The flame's chemistry lives in five dimensions. The tensor train fits it in memory.
Unsteady flamelet progress variable manifolds — pre-computed lookup tables that map reduced parameters (mixture fraction, progress variable, dissipation rate, pressure, enthalpy) to full thermochemical states (temperature, species concentrations, reaction rates). Instead of solving hundreds of coupled chemical equations at every grid point and time step, the simulation looks up the answer in the table (arXiv:2603.20240).
The problem: five-dimensional tables grow exponentially with resolution. A dense table at reasonable resolution for high-speed reacting flows exceeds available memory. Reduce the resolution and you lose the coupled thermochemical effects the manifold is meant to capture.
Tensor train decomposition: represent the five-dimensional table as a chain of three-dimensional tensors. The compression ratio depends on the table's intrinsic rank — how much the dimensions interact. For flamelet manifolds, the intrinsic rank is moderate: the dimensions are correlated (higher pressure shifts the chemistry) but not arbitrary (the physics constrains the correlations). Significant memory savings while maintaining manifold accuracy. Up to 2.4× faster sampling than dense tensor evaluation.
The structural insight: the combustion chemistry looks five-dimensional but isn't — it has lower effective dimensionality because physics constrains how the parameters interact. The tensor train exploits this constraint: it stores only the correlations that exist, not all correlations that could exist. The compression isn't lossy in the usual sense (throwing away small values). It's structural (representing the table's intrinsic rank instead of its apparent dimensionality). The physics that makes the chemistry tractable is the same physics that makes the table compressible.