Researchers develop AI technique to predict yield strength of metals

The research team effectively addressed traditional cost and time limitations and developed an optimal artificial intelligence model for predicting the yield strength of various metals. Acta Materia.

Yield strength is the point at which a material, such as a metal, begins to deform when subjected to an external stress. In materials engineering, accurately predicting yield strength is essential for developing high-performance materials and improving structural stability. However, predicting this property requires considering many variables, such as the material’s grain size and type of impurities, and data collection typically requires extensive experiments over a long period of time.

To address this, the Hall-Petch equation is commonly used to establish the relationship between the yield strength of a material and the grain size, but this equation has limitations in accurately predicting the yield strength of new materials, taking into account their specific properties and various environmental conditions such as temperature and strain rate.

In this research, a team led by Professor Kim Hyun-seop of the Institute of Ferro-Eco-Materials Technology and Department of Materials Science and Engineering and PhD student Lee Jeong-ah of the Department of Materials Science and Engineering, in recent collaboration with Professor Figueiredo of the Department of Metallurgy and Materials Engineering at the Federal University of Minas Gerais in Brazil, combined physical theory and artificial intelligence (AI) techniques to increase accuracy while reducing the cost and time required to predict yield strength.

They developed a machine learning model that applies the mechanism of “grain boundary sliding” to describe how particles within a material move relative to one another, and a machine learning algorithm to predict yield strength.

First, the team used a black-box model to analyze the effect of different material properties on yield strength, then developed a white-box model with explicit inputs and outputs to improve the accuracy of yield strength predictions.

The research team validated the yield strength prediction model using a variety of iron-based alloys that were not included in the training data for the model, and demonstrated that even when predicted using untrained data, the model was highly accurate with a mean absolute error of 7.79 MPa compared to the actual yield strength.

“We have developed a general-purpose AI model that can accurately predict yield strength according to various experimental conditions and types of metal,” said Professor Kim Hyun-seop of POSTECH. “We will continue to actively utilize AI technology to make great strides in materials engineering research.”