Executing realistic earthquake simulations in unreal engine with material calibration

Yitong Sun, Hanchun Wang, Zhejun Zhang, Cyriel Diels, Ali AsadipourComput. Graph 2024 · OpenReview

TL;DR

This paper introduces RESenv, a novel platform built on Unreal Engine 5 that enables high-fidelity, real-time earthquake simulations for urban environments. By calibrating material properties using Ansys and integrating real earthquake data, RESenv significantly enhances the realism of building destruction, providing a crucial tool for training AI and robotics in disaster response scenarios.

What Problem Does This Paper Address?

The increasing integration of AI and robotics in disaster response, particularly for earthquake scenarios, necessitates realistic and high-fidelity simulation environments for effective training. Traditional simulation methods often struggle to provide the real-time performance and visual fidelity required for AI models, especially those relying on computer vision. Existing game engines, while offering real-time capabilities, lack the precision for accurate material simulation and structural damage modeling needed for scientific applications. This gap leaves a need for a platform that bridges the gap between high-fidelity simulation and real-time performance, enabling comprehensive training for AI and robotic systems in complex urban disaster scenarios.

Methods

RESenv employs a four-stage process:

  • Material Calibration: This is a core innovation. The authors use Ansys Explicit Dynamics to simulate material fracture behavior based on real-world test standards. These Ansys results are then used as a benchmark to calibrate the fracture parameters within Unreal Engine's Chaos Destruction system. A genetic algorithm is employed to iteratively optimize these parameters, ensuring that UE's simulations closely match Ansys's high-fidelity results without sacrificing real-time performance. The calibrated material parameters are stored in a RESenv material library.
  • Scenario Preparation: 3D building models are imported into Unreal Engine and assigned fracture settings from the RESenv material library. Real earthquake waveform data is acquired from online databases (IRIS) and imported via a user interface.
  • Data Binding: Buildings are automatically bound to the virtual terrain using Unreal Engine's Physics Constraint Actor (UEPCA). This ensures that terrain movements from earthquake data accurately translate to structural stresses and deformations. Earthquake waveform data is converted into terrain displacements for this binding.
  • Simulation: The RESenv platform operates at various frame rates (40 FPS for desktop, 90 FPS for VR, 240 FPS for high-frame-rate sensor training). As the simulation runs, the earthquake data displaces the terrain, causing the pre-fractured buildings to be destroyed in a realistic manner.

Datasets

  • Earthquake Waveform Data: Acquired from the IRIS online database, providing three degrees of freedom (BH1, BH2, BHZ) for geological motion.
  • Building Models: Recreated using Blender, with surface textures from the Quixel Megascans material library.
  • Real-world Earthquake Damage Scenarios: Four scenarios were selected for comparison: two laboratory shake table experiments and two natural earthquake events.

Key Insights and Ideas

  • Bridging the Simulation Gap: The core insight is that by using Ansys for high-fidelity material simulation and then calibrating Unreal Engine's physics engine with these results, it's possible to achieve both realism and real-time performance. This overcomes a significant limitation of using game engines directly for scientific simulations.
  • Genetic Algorithm for Calibration: The use of a genetic algorithm for material parameter optimization is a novel approach to align disparate simulation engines, demonstrating its effectiveness in achieving desired accuracy.
  • UEPCA for Realistic Binding: The Physics Constraint Actor (UEPCA) is leveraged to realistically bind buildings to the terrain, a crucial step for simulating how ground motion affects structures.
  • AI Training Focus: The platform is explicitly designed for AI training, with a focus on generating synthetic data for tasks like visual recognition and path planning, addressing the data scarcity problem in real-world disaster scenarios.
  • Realism Beyond Ideal Conditions: The experiment highlighting issues with virtual sun glare affecting robot vision underscores the importance of considering environmental factors beyond just structural damage for realistic AI training.

Key Results

  • Material Calibration: RESenv achieved comparable destruction results to Ansys simulations, particularly in terms of crack patterns and stress distribution, for concrete and masonry materials.
  • Similarity Detection: Simulated damaged buildings achieved a high degree of resemblance to real-world scenarios, with similarity scores ranging from 89.34% to 94.80% across four different building materials.
  • Robot Path Planning: In a simulated Japanese neighborhood, 80% of path-planning trials achieved a 100% completion rate, demonstrating the effectiveness of RESenv for training AI path-planning algorithms.
  • Image Segmentation: The Segment Anything Model (SAM) achieved an overall 95% accuracy in detecting edges when compared to ground truth, indicating the quality of synthetic data generated by RESenv for visual recognition tasks.

Limitations and Future Work

Limitations:

  • Focus on Buildings: Current simulations primarily focus on building damage, abstracting more complex dynamics like foundation flexing and land movements.
  • Idealized Materials: Building material parameters are derived from idealized configurations, not fully capturing the diversity of real-world materials and structures.
  • Environmental Factors: The simulation environment does not currently account for variable environmental conditions like changing lighting, smoke, or dust, which can impact AI performance.

Future Work:

  • Online Material Database: Develop a public database of calibrated materials to make RESenv more accessible to users unfamiliar with Ansys.
  • Dynamic Terrain: Implement a two-dimensional dynamic terrain model that accurately reflects seismic wave deformations for enhanced realism.
  • Environmental Factors: Incorporate dynamic environmental factors (lighting, smoke, dust) to further enrich the simulation environment.
  • Broader Disaster Types: Expand RESenv's capabilities to simulate other disaster types, such as floods and hurricanes.

Summary updated

BibTeX
@inproceedings{sun2024executing,
  author = {Yitong Sun and Hanchun Wang and Zhejun Zhang and Cyriel Diels and Ali Asadipour},
  title = {{Executing realistic earthquake simulations in unreal engine with material calibration}},
  booktitle = {Comput. Graph},
  year = {2024},
  url = {https://openreview.net/forum?id=y1uhXJzpvn},
}
Figure