Automatic segmentation of geological structures such as vugs and fractures is an important task for reservoir characterization. Previous research has mainly focused on singular deep neural network architectures for detecting fractures using limited datasets. In this work, we evaluate the performance of different deep neural network backbones (ResNet, VGG16, InceptionV3, DenseNet, and TransUNet), combined through an ensemble strategy. The models were tested on real-world borehole images containing typical noise and artifacts, thus reflecting more realistic scenarios. The results demonstrate that our hybrid architecture improves IoU and F1-score by up to 6% when compared to base architectures commonly used in the literature. The proposed ensemble achieves an average IoU of 64% and an F1-score of 78% across 9 different wells. A 4-fold cross-validation ensures a robust performance analysis. The findings of this study demonstrate how the strengths of convolutions and transformers can be combined to enhance robustness in complex real-world geological scenarios and pave the way for future research on these approaches.

Building Information Modeling (BIM) employs data-rich 3D CAD models for large-scale facility design, construction, and operation. These complex datasets contain a large amount and variety of information, ranging from design specifications to real-time sensor data. They are used by architects and engineers for various analysis and simulations throughout a facility's life cycle. Many techniques from different visualization fields could be used to analyze these data. However, the BIM domain still remains largely unexplored by the visualization community. The goal of this article is to encourage visualization researchers to increase their involvement with BIM. To this end, we present the results of a systematic review of visualization in current BIM practice. We use a novel taxonomy to identify main application areas and analyze commonly employed techniques. From this domain characterization, we highlight future research opportunities brought forth by the unique features of BIM. For instance, exploring the synergies between scientific and information visualization to integrate spatial and non-spatial data. We hope this article raises awareness to interesting new challenges the BIM domain brings to the visualization community.

Building Information Modeling (BIM) provides an integrated 3D environment to manage large-scale engineering projects. The Architecture, Engineering and Construction (AEC) industry explores 4D visualizations over these datasets for virtual construction planning. However, existing solutions lack adequate visual mechanisms to inspect the underlying schedule and make inconsistencies readily apparent. The goal of this paper is to apply best practices of information visualization to improve 4D analysis of construction plans. We first present a review of previous work that identifies common use cases and limitations. We then consulted with AEC professionals to specify the main design requirements for such applications. These guided the development of CasCADe, a novel 4D visualization system where task sequencing and spatio-temporal simultaneity are immediately apparent. This unique framework enables the combination of diverse analytical features to create an information-rich analysis environment. We also describe how engineering collaborators used CasCADe to review the real-world construction plans of an Oil & Gas process plant. The system made evident schedule uncertainties, identified work-space conflicts and helped analyze other constructability issues. The results and contributions of this paper suggest new avenues for future research in information visualization for the AEC industry.