Accurate fold models are of great significance for mineralization control, resource exploration, and underground engineering. However, existing automated modeling methods show difficulty in quantitatively describing fold development patterns and lack the available reference models required for multiple-point statistics and intelligent modeling techniques. This study proposes a novel three-dimensional (3D) fold-modeling method that integrates multiple-point-statistics-based pattern library construction with a long short-term memory (LSTM) network-based modeling framework. The multiple-point geostatistic is employed to quantify spatial distributions and correlations in geological data, thereby identifying the intrinsic structural patterns of folds. The extracted patterns are transformed into a training library that effectively represents the geological semantics and morphological diversity of folds, providing a reliable dataset for LSTM-based model training. An optimized ConvLSTM network is designed to ensure robust representation of fold complexity and variability. Based on the network, 3D models can be rapidly generated from geological profiles. Multiple experiments demonstrate that the proposed method can automatically produce 3D models that conform to realistic geological conditions and accurately reflect true fold geometries. The approach significantly improves modeling efficiency and geological feature representation, providing a reliable tool for geological engineering applications.

Apatite rare-earth-element (REE) compositions can discriminate igneous host-rock types, but systematic comparisons of deep-learning models and alternative REE representations remain limited, while analyses from shared publications or geological bodies complicate estimates of generalization. We evaluated four-class classification in a cohort of 4148 apatite analyses from 76 connected source groups. Six primary models (SVM, RF, XGBoost, MLP, ResNet, and FT-Transformer) were compared across seven representations derived from the same 14 REEs. Logistic regression and linear discriminant analysis served as baselines, and ILR features were examined separately. Source-group-aware nested cross-validation used five outer and three inner folds, with candidate selection and tuning confined to outer training groups. The selected procedure achieved a five-seed pooled out-of-fold Macro-F1 of 0.556 (95% source-group bootstrap CI, 0.398–0.652). With Raw inputs, logistic regression achieved 0.573 (0.436–0.672). Confidence intervals overlap, and paired comparisons after multiplicity adjustment do not support a reliable ranking of the candidates or a consistent advantage for deeper architectures or combined representations. Raw, Geo, CLR, and ILR are alternative representations of the same REE measurements rather than independent chemical information. Logistic regression using Raw inputs remained competitive, while transfer to unseen geological sources remains uncertain, especially for classes represented by few independent sources.