Shape Optimization via CNN Surrogates for Fluid Dynamics
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current shape optimization methods for technical devices exposed to fluids, such as airfoils and turbines, are inefficient due to reliance on genetic algorithms and low-dimensional parametrization, which limit computational efficiency and accuracy.
Innovation Solution
A computer-implemented method using a Convolutional Neural Network (CNN) that discretizes the shape of technical devices into points or surface meshes, allowing for non-parametric input and enabling faster, more accurate optimization of fluid dynamic parameters like pressure profile, drag coefficients, and lift-to-drag ratio without the need for low-dimensional assumptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If genetic algorithms with local search heuristics are used for shape optimization, then the optimization can be performed without access to gradients, but the computational cost becomes very expensive and efficiency is limited
Solution Approach 1:
The patent replaces traditional mechanical optimization algorithms (genetic algorithms with local search) with a neural network-based system. The CNN learns the mapping from shape to fluid dynamic parameters, enabling gradient-based optimization that is computationally efficient while maintaining the ability to optimize without explicit gradient access during the optimization process itself.
Solution Approach 2:
The patent introduces a CNN as an intermediary between the shape representation and fluid dynamic parameter prediction. This intermediary model approximates the complex CFD simulation process, allowing for fast evaluation and gradient computation without running expensive CFD simulations at each optimization step.
2Device complexity
If low dimensional parametrisation of shape is assumed to reduce optimization space complexity, then the optimization problem becomes more manageable, but the method becomes prohibitively inefficient and restricted to a low dimensional manifold
Solution Approach 1:
The patent changes the dimensionality representation by using pixel-based or point-based representations of shapes instead of low-dimensional parametric models. This allows the optimization to explore a much higher dimensional space of possible shapes while the CNN efficiently processes and evaluates each configuration, resolving the trade-off between search space complexity and optimization efficiency.
3Measurement precision
If CFD simulators are used to evaluate each proposed airfoil shape at each time step, then accurate fluid dynamic predictions are obtained, but the computations are very expensive and limit optimization efficiency
Solution Approach 1:
The patent creates a neural network copy or surrogate model of the CFD simulation process. The CNN is trained on CFD data to learn the mapping from shape to fluid dynamic parameters, then uses this learned model for fast prediction during optimization. This copying approach maintains accuracy while dramatically reducing computational cost compared to running full CFD simulations at each step.
4Ease of operation
If many local deformations are evaluated to find a good direction of descent without gradient access, then the optimization can proceed without gradient information, but the number of evaluations required increases computational cost significantly
Solution Approach 1:
The patent replaces gradient-free optimization mechanics with gradient-based optimization by using the CNN's differentiable architecture. The automatic differentiation capability of the neural network provides efficient gradient computation, allowing the optimization to converge in far fewer iterations compared to evaluating many local deformations without gradient information.
Data Source
AI summary
The present invention generally relates to a method, a system and a computer program for shape optimisation of a technical device adapted to be exposed to a fluid flowing around a contour of said device with respect to its fluid dynamic parameters. In order to provide an improved method for shape optimisation of a technical device with respect to its fluid dynamic parameters which is faster, a method is provided comprising discretizing the shape of the technical device into a plurality of points along the contour of the technical device or into a surface mesh, and inputting the plurality of points or the surface mesh into a Convolutional Neural Network (CNN) for computing a prediction of the at least one fluid dynamic parameter.