A staged machine learning framework for daily quality prediction in foundry sand moulding

Metal casting by sand molds is one of the most fundamental manufacturing processes. A foundry is where metal castings are produced, and these castings serve as raw material for subsequent processing. Sand molds have conflicting properties, so this research focuses on an optimization problem: balancing these properties to produce high-quality metal castings. Controlling these sand-quality requirements in foundries through daily sand testing for different quality characteristics is laborious and makes monitoring difficult. This research focuses on a sequential machine learning approach that predicts the eight responses (total clay, active clay, volatile matter, loss on ignition, green compression strength, compactability, permeability, wet tensile strength) from available plant data (1197 samples, 2021–2025) in a way that allows us to discriminate real process-driven predictability from plain persistence. The five developed models are listed in a chronological six-month holdout test: persistence, process-only, lab history-only, process + lagged lab history, and selected features. A separate validation window was used to compare Random Forest and XGBoost; Random Forest outperformed XGBoost on all eight outcomes in that window and was fixed as the sole algorithm for the final holdout evaluation. The process-only model performs poorly with a mean R 2 of −0.43 relative to 0.53 for persistence. Adding level-driven process variables to lab history increases the mean R² by 16%, from 0.43 to 0.496. Integrating process variables with laboratory history and delta-target formulation enhances the R² to 0.559. Response-driven enhancements showed the largest increase in compactability (+0.49), while volatile matter and permeability showed no improvement. The prediction intervals show decent calibration (86–97%), but compactability shows error drift and a 20–50% failed-batch detection rate. Notably, many outputs pose non-normal, autocorrelated, or heteroscedastic residuals. The results showed the developed supervised model is suitable for shadow-mode deployment that enhances decision-making rather than replacing laboratory sand quality testing.