Simulating Three-dimensional Turbulence with Physics-informed Neural Networks Turbulent fluid flows are among the most computationally demanding problems in science, requiring enormous computational resources that become prohibitive at high flow speeds. Physics-informed neural ...

7/7 Read the preprint here: arxiv.org/abs/2507.08972

#PINNs #CFD #Turbulence #ScientificComputing #MachineLearning #DOE #ASCR #AppliedMathematics #Yale #UPenn #PNNL

6/7 Led by the outstanding work of Sifan Wang at Yale, with key contributions from Panos Stinis at PNNL and Shyam Sankaran at UPenn. Supported by DOE Advanced Scientific Computing Research program.

5/7 This work demonstrates that PINNs can handle complex chaotic systems, though computational efficiency remains an important area for future improvements. Opens possibilities for mesh-free modeling and hybrid approaches.

4/7 Validated on three challenging benchmarks:
– 2D Kolmogorov flow (Re = 10⁶)
– 3D Taylor-Green vortex (Re = 1,600)
– 3D turbulent channel flow (Re_τ = 550)

Results accurately reproduce key turbulence statistics including energy spectra, enstrophy, and Reynolds stresses.

3/7 Key ingredients include:
– PirateNet architecture for deep networks
– Causal training strategies
– Self-adaptive loss weighting
– SOAP optimizer for resolving gradient conflicts
– Time-marching with transfer learning

2/7 For the first time, we show that PINNs can simulate fully developed turbulent flows in 2D and 3D by learning solutions directly from the Navier-Stokes equations without training data or computational grids.

1/7 Turbulent flows remain computationally challenging due to their multiscale, chaotic nature. Traditional methods like DNS and LES scale poorly with flow complexity, motivating exploration of alternative approaches.

“Can Physics-Informed Neural Networks (PINNs) simulate 3D turbulence?” A question we've been asked repeatedly since developing the framework in 2017. After nearly a decade of progress, we now have a conclusive answer. Thread below 🧵

Special shoutout to our core contributors for making this possible!

📝Read the full paper: arxiv.org/abs/2405.13063

💻Open-source model & weights: github.com/microsoft/au...

🌏Aurora represents a major step toward making accurate Earth system predictions accessible to everyone. Huge thanks to our collaborators at @msftresearch.bsky.social, SilurianAI, University of Amsterdam, @cambridge-uni.bsky.social, @pennengineering.bsky.social & beyond! 🙏

🚀 These results significantly expand Aurora’s capabilities that already include: 3️⃣ 5-day global air pollution predictions at 0.4° resolution, outperforming CAMS on 74% of targets; 4️⃣ 10-day 0.1° weather forecasts, outperforming IFS HRES on 92% of all targets.

2️⃣ Ocean wave dynamics 🌊: 10-day global forecasts at 0.25° resolution that beat IFS HRES-WAM on 86% of targets. Below: Aurora's accurate prediction of wave patterns during Typhoon Nanmadol👇

1️⃣ Tropical cyclone tracking 🌀: First AI model to outperform seven operational forecasting centers across four basins up to 5 days ahead! Example: Aurora correctly predicted Typhoon Doksuri's landfall when others missed 👇

🎉Excited to announce major new breakthroughs in our Aurora foundation model! Our team (@cbodnar.com, @wessel.ai, @megstanley.bsky.social, @a-lucic.bsky.social, Anna Vaughan) has achieved unprecedented results across multiple Earth system forecasting tasks. Here's what's new... 🧵