Machine learning of material properties: Predictive and interpretable multilinear models.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning models offer rapid material property predictions but frequently lack transparency.
• Interpretability techniques can be applied to black-box models or by creating inherently interpretable models.

Purpose of the Study:

• To demonstrate that simple linear models can achieve accuracy comparable to complex methods.
• To highlight the benefits of directly interpretable models for gaining new scientific insights.

Main Methods:

• Revisiting existing material datasets.
• Developing simple linear combinations of nonlinear basis functions.
• Comparing accuracy against kernel and neural network approaches.

Main Results:

• Linear models accurately predict bandgaps and formation energies for transparent conducting oxides.
• Linear models accurately predict spin states for transition metal complexes.
• Linear models accurately predict formation energies for elpasolite structures.

Conclusions:

• Intrinsically interpretable linear models provide accurate predictions and enhanced transparency.
• Understanding model coefficients and functional forms reveals new scientific insights.
• Recognizing when intrinsically interpretable solutions are optimal is crucial for effective materials discovery.