=

The passage around this formula

A structure-property model asserts a functional dependence $P = f(\mathbf{s}; \theta)$. where P is the target property, $\mathbf{s}$ is a structural descriptor vector (lattice parameters, coordination, bonding topology), and $\theta$ are parameters fit to a training set \{($\mathbf{s}_i$, $P_i$)\} . The equation is trivial; the content is entirely in what $\mathbf{s}$ includes and what range of $\mathbf{s}_i$ the training set spans. A model fit only on cubic perovskites extrapolates poorly to a layered structure even if the underlying physics is continuous, because f was never asked to fit that region.

Read this part in the article →