Download all files (3.66MB)

コレクション

引用

説明:

Recent advances in phosphor informatics have achieved high predictive accuracy with complex machine learning models. However, this often comes at the cost of physical interpretability, creating an interpretability dilemma. To address this challenge, we propose an approach that prioritizes model interpretability. We constructed a simple linear model (ridge regression) for a curated dataset of 118 Eu2+-activated phosphors, using only the chemical composition (atomic fractions) as features. This approach enabled us to successfully quantify the contribution of each constituent element to the peak emission wavelength as a physically interpretable "elemental contribution coefficient" (ECC). The trends derived from the ECCs agree remarkably with fundamental chemical intuitions in phosphor chemistry and with established empirical physical rules, thereby demonstrating the scientific validity of our model. Furthermore, analysis of the systematic discrepancies revealed that a principal limitation of the model is its inability to decouple the competing physical effects of the centroid shift and crystal field splitting. This study demonstrates a pathway to elevate machine learning from a mere predictor to an analytical tool. Such a tool can interpret underlying scientific relationships in data and deepen our understanding of science. This approach bridges the gap between data-driven science and materials science.

権利情報:

• Creative Commons BY Attribution 4.0 International
  

キーワード: materials informatics, phosphor, europium, machine learning, interpretability, linear model, composition-based feature

刊行年月日: 2026-12-31

出版者: Informa UK Limited

掲載誌:

• Science and Technology of Advanced Materials: Methods (ISSN: 27660400) vol. 6 issue. 1 2691688

研究助成金:

原稿種別: 出版者版 (Version of record)

MDR DOI:

公開URL: https://doi.org/10.1080/27660400.2026.2691688

関連資料:

その他の識別子:

連絡先:

更新時刻: 2026-07-07 14:31:04 +0900

MDRでの公開時刻: 2026-07-07 18:24:09 +0900