Casting defects, such as shrinkage porosity, result in rejections, decreasing productivity, and profitability. Physics-based sand casting simulations are widely used for defect prediction, but gating optimization—critical for minimizing porosity and maximizing yield—typically requires extensive simulations with software tools such as ADSTEFAN. This paper presents a machine learning (ML)-driven approach that reduces the number of simulations by faster shrinkage porosity prediction, casting rejections reduction, and casting yield enhancement. A fractional factorial Design of Experiments (DOE) generated a reduced defect dataset, which trained ML models to predict defects and yield. Genetic algorithms were employed to rapidly assess new geometries of gating design, identifying optimal designs. Promising designs were validated through targeted physics-based casting simulations. The hybrid ML simulation approach reduced computational effort and accelerated design optimization, balancing yield and defect minimization, and achieving optimal designs with fewer simulation runs.