Deep Learning Surrogate for Turbulent Flow Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computational fluid dynamics (CFD) simulations for turbomachinery require significant computation time, taking hours, days, or even weeks, which hinders efficient design and optimization processes.
Innovation Solution
A predictive model using deep learning is developed to predict CFD airflow around objects like airfoils and engines, reducing computation time to seconds by training a neural network to replicate traditional CFD simulation results, allowing for rapid prediction of flow attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional CPU-driven CFD simulations are used, then accurate fluid flow predictions are obtained, but computation time becomes excessively long (hours to weeks)
Solution Approach 1:
The patent creates a neural network surrogate model that copies the input-output behavior of traditional CFD simulations. The neural network is trained on CFD simulation data to learn the mapping between geometric parameters and flow characteristics, enabling rapid prediction without re-running full CFD simulations. This copying approach maintains prediction accuracy while reducing computation time from hours/weeks to seconds.
Solution Approach 2:
The patent replaces the mechanical CPU-based numerical computation system with a neural network-based predictive system. Instead of solving partial differential equations through iterative numerical methods on high-performance computing clusters, the system uses a trained neural network model that has learned the underlying flow patterns, substituting heavy mechanical computation with faster neural network inference.
2Loss of time
If simple designs are simulated, then computation time is reduced to a few hours, but design complexity must be limited
Solution Approach 1:
The patent performs preliminary action by training the neural network model on a comprehensive dataset covering a wide range of design configurations and complexities before actual use. This pre-training phase captures the relationships between various geometric parameters and flow characteristics for different levels of design complexity, enabling the model to handle complex designs rapidly without requiring complex computations during the prediction phase.
Data Source
AI summary
The example embodiments are directed to a system and method for predicting a flow about an object through the use of a predictive model instead of a machine simulation. Traditional CFD simulations can take hours, even days. The example embodiments provide a predictive model that can predict a CFD flow in seconds which greatly improves design time. In one example, the method may include receiving input data comprising shape parameters of a geometric object and flow parameters associated with the geometric object, predicting, via execution of a predictive model, a computational fluid dynamic (CFD) flow about the geometric object based on the shape parameters and the flow parameters included in the input data, and outputting one or more attributes of the predicted CFD flow about the geometric object via a display device.