Using neural networks to predict the low curves and processing maps of TNM-B1

Abstract

The ability to predict the behavior of a material is vital in both science and engineering. Traditionally, this task has been carried out using physics-based mathematical modeling. However, material behavior is dependent on a wide range of interconnected phenomena, properties and conditions. During deformation processes, work hardening, softening, microstructure evolution and generation of heat all occur simultaneously, and can either cooperate or compete. In addition, they can vary with the deformation temperature, applied force and process speed. As the complete picture of material behavior from the macroscopic scale to the atomic scale is not yet fully understood, deformation processes such as hot forg-ing can be difficult to handle using physics-based modeling. Usually, modeling the high temperature deformation behavior of metals consists of extracting characteristic points from the experimental flow curve data, and use them to fit the model equations through regression analysis. This is called phenomenological modeling, as it is based on the observations of a phenomena rather than being derived from fundamental theory. Alternatively, the data obtained from experiments could be used for a data-driven or machine learning (ML) approach to model the material behavior. An ML model would require no knowledge of the underlying physical phenomena governing a deformation process, as it can learn a mapping function which connects input to output based purely on the experimental data. In this work, the application of machine learning to modeling the flow curves of two different states of the titanium aluminide (TiAl) TNM-B1; hot isostatically pressed (HIPed) and heat treated, is investigated. Neural networks were used to learn a mapping function which predicted flow stress based on the inputs temperature, strain and strain rate. In addition, strain rate sensitivity maps and processing maps based on the experimental and the predicted data are analysed and compared. The results revealed that the neural networks were able to produce realistic and accurate flow curves, which fitted to the underlying behavior of the experimental data rather than the noise. The strain rate sensitivity and processing maps showed conflicting results. Good correlation was found for the HIPed material state between the ones based on experimental data and the ones based on predicted values, while there was a significant difference for the heat treated state.