Dirty engineering data-driven inverse prediction machine learning model

Jin-Woong Lee,Seonghwan Kim,Nam Hoon Goo,Byung Do Lee,Woon Bae Park,Kee-Sun Sohn

doi:10.1038/s41598-020-77575-0

Jin-Woong Lee, Seonghwan Kim + Show 4 more

Open Access

https://doi.org/10.1038/s41598-020-77575-0

Copy DOI

Abstract

Most data-driven machine learning (ML) approaches established in metallurgy research fields are focused on a build-up of reliable quantitative models that predict a material property from a given set of material conditions. In general, the input feature dimension (the number of material condition variables) is much higher than the output feature dimension (the number of material properties of concern). Rather than such a forward-prediction ML model, it is necessary to develop so-called inverse-design modeling, wherein required material conditions could be deduced from a set of desired material properties. Here we report a novel inverse design strategy that employs two independent approaches: a metaheuristics-assisted inverse reading of conventional forward ML models and an atypical inverse ML model based on a modified variational autoencoder. These two unprecedented approaches were successful and led to overlapped results, from which we pinpointed several novel thermo-mechanically controlled processed (TMCP) steel alloy candidates that were validated by a rule-based thermodynamic calculation tool (Thermo-Calc.). We also suggested a practical protocol to elucidate how to treat engineering data collected from industry, which is not prepared as independent and identically distributed (IID) random data.

Highlights

Most data-driven machine learning (ML) approaches established in metallurgy research fields are focused on a build-up of reliable quantitative models that predict a material property from a given set of material conditions
Over the past decade we have successfully discovered many inorganic functional materials using metaheuristics strategies[17,18,19], but this approach has never been used for metal alloy design
The present investigation deals with the most common quantitative material conditions and performance relationships (QCPR) models with a fully labeled dataset, such as deep neural network (DNN)[21], k-nearest neighbors (KNN)[22], random forest (RF)[23], support vector machine (SVM)[24], and Gaussian process regression (GPR)[25], because these are considered a prerequisite for the major focus of the present investigation—inverse design

Summary

Introduction

Most data-driven machine learning (ML) approaches established in metallurgy research fields are focused on a build-up of reliable quantitative models that predict a material property from a given set of material conditions. One approach involves a metaheuristics-assisted inverse prediction using a conventional forward DNN model, and the other approach depends on plausible input data generation via a so-called modified variational autoencoder (MVAE).

Results

Conclusion