applications for AI. They allow us to address a new more complex set of problems that AI was so far not able to find generally applicable solutions for, for example motion planning and control or fast simulation. Moreover, we can build AI systems that act in accordance with the physics of our world.

This ability becomes increasingly more important as we move industrial applications into the metaverse. The metaverse, a virtual, photo-realistic and real-time representation of our physical world, adequately represents how things behave in the real world and, thus, understands the physics behind it. AI systems acting in the metaverse will have to do the same.

References:

[1] Adams, Gryphon. “How Do Eagles Learn to Fly?” Pets on Mom.com, 19 Nov. 2020, https://animals.mom.com/how-do-eagles-learn-to-fly-3179849.html.

[2] Solowjow, Eugen, et al. “Industrial robot grasping with deep learning using a programmable logic controller (plc).” 2020 IEEE 16th International Conference on Automation Science and Engineering (CASE). IEEE, 2020.

[3] Raissi, Maziar, Paris Perdikaris, and George E. Karniadakis. “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.” Journal of Computational physics 378 (2019): 686-707.