Leveraging multi-task model for improving mechanical property predictions of high entropy alloys (HEAs)

High Entropy Alloys (HEA), a special class of alloys containing 5 or more principal elements with molar fraction of each element between 5 and 35%, have been reported to show combinations of exceptional mechanical properties that are not attainable in conventional alloys. However, designing HEAs with target properties is a challenging task due to the large and mostly uncharted design space and the difficulty in predicting the non-linear relationships between structure, property, and processing parameters – especially well outside the temperature range of conventional alloys. Machine Learning (ML) surrogate models allow for faster, more reliable, and accurate predictions of material properties, but are contingent on the availability of abundant training data. 

 

However, data sparsity arises due to destructive mechanical tests and high manufacturing costs. This restricts researchers in performing multiple tests on a material system, resulting in some properties having significantly less data compared to others. Training ML models on such small datasets could lead to overfitting. However, the correlation between property, chemistry, and processing can potentially be leveraged using methods like multi-task learning (MTL). Therefore, we will compare a MTL model vs single task models on some HEA mechanical properties. Their performances will be assessed first on a synthetic dataset with varying amounts of missingness of these properties, and finally, on a real dataset.