Reproductions of published results in flow physics and data-driven modelling of fluid systems. Each folder targets one paper, reports a specific number against the published value, and documents the choices the paper left unspecified.
Failed and partial reproductions are kept, not deleted. The gap between a paper and a working implementation is usually the interesting part.
| # | Paper | Target metric | Published | Reproduced | Status |
|---|---|---|---|---|---|
| 1 | Fukagata & Fukami — CNN-AE | (fill in) | (fill in) | (fill in) | — |
| 2 | Pathak et al., PRL 120, 024102 (2018) — reservoir computing on Kuramoto–Sivashinsky | valid prediction time, Fig. 2 config | no published number | 2.6 ± 0.4 Lyapunov times | method reproduced; Fig. 4 (the ~8 Lyap result) not attempted |
Each reproduction folder follows the same layout:
<author>-<year>-<method>/
├── README.md paper reference, parameter table, results, deviations
├── src/ implementation
├── figures/ generated output
└── notes/ theory worked through, derivations, dead ends
Every folder README states three things explicitly:
- What the paper specifies — the parameters lifted directly from the text
- What it doesn't — hyperparameters absent from the paper, and the value chosen
- What matched and what didn't — with numbers, not adjectives
Python ≥ 3.10, numpy, scipy, matplotlib. PyTorch only where a reproduction
requires it. Generated data (*.npz, *.npy) is gitignored — every dataset here
is reproducible from the scripts with a fixed seed.