CFD Super-Resolution Prediction With Uncertainty-Aware Mesh Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Computational Fluid Dynamics (CFD) simulations at high mesh resolutions are computationally intensive and time-consuming, and existing super-resolution techniques fail to accurately reflect real-world scenarios due to the lack of high-resolution mesh data and limitations in training data generation.
Innovation Solution
A method and system for high resolution simulation prediction in CFD that utilizes a rule-engine technique to select primary and secondary features, identifies a network architecture, and performs uncertainty analysis using aleatoric and epistemic uncertainty analysis to predict fine mesh data from coarse mesh data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high mesh resolution simulations are conducted to improve measurement precision, then simulation precision is improved, but computational intensity and time consumption increase
Solution Approach 1:
The patent uses machine learning models to create a computational shortcut that copies the essential physics relationships from high-resolution simulations. Instead of performing computationally intensive high-resolution simulations for every prediction, the system trains a neural network on a subset of high-resolution data and uses this trained model to quickly predict results for new cases, effectively copying the physics behavior without re-running expensive simulations.
Solution Approach 2:
The system performs preliminary action by pre-training machine learning models on a subset of high-resolution simulation data before actual predictions are needed. This preliminary training phase allows the model to learn the complex physics relationships once, and then subsequent predictions can be made quickly without reperforming the expensive high-resolution simulations for each new case.
2Productivity
If conventional super-resolution techniques are applied to reconstruct high-resolution data from low-resolution data, then computational efficiency is improved, but measurement precision deteriorates due to lack of alignment with physical PDEs
Solution Approach 1:
The patent changes the fundamental parameters of how super-resolution is achieved by moving from traditional image-processing-based methods to physics-informed machine learning. Instead of using conventional super-resolution algorithms that treat data as images, the system uses neural networks trained on physics-based governing equations, changing the approach from empirical image reconstruction to physics-constrained prediction.
Solution Approach 2:
The system replaces the mechanical image-processing mechanisms of conventional super-resolution with a machine learning-based approach. Rather than using traditional signal processing techniques that operate on pixel data, the patent substitutes a neural network that learns from physics equations and simulated data, replacing the mechanical image reconstruction process with an intelligent prediction system.
3Ease of manufacture
If down sampling high-resolution data to create training data is performed, then ease of manufacture of training data is improved, but loss of information increases due to retention of governing physics
Solution Approach 1:
The patent inverts the traditional super-resolution approach by not down-sampling high-resolution data to create training data. Instead, the system generates training data by running high-resolution simulations directly and using these complete high-resolution datasets for training the neural network. This inversion avoids the information loss that would occur during down-sampling while still enabling efficient training through modern ML techniques.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
This disclosure relates generally to high resolution simulation prediction for Computational Fluid Dynamics (CFD). CFD plays a crucial role in comprehending intricate physical phenomena spanning across scientific and engineering domains, hence it is essential to conduct simulations at high mesh resolutions for the governing equation of fluid flow. The current state-of-the-art super-resolution techniques involve reconstructing high-resolution data from down sampled low-resolution is limited to single scenario and does not accurately reflect real-world scenarios. The disclosed techniques enable prediction of fine-resolution data from low-resolution inputs from a variety of real-world CFD scenarios. Further the disclosed technique also identifies the most relevant network architecture for any CFD scenario and enabling accurate prediction of high-resolution data from low-resolution inputs by training the network architecture. Furthermore, the disclosure ensures also the robustness of the disclosed system through uncertainty analysis, encompassing both aleatoric and epistemic uncertainty analyses.