AI 中文总结
研究针对CFD中机器学习代理预测无信任判断问题，通过计算物理残差实现双向分离，残差作信任信号效果良好，还部署监督深度平衡校正器降低MSE，贡献了自我审计信任层等内容。
AI 中文摘要
用于计算流体动力学（CFD）的机器学习代理预测稳定流场的速度比传统求解器快几个数量级，但只输出单个场，且没有内置方法来判断是否值得信任，尤其是在分布外情况。我们结合控制物理来闭环：计算预测的离散稳态RANS残差并探究其作用。核心发现是一种双向分离：物理残差是可靠的、骨干鲁棒的信任信号，但却是较差的校正目标。作为信任信号，残差与场误差的逐例秩相关性在三种不同架构骨干上始终为正，并能推广到第二个数据集和流态。一个分裂共形层达到目标覆盖率，与深度集成西格玛配对产生输入自适应带。作为校正目标或接受门时残差失败。我们还部署了一个监督深度平衡校正器，它能降低SOTA骨干上所有三个种子的体积场MSE。我们明确报告了注意事项：校正质量依赖于骨干，覆盖率保证在可交换性下成立。贡献在于一个自我审计信任层、残差的两个作用以及伴随的学习自校正。
英文摘要
Machine-learning surrogates for computational fluid dynamics (CFD) predict steady flow fields far faster than classical solvers, but emit a single field with no built-in way to know whether to trust it -- especially out of distribution. We make the surrogate audit itself against the governing physics: we compute the discretised steady-state Reynolds-averaged (RANS) residual of the prediction and ask what jobs it can do. Our central finding is a clean two-way dissociation: the physics residual is a reliable, backbone-robust trust signal (it tells you where the prediction is wrong) but a poor correction objective (it does not tell you how to fix it). As a trust signal, the residual's per-case rank correlation with field error is consistently positive across three architecturally distinct backbones and a second dataset of laminar bluff bodies; it flags the worst-decile cases with AUROC $\approx 0.9$, and a distribution-free split-conformal layer attains its target coverage. As a correction objective, the residual fails: reducing it does not reduce field error. The monotone-residual acceptance gate built on it, by contrast, does help. Separately, a learned deep-equilibrium corrector trained toward ground truth lowers volume-field error on a state-of-the-art backbone; a controlled ablation that removes the residual input matches this gain, so the improvement comes from the learned correction, not from conditioning on the residual. The contribution is a self-auditing, calibrated trust layer, the residual's two roles (trust signal yes, correction objective no), and the learned correction it accompanies, demonstrated on a competitive surrogate and released as the open-source neuroforge-cfd package.
Comments31 pages, 10 figures, 9 tables. v3 corrects v2's claim that the eddy-viscosity channel is effectively unlearned: measured, it is learned (R^2 0.996 grid backbone, 0.67-0.70 MeshGraphNet) and supplies 84% of nu_eff. Also closes the no-slip-weight objection to the residual-floor result in closed form, and adds a wall-clock cost measurement. Code: https://github.com/ali-kin4/neuroforge-cfd
