Predicting strain-induced martensite in austenitic steels by combining physical modelling and machine learning


Technological advances have enabled the design of advanced materials to meet current needs. Austenitic stainless steels attract significant research attention as they are important engineering materials already with further potential to be used in additional numerous applications owing to their good corrosion resistance and mechanical properties. Their properties can be tuned by controlling the strain-induced martensitic transformation (SIMT). Thus, the control of SIMT is of great importance in optimizing its properties for various applications. Therefore, with the increasing computational materials design, accurate SIMT modeling is critical in current materials engineering. It is worth noting that SIMT and the related TRIP (transformation induced plasticity) effect is important also in other alloys containing the austenite phase.

Computational materials design has been extensively studied. As such, models for optimizing SIMT in steels, such as the Olson-Cohen (O-C), have been developed. However, these models are costly and time-consuming as they require experimental data fitting before application to different test conditions and materials. Based on the previous research findings, this problem can be solved by using a predictive model that only requires input on the specified test conditions and chemical composition to simulate the austenite stability towards SIMT. Lately, machine learning (ML) technology has been identified as a promising approach for predicting various material properties. In fact, ML (in the form of artificial neural networks) have been applied to model SIM formation in austenitic steels in some studies, even though the obtained results are unsatisfactory it still shows proof of concept. It is believed that combining physically-based modelling and ML can solve the remaining deficiencies and contribute to the development of an accurate predictive model.

Motivated by the previous results, researchers at the KTH Royal Institute of Technology and Ferritico AB, i.e. Dr. Wangzhong Mu, Professor Joakim Odqvist, Dr. Moshiour Rahaman, Dr. Felix Rios and Professor Peter Hedström utilized a combination of O-C based physical modeling and ML techniques to model SIMT in austenitic steels. Most importantly, they proposed a fully predictive model for predicting the SIMT fraction in austenitic stainless steels and associated alloys, taking into account the effects of strain, temperature and alloy composition. The original research article was published in the journal, Materials and Design.

In brief, the Swedish study was based on the three previously identified areas of improvement: a suitable size database to enhance the accuracy and applicability of ML, using state-of-the-art ML algorithms to treat small datasets, and using physically-based models to smoothly describe the relationship between SIMT fraction and applied temperature and strain. First, an experimental dataset correlating SIMT with strain, temperature, and chemical composition was collected from the open literature. Second, the dataset was expanded to 16500 entries using the O-C model predictions. Finally, ensemble ML methods were applied to model the data to produce a final model that was validated against the holdout dataset.

Using strain, temperature and chemical composition as input and SIM fraction as output, the results demonstrated the feasibility of the resulting models to predict both single- and multi-phase steels containing austenite. The final model accurately predicted SIM fractions for different strain values in the temperature range from -196 to 100 °C. The proposed model can readily be extended to account for more factors such as stress state, austenite grain size and strain rate. Furthermore, the authors noted that combining the model with kinetic and thermodynamics calculation could facilitate its application in evaluating the stability of austenite in multi-phase steels. In addition, combining the model with a finite-element code would enable the prediction of structure and properties of steels after processing such as cold forming.

In a nutshell, a predictive ML model for predicting SIM formation in austenitic steels was proposed based on a combination of ML ensemble technology and O-C physical modeling. The resulting model outperformed the existing ones in terms of accuracy and reliability. Moreover, it is versatile and can be adjusted to accommodate more critical features. It can also be combined with kinetic and thermodynamic simulations to facilitate steel and components design. In a statement to Advances in Engineering, the authors said the study can aid computational materials design, allowing for the development of high-performance austenitic stainless steel materials and related alloys.

About the author

Wangzhong Mu is a senior research scientist in metallurgy at the Department of Materials Science and Engineering, KTH Royal Institute of Technology, Stockholm, Sweden. He has obtained his PhD degree at KTH in 2015, subsequently he has worked in McMaster University, Canada, Tohoku University, Japan, and HiMat Engineering (now ‘Ferritico’, part-time). His research interest focuses on the non-metallic inclusion particles behavior in metallurgy, process-structure-property correlation in steels alloys and advanced material design and characterization. His aim is to apply the fundamental study to industrial-driven research.

About the author

Joakim Odqvist is professor in materials science at KTH Royal Institute of Technology, Stockholm, Sweden. He is the head of Unit of Structures. Also, he is the CSO and Co-founder of Ferritico, Sweden. His research interest focuses on materials modelling, correlation of structure and property in engineering materials, microstructure evolution in metallic materials, ceramics and hard metals, materials design, etc.


About the author

Peter Hedström is professor in materials science at KTH Royal Institute of Technology, Stockholm, Sweden. At the Department of Materials Science and Engineering he heads one of the research units, Unit of Properties. Hedström is also Director of the Center for X-rays in Swedish materials science (CeXS). His research interests relate to advanced materials characterization, structure-property relations in engineering materials, deformation of metals, phase transformations in inorganic materials, powder metallurgy, etc. Hedström is also active in bringing materials research into innovations, e.g. as co-founder of two companies Ferritico and Scatterin.


Mu, W., Rahaman, M., Rios, F., Odqvist, J., & Hedström, P. (2021). Predicting strain-induced martensite in austenitic steels by combining physical modelling and machine learningMaterials & Design, 197, 109199.

Go To Materials & Design

Check Also

A.I. Improves the Search for Next Generation Superhard Materials