Alchemical learning capability within the PET architecture

Ascertain whether the Point Edge Transformer (PET) architecture can be adapted to perform alchemical learning in a straightforward manner, enabling interpolation or embedding for unseen chemical elements to extrapolate in high-entropy alloy datasets, and, if feasible, develop a concrete procedure to implement such alchemical compression within PET.

Background

The authors demonstrate alchemical learning in CACE by using low-dimensional element embeddings and linear interpolation to extrapolate to unseen elements (Re and Os) in the HEA25 dataset, achieving transferability comparable to test-set performance.

In contrast, while PET shows state-of-the-art test errors, the authors raise uncertainty about PET’s ability to support alchemical learning straightforwardly, highlighting a gap in adapting PET for element interpolation or compression across chemical space.