Finite Element-Based Machine Learning for Modeling and Prediction of Manufacturing Processes: Applications in Rolling and Milling

Abstract

The global transition toward smart manufacturing is increasing the demand for modeling approaches that can estimate process parameters with high accuracy and computational efficiency. Conventional methods, including physical experimentation, empirical or analytical formulations, and finite element (FE) simulations, remain essential but are often costly, time-consuming, or limited in their generalizability. These limitations can hinder rapid process optimization and adaptive decision-making in modern manufacturing. This thesis develops machine learning (ML) frameworks that combine FE simulations with domain knowledge to enhance predictive performance and reduce reliance on repetitive physical testing and large-scale numerical simulations. Two representative processes are investigated in depth: wire rod rolling and milling. In wire rod rolling, a dimensionless ML framework is proposed to predict the lateral spread after each rolling stand which is a critical parameter governing roll-pass design and process optimization. FE simulations are employed to generate a diverse dataset that captures the influence of geometric, material, and process variables. Through a physics-based feature design that ensures dimensionless inputs and outputs, the trained ML models achieve high accuracy and consistent generalization across the studied rolling stands. The results demonstrate that the framework can serve as a transferable and data-efficient alternative to conventional empirical or resource-intensive simulation methods. In milling process, chatter vibrations are a primary barrier to productivity. Manufacturers often adopt conservative cutting conditions to avoid chatter, which limits material removal rates. Although chatter stability lobe analysis can improve performance, its dependence on complex measurements limits its industrial application. This thesis proposes two complementary ML approaches that reduce measurement requirements while maintaining predictive capability. The first develops ML models trained on FE data to estimate milling force coefficients, minimizing the need for table dynamometers and extensive cutting tests. The second introduces an ML-based substructure coupling framework that predicts tool tip dynamics and chatter stability through a combination of simulation data and limited experimental input, allowing rapid tool tuning and selection of stable cutting parameters. Validation results show that both models achieve accurate predictions for their respective tasks.