Ml-driven approaches to enhance inventory planning: Inoculant weight application in casting processes

Abstract

This study addresses the complexities of inventory management in the casting industry, focusing on a proprietary inoculant crucial for casting process. The unpredictable nature of inoculant usage, influenced by various casting variables, challenges conventional calculations. To overcome these, expert opinions were gathered to identify features for accurate prediction of inoculant requirements, and an annual dataset was created. This dataset underwent data preprocessing to introduce new features and mitigate outliers. Machine learning models, including decision tree, random forest, least absolute shrinkage and selection operator (LASSO), elastic net regression, ridge regression, multiple linear regression, gradient boosting, extreme gradient boosting, and voting regression, were trained and tested. The voting regression algorithm, which combines decision tree and LASSO through hard voting, demonstrates that combining high-variance low-bias and low-variance high-bias algorithms can achieve competitive performance with low errorwithout overfitting. Predictions obtained using the voting regression algorithm for the entire year’s data were analyzed using batch size determination algorithms, Wagner-Whitin and Silver-Meal, with Wagner-Whitin providing cost advantage. This study highlights the potential of predicting materials with uncertain usage through AI techniques and historical data, minimizing inventory cost. The findings offer practical insights for managers, demonstrating how advanced predictive models can enhance decision-making, improve inventory planning, and achieve cost savings. This approach is valuable without needing additional investment in material requirement planning, especially where bill of materials’ reliability is reduced by environmental factors.