Motivation

The representation of materials as data, i.e., a numerical feature vector, is not straight-forward and requires decisions on the captured properties (e.g., mechanical properties, composition and crystal structure) and their encryption (e.g., human-readable). Applied on large data sets a representation that uses little-to-no domain knowledge might perform as well as an informed feature selection. However, for scarce data, which are common in materials science, the featurization has significant impact on the prediction performance of machine learning (ML) models [1]. Similarly, the model design becomes more important with limited data and should consider aspects of overfitting and model interpretability.

Objectives

At BAM we aim to use the in-house expertise on material compositions, structures and properties to improve data driven materials science and engineering. There is no single method of building in domain knowledge in ML approaches. Our benefit is the close collaboration between material and data scientists, that enables the development of efficient and creative solutions, e.g., to

• derive domain specific feature selections, such as composition-based feature vectors, to improve the prediction performance of ML models
• develop human-readable feature formats that increase interpretability and may help to provide physical and chemical insights
• improve the model performance by capturing known correlations and analytical relations, e.g., via transfer learning
• complement time and cost intensive experimental data with simulated data, which are based on physical principles

References

[1] Murdock, R. J., Kauwe, S. K., Wang, A. Y.-T. & Sparks, T. D. “Is Domain Knowledge Necessary for Machine Learning Materials Properties?” Integrating Materials and Manufacturing Innovation 9, 221–227 (2020).