Abstract

The martensite start temperature (Ms) is a critical parameter when designing high-performance steels and their heat treatments. It has, therefore, attracted significant interest over the years. Numerous methodologies, such as thermodynamics-based, linear regression and artificial neural network (ANN) modeling, have been applied. The application of data-driven approaches, such as ANN modeling, or the wider concept of machine learning (ML), have shown limited technical applicability, but considering that these methods have made significant progress lately and that materials data are becoming more accessible, a new attempt at data-driven predictions of the Ms is timely. We here investigate the usage of ML to predict the Ms of steels based on their chemical composition. A database of the Msvs alloy composition containing 2277 unique entries is collected. It is ensured that all alloys are fully austenitic at the given austenitization temperature by thermodynamic calculations. The ML modeling is performed using four different ensemble methods and ANN. Train-test split series are used to evaluate the five models, and it is found that all four ensemble methods outperform the ANN on the current dataset. The reason is that the ensemble methods perform better for the rather small dataset used in the present work. Thereafter, a validation dataset of 115 Ms entries is collected from a new reference and the final ML model is benchmarked vs a recent thermodynamics-based model from the literature. The ML model provides excellent predictions on the validation dataset with a root-mean-square error of 18, which is slightly better than the thermodynamics-based model. The results on the validation dataset indicate the technical usefulness of the ML model to predict the Ms in steels for design and optimization of alloys and heat treatments. Furthermore, the agility of the ML model indicates its advantage over thermodynamics-based models for Ms predictions in complex multicomponent steels.

Highlights

  • MATERIALS development is currently undergoing large changes with a transition from the previously dominating empirical development methodologies toward methodologies with more computational components

  • Key for both these areas is the use of databases where the integrated computational materials engineering (ICME) methods to a large extent rely on the so-called CALPHAD databases that collect thermodynamic and kinetic data essential for the modeling of phase transformations and related phenomena, while the machine learning (ML) approaches are more flexible to use any database that contains data of relevance for the parameter that should be predicted

  • It can be seen that the ensemble methods consistently perform better than the multilayer perceptron (MLP) method based on all the different quality metrics

Read more

Summary

INTRODUCTION

METALLURGICAL AND MATERIALS TRANSACTIONS A experimental input with physically based modeling on different length- and timescales, often referred to as integrated computational materials engineering (ICME).[1]. The data are taken from the open literature starting with the dataset made openly available by prior works of Capdevila and Andres,[25] Capdevila et al.,[26] and Garcia Matteo and co-workers,[27,28,29] whom developed ANN models for the prediction of the Ms Prior works, have not been able to predict the Ms for a large set of steel grades without significant scattering of the predictions, and to date, thermodynamics-based models with the mature commercial CALPHAD databases have been providing the most reliable predictions We challenge this in the present work

METHODOLOGY
ML Approach
Evaluation of Predictability for Statistical Modeling
Model Evaluation
Interactions in Data and Physical Interpretation
Benchmarking of ML Model with Thermodynamics-Based Models
CONCLUSIONS
79. Python Data Analysis Library–Pandas
Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call