Predicting Time-Dependent Flow Over Complex Geometries Using Operator Networks

Summary

This research introduces a new way to quickly and accurately predict how fluids flow around different shaped objects, even as conditions change over time.

What's the problem?

Predicting fluid flow, especially around complex shapes and as time progresses, is computationally expensive and slow using traditional methods like Computational Fluid Dynamics (CFD). Existing attempts to speed things up with machine learning often struggle to generalize to new shapes or accurately capture the flow's evolution over longer periods.

What's the solution?

The researchers developed a 'Deep Operator Network' – a type of artificial intelligence – that learns to predict velocity fields in fluids. It takes into account both the shape of the object (using something called a signed distance field) and the flow's history (using a convolutional neural network). They trained this AI on a large dataset of accurate fluid flow simulations and then tested it on new, unseen shapes. The result is a model that can predict flow fields with about 5% error and is up to 1000 times faster than traditional CFD.

Why it matters?

This work is important because it offers a significantly faster way to simulate fluid flow, which has applications in many fields like engineering design, weather forecasting, and even medical simulations. The speedup allows for quicker testing of designs and a better understanding of complex fluid dynamics. The researchers also openly shared their code, allowing others to build upon and improve their work.