FlowReconBench benchmarks flow reconstruction from sparse sensors
Original title: FlowReconBench: Accuracy, calibration and cost of learned and classical flow reconstruction from sparse sensors
FlowReconBench shows classical methods lead in low-sensor regimes while generative models improve ensembles at higher compute cost.
FlowReconBench evaluates ten learned and five classical methods on laminar wakes, Kolmogorov turbulence and 3D isotropic flows. Classical methods perform strongly in dense-sensing regimes and for unobserved variables like pressure. Data-driven methods become superior at decreasing sensor density or observability. Generative models balance pointwise error and ensemble spread via sampler budgets.
- Classical methods lead in low-order regimes
- Generative models outperform at sparse sensors
- Scaling shows memory-efficient queryable methods
- Continuous ranked probability score ranks differently from pointwise error
From the original article
1 min read at arxiv.orgReconstructing full flow fields from sparse sensors requires prior assumptions about unmeasured regions and physical variables. Yet it remains unclear when learned models outperform classical estimators and which aspects of their predictions can be trusted.
By Nicolas Tricard · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.