Investigating representation schemes for surrogate modeling of High Entropy Alloys

TL;DR

This paper systematically compares different representation schemes for surrogate modeling of High Entropy Alloys, finding that chemically meaningful spatial arrangements improve deep learning models, while tree-based models using atomic fractions excel in transfer learning.

Contribution

It provides a comprehensive comparison of composition representation schemes and highlights the effectiveness of spatial arrangements and tree-based models in alloy surrogate modeling.

Findings

Chemically meaningful spatial schemes improve deep learning performance.

Tree-based models with atomic fractions outperform in transfer learning.

Structured representations generally lead to better surrogate models.

Abstract

The design of new High Entropy Alloys that can achieve exceptional mechanical properties is presently of great interest to the materials science community. However, due to the difficulty of designing these alloys using traditional methods, machine learning has recently emerged as an essential tool. Particularly, the screening of candidate alloy compositions using surrogate models has become a mainstay of materials design in recent years. Many of these models use the atomic fractions of the alloying elements as inputs. However, there are many possible representation schemes for encoding alloy compositions, including both unstructured and structured variants. As the input features play a critical role in determining surrogate model performance, we have systematically compared these representation schemes on the basis of their performance in single-task deep learning models and in transfer learning scenarios. The results from these tests indicate that compared to the unstructured and randomly ordered schemes, chemically meaningful arrangements of elements within spatial representation schemes generally lead to better models. However, we also observed that tree-based models using only the atomic fractions as input were able to outperform these models in transfer learning.