An Evaluation of Physics-Informed Learning With General Neural Operator Transformers

What is it about?

his study tested a advanced AI model (a transformer) to predict how air or fluid moves in a specific low-speed flow scenario (a cavity flow). The researchers used two methods: one where the AI was trained with data alone (supervised learning), and another where they added physics-based rules to guide the training (physics-informed loss). The purely data-trained AI had a very small error (0.63%) in its predictions. When they added the physics rules, the error increased slightly (1.54%), but the AI did a better job predicting important flow features, like small swirling patterns (secondary vortices) and behaviour near the edges of the flow. The higher error with the physics rules was due to limitations in the computational grid used, which was constrained by the AI's memory needs. The AI did well overall, and adding physics rules helped it capture more realistic flow details, even though it introduced a bit more error due to technical limitations.

Featured Image

Photo by Pawel Czerwinski on Unsplash

Photo by Pawel Czerwinski on Unsplash

Why is it important?

A lot of generative AI focuses on creating picture perfect predictions. With fluid dynamics however, a visually similar solution does not necessarily imply a physically accurate solution. This paper is novel in that it not only studies the effectiveness of introducing physics informed training to a contemporary architecture, but also the efficacy of using traditional machine learning metrics to assess models for physical problems.