Accurate yet rapid defect prediction is vital for cost-effective sheet metal design. Using the Deep Drawing and Cutting Simulations (DDACS) dataset of DP600 deep drawing, data-driven regressors emulate finite-element simulations to predict three targets: maximum equivalent plastic strain (EPS_Max), minimum thickness, and maximum springback. Ten machine learning regressors were screened across four preprocessing pipelines. Top transformers advanced to Stage-2 validation with 15 randomized 70/30 splits. Extreme gradient boosting showed the highest fidelity (CM = 0.045 for minimum thickness and 0.106 for springback). Gaussian process regression performed best for EPS_Max (CM = 0.089). CM is a Euclidean score combining RMSRE, Max Error, MAE, and (1−R²), where lower values are better. Feature-importance analysis highlighted blank-holder force, friction, and radius as dominant predictors. Relative to a quadratic response surface methodology baseline, CM fell by 50–90% and MAE reached 10⁻³. These results support reliable, near-real-time defect prediction for sheet metal design.
