# Investigating representation schemes for surrogate modeling of High Entropy Alloys

**Authors:** Arindam Debnath, Wesley F Reinhart  
**arXiv:** [2301.00179](https://arxiv.org/abs/2301.00179)  
**Date:** 2022-12-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.00179v1/

## Key points

- High Entropy Alloys (HEAs) in materials science have exceptional mechanical properties and are a focus of research.
- Traditional methods for designing HEAs are challenging, leading researchers to use machine learning, particularly surrogate models.
- Surrogate models involve screening candidate alloy compositions and the choice of representation scheme is crucial.
- Representation schemes can be unstructured or structured, with spatial representation schemes showing superior performance compared to randomly ordered ones.
- Tree-based models using only atomic fractions as input have shown to outperform other models in transfer learning tasks.
- A study by Arindam Debnath and Wesley F Reinhart highlights the importance of representation schemes in surrogate modeling for HEAs.

## Summary

In the field of materials science, the design of High Entropy Alloys (HEAs) with exceptional mechanical properties has garnered significant attention. Traditional methods for designing these alloys have proven to be challenging, leading researchers to turn to machine learning as a valuable tool in this endeavor. Surrogate models, which involve screening candidate alloy compositions, have become integral to the materials design process in recent years. One key aspect of surrogate modeling is the representation scheme used to encode alloy compositions. This includes both unstructured and structured variants, with many models relying on atomic fractions of alloying elements as inputs. The choice of input features is crucial in determining the performance of surrogate models. To shed light on this topic, a systematic comparison was conducted on various representation schemes based on their efficacy in single-task deep learning models and transfer learning scenarios. The findings from these tests revealed that spatial representation schemes incorporating chemically meaningful arrangements of elements generally yielded superior model performance compared to unstructured and randomly ordered schemes. However, an interesting observation was made regarding tree-based models utilizing only atomic fractions as input – they were able to outperform other models in transfer learning tasks. This study, led by authors Arindam Debnath and Wesley F Reinhart, provides valuable insights into the importance of representation schemes in surrogate modeling for HEAs. By understanding how different encoding methods impact model performance, researchers can optimize their approach to designing new high entropy alloys with exceptional mechanical properties.