Exploration of Characteristic Temperature Contributions to Metallic Glass Forming Ability

TL;DR

This study uses machine learning to evaluate if characteristic temperatures can predict the glass forming ability of metal alloys, finding limited success with regression models but some promise in categorization tasks.

Contribution

It systematically assesses the predictive power of characteristic temperatures for glass forming ability using a large database and machine learning, revealing their limited effectiveness for regression.

Findings

Regression models perform poorly with high RMSE.

Categorization models achieve a mean F1 score of 0.77.

Larger databases may improve predictive accuracy.

Abstract

Various combinations of characteristic temperatures, such as the glass transition temperature, liquidus temperature, and crystallization temperature, have been proposed as predictions of the glass forming ability of metal alloys. We have used statistical approaches from machine learning to systematically explore a wide range of possible characteristic temperature functions for predicting glass forming ability in the form of critical casting diameter, $D_{max}$. Both linear and non-linear models were used to learn on the largest database of $D_{max}$ values to date consisting of 747 compositions. We find that no combination of temperatures for features offers a better prediction of $D_{max}$ in a machine learning model than the temperatures themselves, and that regression models suffer from poor performance on standard machine learning metrics like root mean square error (minimum value of $3.3 \pm 0.1$ $mm$ for data with a standard deviation of 4.8 $mm$). Examination of the errors vs. database size suggest that a larger database may improve results, although a database significantly larger than that used here would likely be required. Shifting a focus from regression to categorization models learning from characteristic temperatures can be used to weakly distinguish glasses likely to be above vs. below our database's median $D_{max}$ value of 4.0 $mm$, with a mean F1 score of $0.77 \pm 0.02$ for this categorization. The overall weak results on predicting $D_{max}$ suggests that critical cooling rate might be a better target for machine learning model prediction.