Machine learning of material properties: Predictive and interpretable multilinear models

Problem

Machine learning models predict material properties accurately but lack transparency, limiting scientific insight.

Why it matters

Opaque models hinder trust and scientific understanding in materials design decisions.

What the AI did

Simple linear combinations of nonlinear basis functions were shown to match kernel and neural network accuracy while remaining directly interpretable via their coefficients.

Results

Comparable accuracy to kernel/neural network approaches for bandgap, formation energy, and spin state prediction (no numbers given)

Impact

Evidence supports interpretable linear models matching black-box accuracy on revisited public material datasets.

Record

Importance rating

More in Engineering & Materials