An Integrated NLP-ML Framework for Property Prediction and Design of Steels

This paper introduces a data-driven framework designed to accelerate the property prediction, and hence the discovery and development of steels. The study leverages natural language processing and machine learning techniques to analyse a dataset containing steel compositions, (thermo)mechanical processing, and mechanical properties. By integrating unsupervized machine learning for process classification (via natural language processing-assisted clustering) and supervised regression models for property prediction, the framework enables an efficient exploration method to study steels. A predictive accuracy of R² > 0.85 was achieved, with mean absolute errors <15 MPa for both yield and ultimate tensile strength. A cloud-based graphical user interface was developed to facilitate user interaction, allowing researchers to input steel processing and composition to receive predictive insights on mechanical properties. The findings demonstrate the framework to support circular economy principles by reducing trial-and-error experimentation, with a view to accelerating the design of steels, and promoting sustainable steel innovation.