Fig. 8: Overall CrabNet architecture and prediction of material property and uncertainty.
From: Compositionally restricted attention-based network for materials property predictions
a Schematic of the CrabNet architecture including the input EDM, the self-attention layers (repeated N times), the updated and final element representations (EDM\(^{\prime}\) and EDM″), the residual network, and the final model output. b The calculation steps for element contributions and prediction of the targets and uncertainties. The \(p^{\prime}\) and \(u^{\prime}\) vectors represent the element-proto-contributions and the element uncertainties, respectively. \(y^{\prime}\) represent the element contributions. The material property is obtained by taking the mean of element-contributions (\(y^{\prime}\)) for each compound. Similarly, the mean of the element-uncertainties (\(u^{\prime}\)) gives us the estimated aleatoric uncertainty.
