HARMAP: Hierarchical Atomic Representation for Materials Property Prediction

Haocheng Yuan, Qiangqiang Shen, Yudong Mao, Peilin CHEN, Shiqi Wang

Keywords: AI for Materials, Atomic Representation, Material Property Prediction

TL;DR: A Hierarchical Atomic Representation for Materials Property prediction.

Abstract: Accurate prediction of material properties is a key step toward rapid materials discovery and cost-effective exploration of vast chemical spaces. Recent advances in machine learning (ML) offer a data-driven alternative that enables fast and scalable property estimation. However, prevailing graph-based pipelines use one-hot or shallow element embeddings and simple distance-based edges, which under-encode element-specific characteristics and cannot faithfully capture bond relations. Thus, we develop HARMAP, a Hierarchical Atomic Representation for Materials Property prediction. First, we build a chemistry-informed Hierarchical Element Knowledge Tree (HEK-Tree) that classifies elements from coarse to fine (e.g., metal vs. non-metal, subgroupings), producing atomic embeddings that preserve unique identities and inter-atomic relations. Second, we map these features into hyperbolic spaces that preserve hierarchical structure, enabling compact separation of levels and smooth similarity across related elements. Finally, we construct a compound graph whose nodes use the learned atomic embeddings and whose edges combine geometric proximity, providing bond-aware connectivity. Across three large public datasets, HARMAP consistently improves over formula-only, structure-only, and standard graph baselines, indicating the effectiveness of HARMAP's unique atomic and bond representations.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 3745