Overview
In modern cities, tackling complex environmental and energy issues through experiments can be impractical due to cost, risk, or scale. Extreme-scale computing now enables the creation of highly accurate predictive models and allows for the analysis of vast datasets, thanks to advancements in AI. Integrating predictive modelling with machine learning opens new opportunities in science, shifting the focus from traditional, human-centered approaches to hybrid models where human and artificial intelligence collaborate on design, discovery, and evaluation.
The Exa-MA project emphasizes scalable numerical methods on current and future hardware. This cross-disciplinary initiative combines modeling, data analysis, and AI, addressing challenges in various sectors like energy, health, and the environment. Exa-MA is part of the French Exascale program NumPEx, which aims to prepare for exascale computing with advanced software and training, ensuring France’s leadership in high-performance computing (HPC).
Collaboration and Goals
Cemosis (Strasbourg University) is a key player in Exa-MA, collaborating with partners like CEA, Inria, Ecole Polytechnique, and Sorbonne Université. The partnership also connects NumPEx with the CoE HiDALGO2, which focuses on modelling, data simulation, and visualization. Together, they work toward solutions that leverage pre-exascale systems to solve real-world challenges.
As part of this initiative, Cemosis is leading a project to develop the Urban Building Model (UBM). This pilot, part of the HiDALGO2 project, focuses on enhancing buildings’ energy efficiency and indoor air quality by incorporating vegetation into energy simulations.
Methodology: Generating 3D Vegetation Models
Urban vegetation, particularly trees, plays a crucial role in energy and thermal simulations for cities. The methodology for generating 3D tree models involves:
- Data Acquisition: Tree metadata, such as GPS coordinates, height, trunk
circumference, and species, is extracted from OpenStreetMap using the cpr
framework. - Tree Modeling: Trees are classified by shape (cone, oval, round). Basic models are created with Gmsh for lower Levels of Detail (LOD 0), while higher LODs use pre- processed tree meshes sourced from the SketchUp 3D Warehouse.
- Tree Scaling: The models are scaled based on data from OpenStreetMap, adjusted
for height, trunk circumference, and crown diameter. - CGAL 3D Alpha Wrapping: This algorithm generates runtime tree meshes, optimized
for LOD 1, 2, and 3 by adjusting the α values. - Tree Placement: GPS coordinates are converted to Cartesian coordinates, and the
trees are placed accordingly in the simulation. - Mesh Merging and Optimization: Tree meshes are merged with other elements
(buildings, terrain) for optimized rendering. - Parallelization: Multi-threading is used to parallelize tree placement, enhancing the
performance of large dataset management.
Results and Discussion
Our approach was tested in the Strasbourg city centre, where trees were accurately placed and scaled within the urban terrain mesh. The foliage features seasonal changes, with different colours corresponding to varying leaf densities.
We also conducted a performance analysis, assessing the impact of auto-refinement on processing speed. The results highlight the balance between accuracy and computational performance, demonstrating that auto-refinement plays a significant role in optimizing the modeling process for large-scale urban simulations.
We carried out a performance analysis to assess the efficiency of our implementation. The following figure displays the execution time under various conditions, specifically comparing the impact of enabling versus disabling the auto-refinement feature. The data reveals how auto-refinement influences processing speed, offering insights into the trade- offs between accuracy and performance.
Author: UNISTRA – Cemosis
Editor: Kyriaki Daskaloudi
