Overview
The "Nuberisim2AI" project focuses on accelerating flow-induced noise prediction through AI-driven time series extension. Turbulent flows around objects (e.g., pipes, turbines, vehicles, or aircraft) generate noise that must be minimized during development. Accurate 3D simulations are time-consuming, often requiring up to 14 days on high-performance computers (HPC).
By using AI to extend existing time series simulations—potentially by 50%—the project aims to reduce the overall simulation time by up to 75%, enabling engineers to analyze noise sources and optimize device designs more efficiently. The approach utilizes the KIT-Service Physical Intelligence: Kopplung von physikalischen und KI-basierten Modellen für Regression, Interpolation und Optimierung, which integrates physical models with AI for regression, interpolation, and optimization. This potentially allows engineers to investigate more design variants and enhance design optimization processes across multiple industries.
Our Goal
The primary goal is to develop and validate AI-based time series extension strategies that maintain high-quality data for noise source and intensity analysis. Challenges include:
- Ensuring data quality for training AI models by reducing high-dimensional datasets while preserving spatial and temporal fluctuations.
- Predicting at least 20 additional time steps (ideally 40–50) from an existing series of ~100 time steps, with a focus on acoustically relevant regions (e.g., high-gradient flow areas).
- Leveraging multiple datasets with varying time steps and spatial resolutions to improve AI prediction capability and reduce conventional simulation time.
- Creating a transferable system adaptable to different objects and industries, ensuring versatility for real-world applications.
