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.
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.