OP here. I’m an independent researcher working on computational fluid dynamics.
The Problem: Direct Numerical Simulation (DNS) of turbulence usually scales as O(N^3), meaning memory requirements explode as you increase resolution.
The Solution: I wrote a solver using Quantized Tensor Trains (QTT) and found that the representational rank of the velocity field saturates (\alpha \approx 0).
Result: I simulated 256^3 turbulence on a single consumer GPU with >10,000x compression compared to standard dense grids, while maintaining spectral accuracy.Happy to answer questions about the spectral solver, the PyTorch implementation, or the QTT format.
INVARIAN•2h ago
The Problem: Direct Numerical Simulation (DNS) of turbulence usually scales as O(N^3), meaning memory requirements explode as you increase resolution.
The Solution: I wrote a solver using Quantized Tensor Trains (QTT) and found that the representational rank of the velocity field saturates (\alpha \approx 0).
Result: I simulated 256^3 turbulence on a single consumer GPU with >10,000x compression compared to standard dense grids, while maintaining spectral accuracy.Happy to answer questions about the spectral solver, the PyTorch implementation, or the QTT format.