Dirty engineering data-driven inverse prediction machine learning model

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.

A machine-learning-based alloy design platform that enables both forward and inverse predictions for thermo-mechanically controlled processed (TMCP) steel alloys

“…We incorporated 16 ML algorithms and provided full details for each algorithm involving a systematic hyper-parameter optimization and NSGA-II-driven inverse design (prediction) as well as visualization in the input feature space. Note that ANN, which was a major ML algorithm in our previous report 19 , is not mentioned in our account of the present investigation.…”

“…Herein, we report an integrated ML platform, which is a much-improved version of our previous ML strategy 19 . Most ML-based studies of metallic alloy design adopt, at best, a few ML algorithms.…”

“…Shen et al 5 have also used GAs successfully for inverse design. We used a deep neural networks (DNNs)-based inverse design 19 in association with an elitist-reinforced non-dominated sorting genetic algorithm (NSGA-II) 38 . Our NSGA-II-related inverse design approach has a lucid distinction from others based on the fact that we deal with more than one objective function, and two output features are simultaneously optimized with our approach.…”

“…As clearly described above, an NSGA-II was used to optimize the seven fully trained ML models. Using the NSGA-II algorithm for inverse prediction problems for the seven multi-output-feature ML models is an unprecedented approach, although conventional single-objective genetic algorithms in association with a particular ML algorithm have more often been used for the same purpose 5 , 19 , 35 – 37 , 42 – 44 . Since two objective functions (= two output features, YS and UTS) were of concern in the present investigation, both had to be simultaneously optimized, and the NSGA-II was the best option for that optimization process.…”

“…This industry-produced, real-world data was compromised by the effect of the so-called ‘dirty nature’ of data due to human intervention during the data production, which leads to data that is non-identically and independently distributed (Non-IID). In our previous report, the nature of the data was described well and included conventional data analysis based on both a pair-wise scatter plot and a Pearson correlation coefficient matrix 19 . For this study, we adopted an alternative, more comprehensive data visualization strategy, a so-called holographic research strategy (HRS) 45 – 48 , that is capable of a simpler visualization of multi-dimensional data in contrast to the well-known data dimension reduction and visualization methodologies such as the principal component analysis (PCA) that accompanies considerable amounts of information loss.…”