Detecting secondary-phase in bainite microstructure through deep-learning based single-shot approach

Vinod Kumar, Sharukh Hussain, Vishwas Subramanian, P G Kubendran Amos · 2025

Read Detecting secondary-phase in bainite microstructure through deep-learning based single-shot approach by Vinod Kumar, Sharukh Hussain, Vishwas Subramanian, P G Kubendran Amos (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-22757.txt.

Introduction

Vinod Kumar and co-authors present a 2025 deep learning method to find secondary phase islands in bainite steel. Bainite is a mixed microstructure whose small martensite austenite islands shape the strength and toughness of the steel. Spotting those islands by eye is slow and hard due to their tangled shapes.

The team recast object detection models for the task and trained a regression based detector at length. The detector marks the islands in one pass, with no need to resize the micrograph. It scales to automated scans of many samples. The paper prefaces the method with a study that shows plain classification tools fall short at linking process settings and composition to these mixed structures.

Catalog record

CREATORVinod Kumar, Sharukh Hussain, Vishwas Subramanian, P G Kubendran Amos
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-22757

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Detecting secondary-phase in bainite microstructure through deep-learning based single-shot approach Vinod Kumar, Sharukh Hussain, Vishwas Subramanian, P G Kubendran Amos Abstract Relating properties and processing conditions to multiphase microstructures begins with identifying the constituent phases. In bainite, distinguishing the secondary phases is an arduous task, owing to their intricate morphology. In this work, deep-learning techniques deployed as object-detection algorithms are extended to realise martensite-austenite (MA) islands in bainite microstructures, which noticeably affect their mechanical properties. Having explored different techniques, an extensively trained regression-based algorithm is developed to identify the MA islands. This approach effectively detects the secondary phases in a single-shot framework without altering the micrograph dimensions. The identified technique enables scalable, automated detection of secondary phase in bainitic steels. This extension of the detection algorithm is suitably prefaced by an analysis exposing the inadequacy of conventional classification approaches in relating the processing conditions and composition to the bainite microstructures with secondary phases. keywords Bainite microstructure, secondary phase, MA island, deep learning, object detection 1 Introduction Despite being hard to define and comprehend aaronson1990bainite ; aaronson2002progress , bainite steels continue to be of huge interest to engineers adamczyk2021studies ; ramachandran2021role . This interest is primarily driven by the wide range of properties they offer through their secondary phases tomita1983improvement ; edmonds1990structure . The secondary phases, while yielding the desired set of properties, significantly increase the complexity of the bainite microstructure zajac2005characterisation . Achieving the desired properties demands meticulous control over the type and volume fraction of the secondary phases in bainite steels bleck2011improved . A control over the nature of these secondary phases can only be gained through a proper understanding of the effect of composition and/or processing in bainite. Generally, the influence of composition and/or processing on a microstructure is largely evident. Consequently, a required phase-fraction is achieved by exposing an alloy of definite composition to a suitable processing treatment kumar2018toughness . However, owing to the intricacy of the bainite microstructure, understanding the effect of composition and processing treatments becomes challenging. Therefore, human experts are invariably involved to realise the features of the bainite microstructure, particularly the type and configuration of the secondary phases zajac2005characterisation . But recently, a paradigm shift is noticeable in the analysis of the microstructure with respect to its composition and processing conditions durmaz2021deep ; azimi2018advanced . Lately, deep learning has increasingly been employed to realise the change induced in the intricate bainite microstructure by the processing history and/or chemical make-up muller2020classification ; muller2021microstructural ; ackermann2022automated ; kumar2025interpretable . A wide range of computer vision, machine learning, and deep learning approaches have been adopted to distinguish various steel microstructures, including bainite, pearlite, martensite, and ferrite. Realising the distinction among these steels is rather straightforward, owing to the characteristic differences in the morphology of the phases. Correspondingly, after adequate computer vision-based pre-processing, machine learning techniques like support vector machines are employed to classify the different microstructures of steels, even distinguishing between upper and lower bainite muller2020classification ; muller2021microstructural . However, when the intent is to relate the processing condition and/or composition to the microstructure, these classifications offer only a superficial understanding. In other words, the processing for a given steel rarely brings about a significant change in the type of the constituent phases; instead, it alters the phase-fraction and the morphology of the secondary phase(s). This means the adopted technique should be capable of distinguishing microstructures with the same combination of phases, primarily based on their intricate morphology and distribution. The inadequacy of conventional machine learning techniques in classifying such microstructures with similar phases necessitates further analysis, such as detecting the critical phase(s) that dictate the properties and are influenced by the chemical composition and/or processing conditions kumar2024inadequacy . In bainite microstructures resulting from non-isothermal transformation, a secondary phase referred to as martensite-austenite (MA) islands is introduced, which establishes significant changes in the mechanical properties depending on their morphology and distribution. Therefore, any attempts to understand the effect of composition and/or processing conditions on bainite microstructures containing MA islands demand realising this characteristic secondary phase xu2021significant ; jiang2021tempering ; lu2023structure . This necessity arises from the assumption that conventional machine-learning-based classification would be insufficient. To that end, recently, a masked region-based convolutional neural network (RCNN) within the framework of Detectron2 has been employed to segment MA islands ackermann2022automated . This work, while pioneering the effort in realising the MA islands in intricate bainite microstructures, still offers scope for further refinement. Firstly, the approach involves a combination of classification and detection for realising the MA islands. Secondly, the accuracy of the detection is dependent on the size of the micrographs, with the models offering greater precision when the images are smaller than the original. The former component of the study indicates the computational demand, whereas the latter limits the general usability of the approach. Given these apparent limitations, in this work, a single-shot tool is developed to efficiently detect the MA islands in the bainite microstructure. Before expose various techniques to identify an efficient approach for realising the secondary phase, the bainite microstructures are analysed through machine-learning tools to expose the inadequacy of the conventional treatment and to establish the need for detecting the critical MA islands. 2 Overall Workflow Figure 1: The overall work flow of the present investigation including processing of the original bainite microstructure, feature extraction, classification using range of machine-learning techniques and development of secondary-phase detection models. 2.1 Bainite Dataset The present model to detect MA islands in bainite microstructures is developed by training on the known dataset published in ref iren2021aachen . This dataset includes micrographs of nine different steels distinguished by their composition, and exposed to two different heat treatment cycles. Of the nine, only five steels are considered in this work (CFB2, Type-1, -5, -10, and -11), given the number of microstructures associated with each type. This consideration yields a set of 1476 bainite micrographs with 8539 MA islands in them. All steel types, before being hot rolled at 900 0 C, are austenised at 1200 0 C for 10 minutes. During hot rolling, a strain of 0.3 is applied at a rate of 10s -1 . The difference in the processing is introduced by cooling some samples, across compositions, at the rate of 5Ks -1 and others at 0.3Ks -1 to the quench-stop temperatures of 400 0 C and 500 0 C, respectively. These disparities in the cooling rate and quench-stop temperature are treated as the distinguishing processing conditions in the present investigation. High quality images of the bainite microstructures from different steels are obtained through Scanning Electron Microscopy (SEM) at 4000 \times magnification. A duplicate of the 1476 bainite micrographs, called the overlaid dataset, is created by superimposing the polygons that highlight MA islands. The coordinates of the polygons are assigned by human experts after careful consideration. Therefore, besides the raw dataset of 1476 bainite micrographs, there exists a duplicate overlaid dataset highlighting the MA islands. 2.2 Deep-learning approach The entirety of the current work is graphically presented in Fig. 1 . This workflow can be broadly categorised into two parts. In the preliminary analysis, the bainite microstructures containing MA islands are classified through a wide range of machine-learning models to explicate the inadequacy of classification in realising

Read the complete text