Is Domain Knowledge Necessary for Machine Learning Materials Properties?

Abstract

<div>New methods for describing materials as vectors in order to predict their properties using machine learning are common in the field of material informatics. However, little is known about the comparative efficacy of these methods. This work sets out to make clear which featurization methods should be used across various circumstances. Our findings include, surprisingly, that simple one-hot encoding of elements can be as effective as traditional and new descriptors when using large amounts of data. However, in the absence of large datasets or data that is not fully representative we show that domain knowledge offers advantages in predictive ability.</div><div><br></div>

Is Domain Knowledge Necessary for Machine Learning Materials Properties?

“…Band gap, formation energy, shear modulus, bulk modulus, Debye temperature, thermal expansion, and thermal conductivity data were then collected from the ICSD catalogue of the AFLOW database [16]. Duplicate entries were removed and each material property's formulae and ground-truth values were randomly partitioned into training, validation, and test sets (the full code is available in the GitHub repository [22]. Note, for this work, the associated Crystal Information Files (CIF) were discarded.…”

“…The model was then tested on these withheld formulae. The code for these methods is also available on GitHub [22].…”

Invited review: Machine learning for materials developments in metals additive manufacturing

“…The preprocessing step of featurizing data is crucial for successful implementation of machine learning algorithms. Improper featurization of data can impact prediction and classification errors [30].…”

Machine Learning for Materials Scientists: An Introductory Guide Towards Best Practices

“…In general, a good ML project should do one or more of the following: screen or downselect candidate materials from a pool of known compounds for a given application or property, [1][2][3] acquire and process data to gain new insights, 4,5 conceptualize new modeling ap-proaches, [6][7][8][9][10] or explore ML in materials-specific applications. 1,[11][12][13] Consider these points when you judge the applicability of ML for your project.…”

“…For sufficiently large datasets and for more "capable" learning architectures like very deep, fully-connected networks 7,122 or novel attention-based architectures such as CrabNet, 6 feature engineering and the integration of domain knowledge (such as through the use of CBFVs) in the input data becomes irrelevant and does not contribute to a better model performance compared to a simple onehot-encoding. 11 Therefore, due to the effort required to curate and evaluate domain knowledge-informed features specific to your research, you may find it more beneficial to seek out additional sources of data, already-established featurization schemes, or use learning methods that don't require domain-derived features 6 instead.…”