The development of scalable models for automated BIM element classification is hindered by the reliance on supervised learning, which requires expensive and laborious manual data annotation. BIM-JEPA introduces a pre-trained model that leverages a Joint Embedding Predictive Architecture for self-supervised learning on unlabeled 3D point cloud representations of individual BIM elements. Unlike conventional approaches, which train separate and task-specific models from scratch, a single pre-trained model can be efficiently fine-tuned for a wide range of downstream tasks.

BIM-JEPA learns by predicting the embeddings of masked portions of a BIM element from a given context, entirely within a shared latent space. From an ordered sequence of geometric patches, the model samples one larger context block and four smaller target blocks. The context block represents a partial view of the object (i.e., 40% to 75% of the patch embeddings), while the target blocks represent different masked regions (i.e., 15% to 20% of the patch embeddings) that the model must predict. The indices of patch embeddings chosen for the context differ from those for the targets to avoid trivial solutions. Predicting in the latent space rather than reconstructing raw points lets the model focus on high-level geometry and avoids the need for heavy data augmentation.

BIM-JEPA is pre-trained on 907,349 samples of individual BIM elements, drawn from three large-scale sources of real-world IFC/BIM geometry. Although these are labeled datasets, the labels were not used for the pre-training task. For the downstream classification task, the model is evaluated on two datasets: IFCNetCore, a curated subset of 20 balanced classes that provides a standard for general classification accuracy, and BIMGEOM, a more imbalanced set of 13 classes that assesses the robustness and generalization capabilities of the model.

BIM-JEPA establishes the best performance across all metrics on IFCNetCore. The model achieves an overall accuracy of 89.37% and a mean class accuracy of 86.63%, surpassing all baseline methods with minimal data augmentation during the fine-tuning phase (i.e., only scaling), thus relying primarily on the rich features learned during pre-training.

On the more imbalanced BIMGEOM dataset, BIM-JEPA attains 92.43% overall accuracy and 89.53% mean class accuracy, exceeding SpaRSE-BIM (i.e., 90.47% and 87.66% respectively). On weighted average metrics, BIM-JEPA surpasses SpaRSE-BIM across precision, recall, and F1, while its stronger macro average performance indicates more even performance across under-represented BIM elements.

To provide qualitative insight into the structure of the learned feature spaces, the test-set embeddings are projected onto a two-dimensional plane using Principal Component Analysis (PCA), with the background regions representing the linear decision boundaries learned by an SVM. The pre-trained encoder yields features that appear as a single, undifferentiated cluster, typical of a self-supervised model before it has been adapted to a downstream task. Upon fine-tuning, this structure becomes visually apparent: the features coalesce into distinct class-specific groups, culminating in the fine-tuned prelogits which form tighter, denser, and separated clusters compared to other models.

The data efficiency curves demonstrate that the model is highly sample-efficient, capturing a large fraction of its final performance with only a small subset of the labeled data. Notably, with just 25% of the training set, BIM-JEPA reaches a mean overall accuracy of approximately 75% on IFCNetCore and 82% on BIMGEOM. This rapid convergence strongly suggests that the self-supervised pre-training effectively equips the model with robust and generalizable representations, thereby reducing its dependency on extensive labeled datasets for the downstream task.

The N-shot learning experiments further evaluate the ability of BIM-JEPA to generalize from a very small number of labeled training examples per class. For BIMGEOM, the mean overall accuracy steadily climbs from approximately 50% with 5 shots to over 80% with 140 shots. A similar, albeit more compressed, trend is observed on IFCNetCore, where accuracy increases from around 40% to nearly 70% as the number of shots grows from 5 to 36.