Bracketing Inference with Uncertainty Quantification: A Reliability Pipeline for Neural Aerodynamic Surrogates | Research Square
Data-driven surrogates for computational fluid dynamics (CFD) offer dramatic acceleration of the automotive design cycle, but their adoption requires assessing when a prediction can be trusted. As surrogate accuracy on public benchmarks approaches saturation, the barrier to deployment shifts from...
Data-driven surrogates for computational fluid dynamics (CFD) offer dramatic acceleration of the automotive design cycle, but their adoption requires assessing when a prediction can be trusted. As surrogate accuracy on public benchmarks approaches saturation, the barrier to deployment shifts from performance to reliability: a trained model never refuses to predict, and silent extrapolation on novel geometries poses a direct risk to engineering decisions. This study proposes a two-stage reliability pipeline that brackets inference with uncertainty quantification (UQ): a pre-inference…
saved by
related reading
- David Duvenaudcs.toronto.edu
- On Getting Confidence Estimates from Neural Networks | Bharath's notesbharathpbhat.github.io
- How well do truth probes generalise? — LessWronglesswrong.com
- Foundation Models for Oversight | Transluce AItransluce.org
- Predicting LLM Safety Before Release by Simulating Deploymentcdn.openai.com
- Large Language Models Must Be Taught to Know What They Don't Knowarxiv.org
- GRaM Competition @ ICLR 2026 | Competition track of the GRaM workshopgram-competition.github.io
- FrontierSWEfrontierswe.com
- Building confidence in Similesimile.com
- Scaling Laws That Extrapolate 300× Past the Fitopenathena.ai
- Not Pinning Your OpenRouter Provider Might Invalidate Your Research — LessWronglesswrong.com
- [2606.02609] Building Better Activation Oraclesarxiv.org