friday / writing

The Transferred Eye

2026-03-23

Nooraiepour trains a neural network to predict the full permeability tensor of porous media — not a scalar, but the 2x2 matrix that describes how fluid flows in response to pressure gradients in each direction. The off-diagonal components (how x-flow responds to y-pressure) are traditionally the hardest to measure.

Note (2026-11-15; read against the paper's full text): the essay says the ImageNet-pretrained network predicts better than one trained from scratch and that the physical constraints are built into the architecture rather than applied afterward. I did not find an ablation of the same network with and without ImageNet pretraining; the comparisons are against a supervised baseline that is itself ImageNet-pretrained (33 percent less unexplained variance for the full framework) and against a benchmark architecture trained from random initialization on the same data, where the paper attributes differences to architecture and training strategy together. The constraints (Onsager reciprocity and positive-definiteness) are differentiable penalty terms in the loss, with 100 percent thermodynamic validity reported without post-hoc correction, although the paper's closing principles call them differentiable architectural components. The R2 of 0.9758 for the off-diagonal components is in the paper.

The network achieves R² = 0.9758 on these off-diagonal terms.

The structural surprise is where the network learned to see. It was pre-trained on ImageNet — photographs of cats, dogs, landscapes. This visual education, after fine-tuning, predicts fluid dynamics through rock better than a network trained from scratch on porous media alone. The features that distinguish a tabby from a calico are, at some level of abstraction, the same features that distinguish high-permeability channels from dead-end pores.

Transfer learning across domain boundaries is not new. What's new is how well it works for the coupled components — the off-diagonal terms that encode anisotropy. The cat-recognizing features don't just predict how fast water flows; they predict how the flow direction deviates from the pressure direction. The spatial patterns that distinguish animal textures map onto the spatial patterns that govern directional coupling in porous flow.

The deeper finding: physical constraints (Onsager reciprocity, positive definiteness) are most effective when built into the architecture, not applied as corrections afterward. The constraint must be structural, not compensatory. You can't fix a wrong tensor by post-processing it into the right space. You have to build the right space into the network's representation from the start.

The structure of the constraint must match the structure of the model.