Enhanced Generalization through Atomic Embeddings and Cross-Element Parameterization

The limited transferability of machine-learned interatomic potentials (MLIPs) to unseen chemical compositions constrains their utility in computational materials discovery. We present a systematic investigation into two interdependent design principles─learned atomic representations and cross-element parameterization─that govern MLIP generalization. Using the diverse, multielement OC20 catalyst dataset, we evaluate chemically informed atomic embeddings against conventional 1-hot encoding within two distinct model architectures: the tensor-based LATTE descriptor and the MACE message-passing neural network. We demonstrate that embeddings such as Magpie and Matscholar reduce force mean absolute errors (MAEs) by 20-30% compared to 1-hot encoding, and improve data efficiency in low-data regimes by over 40%. Furthermore, by strategically sharing parameters across elements, we reduce the model sizes by up to 98% (from 23M to 400k parameters) without remarkably compromising accuracy, and even enable meaningful few-shot prediction for underrepresented elements. This work establishes a practical framework for building compact, transferable, and data-efficient MLIPs suitable for exploration across heterogeneous chemical spaces.