用机器学习生成更优初始流场，让瞬态流体模拟快一半。

Transient computational fluid dynamics (CFD) simulations are essential for many industrial applications, but suffer from high compute costs relative to steady-state simulations. This is due to the need to: (a) reach statistical steadiness by physically advecting errors in the initial field sufficiently far downstream, and (b) gather a sufficient sample of fluctuating flow data to estimate time-averaged quantities of interest. We present a machine learning-based initialization method that aims to reduce the cost of transient solve by providing more accurate initial fields. Through a case study in automotive aerodynamics on a 17M-cell unsteady incompressible RANS simulation, we evaluate three proposed ML-based initialization strategies against existing methods. Here, we demonstrate 50% reductions in time-to-convergence compared to traditional uniform and potential flow-based initializations. Two ML-based initialization strategies are recommended for general use: (1) a hybrid method combining ML predictions with potential flow solutions, and (2) an approach integrating ML predictions with uniform flow. Both strategies enable CFD solvers to achieve convergence times comparable to computationally-expensive steady RANS initializations, while requiring far less wall-clock time to compute the initialization field. Notably, these improvements are achieved using an ML model trained on a different dataset of diverse automotive geometries, demonstrating generalization capabilities relevant to specific industrial application areas. Because this Hybrid-ML workflow only modifies the inputs to an existing CFD solver, rather than modifying the solver itself, it can be applied to existing CFD workflows with relatively minimal changes; this provides a practical approach to accelerating industrial CFD simulations using existing ML surrogate models.