Materials Science and Engineering A 434 (2006) 237–245 Neural network analysis of strain induced transformation behaviour of retained austenite in TRIP-aided steels Monideepa Mukherjee a,∗ , Shiv Brat Singh b , Omkar Nath Mohanty c a Tata Steel, Jamshedpur 831001, India b Department of Metallurgical and Materials Engineering, Indian Institute of Technology, Kharagpur 721302, India c Biju Patnaik University of Technology at Orissa, CET Campus, Ghatikia, Bhubaneswar 751 003, India Received 28 April 2006; received in revised form 26 June 2006; accepted 26 June 2006 Abstract Experimental data published in literature were used to develop a neural network model to predict the strain induced transformation behaviour of retained austenite as a function of 13 input variables including chemical composition of the steel, initial retained austenite content, matrix microstructure and forming conditions. The model was found to make reasonable predictions with respect to established metallurgical principles and other published data. © 2006 Elsevier B.V. All rights reserved. Keywords: TRIP; Neural networks; Retained austenite; Transformation; Martensite 1. Introduction During the past few years, the automobile industry has shifted its focus towards the design of light, environment friendly, fuel- efficient vehicles that has led to the development of a variety of high strength steels. These steels are essentially ‘multiphase steels’ (at least two different microstructural components) and are designated as advanced high strength steels (AHSS) that combine attractive strength and formability properties. Transformation induced plasticity (TRIP)-aided [1] steels belong to this AHSS genre and possess a complex microstructure typically consisting of soft polygonal ferrite as the matrix phase and bainite and metastable retained austenite as second phase. In recent years, two more varieties of TRIP-aided steels with bainitic ferrite and annealed martensite matrix have been devel- oped in response to the complex formability requirements of the present day auto-body manufacturers [2,3]. The most impor- tant phase, however, is retained austenite, which transforms to martensite on straining. This transformation is accompanied by a volume expansion and results in a localized increase of strain ∗ Corresponding author at: Department of Metallurgical and Materials Engi- neering, Indian Institute of Technology, Kharagpur 721302, India. Tel.: +91 3222 281796; fax: +91 3222 282280. E-mail addresses: monideepa13a@gmail.com (M. Mukherjee), sbs22@iitkgp.ac.in (S.B. Singh), omkarmohanty@yahoo.com (O.N. Mohanty). hardening coefficient which delays the onset of necking and ultimately leads to higher uniform and total elongation. This is called the transformation induced plasticity (TRIP) effect. The beneficial effect of this transformation depends to a great extent on the resistance of austenite to the same. This in turn depends on the driving force available for transformation and the acti- vation energy needed for transformation. The lack of activation energy is referred to as the stability of retained austenite and must be compensated by the mechanical driving force in order to induce transformation. If stability is too low or too high the beneficial effect of progressive transformation during forming cannot be utilized [4,5]. Therefore it is not only the amount of retained austenite but also its stability in the presence of strain, which is responsible for the good formability [6]. Stability of austenite depends on a number of factors viz. chemical compo- sition, annealing parameters, morphology and size of retained austenite and forming conditions like test temperature, strain, strain rate and stress state. A number of empirical as well as semi-empirical models have been used to describe the variation of the amount of retained austenite with strain [6–11]. Most of these models use constants, which can explain to a certain extent the effect of some of the above parameters that affect retained austenite stability. But to get a general idea about the effect of a large number of parameters simultaneously, an attempt was made to develop a neural network based model which could pre- 0921-5093/$ – see front matter © 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.msea.2006.06.076