A PCA-GMM-based adaptive probabilistic boundary model for chemical-group classification of ordinary chondrites
Xing Liu, Zhipeng Xia, Bingkui Miao, Cheng Xu
Icarus, published: 21 October 2026
“Highlights
- PCA-GMM resolves oblique Fa-Fs trends in ordinary chondrites.
- Adaptive probabilistic boundaries separate H, L and LL groups.
- Boundary topology reveals an H/L hiatus and L/LL overlap.”
“The chemical-group classification of ordinary chondrites is primarily based on the fayalite content of olivine (Fa) and the ferrosilite content of low-Ca pyroxene (Fs). However, the strong covariation between these two parameters makes conventional classification boundaries based on univariate empirical ranges insufficient for describing the oblique distribution of ordinary chondrites in Fa-Fs space. In this study, we compiled 17,533 curated Fa-Fs records of ordinary chondrites from the Meteoritical Bulletin Database and developed a PCA-GMM-based Adaptive Probabilistic Boundary (APB) model. First, principal component analysis (PCA) was used to extract PC1 as the principal covariation axis of the Fa-Fs data. A one-dimensional Gaussian mixture model (GMM) was then applied along PC1 to identify the latent core distributions of the H, L, and LL groups. The core intervals were defined adaptively according to the distributional characteristics of each group: a 2.807σ criterion was applied to the near-Gaussian H and L groups, whereas a 0.25%–99.75% percentile range was used for the positively skewed LL group. By back-projecting the PC1 thresholds into the original Fa-Fs space, six statistical boundaries were obtained, and an empirical silicate FeO-enrichment index, IFeO, was defined. The results show that the APB boundaries conform well to the oblique covariation pattern of the Fa-Fs data, and that the boundary parameters exhibit good stability under repeated subsampling. The APB reclassification results are 97.98% consistent with the original database labels, while also allowing boundary samples and potential outliers to be identified. Further analysis reveals two distinct boundary topologies: the H/L boundary is characterized by a clear low-density hiatus, whereas the L/LL boundary shows local overlap and a more continuous transition. A comparison with published oxygen-isotope data shows a similar relative pattern, with clearer H/L separation than L/LL separation. The proposed APB model and IFeO-based classification workflow provide an objective and reproducible auxiliary tool for chemical-group discrimination, boundary-sample identification, and database quality control of equilibrated ordinary chondrites. Because the final results are expressed as explicit Fa-Fs boundary equations and an empirical IFeO index, the workflow can also be applied to routine mean Fa-Fs data without requiring users to implement the full computational procedure.”
