Fourier Neural Operators for Rayleigh-Bénard Convection

The paper's core finding isn't the accuracy improvement itself, but the explicit demonstration that resolution generalization fails not because the model lacks capacity, but because training data resolution sets a hard ceiling. This reframes the problem from 'we need bigger networks' to 'we need denser training grids', a shift with direct implications for data collection strategy.

This connects directly to the optimization framework paper from two days ago (DSGNAR), which tackled ill-conditioned training in physics-informed neural networks. Both papers identify a specific bottleneck that architectural innovation alone cannot solve: DSGNAR showed that loss landscape conditioning was the constraint; this work shows that data resolution is the constraint for neural operators. Together they suggest the next frontier in scientific ML is not model design but training infrastructure (optimization, data generation, sampling strategy). The self-explainable operator learning paper also becomes more relevant here, since if resolution limits generalization, interpretability of what the model learned within those limits becomes more valuable for practitioners deciding whether to deploy.

If the authors release code and someone successfully trains the same architecture on higher-resolution Rayleigh-Bénard data (say, 2x or 4x the grid density) and achieves generalization to unseen resolutions, that confirms the resolution hypothesis. If they don't, or if accuracy plateaus despite denser training data, the bottleneck is elsewhere and the paper's claim needs revision.