Background
Accurate simulation of fluids is important for many science and engineering problems. Computational fluid dynamics (CFD) solutions (also known as fluid flow profiles) are typically calculated by numerically solving the partial differential Navier Stokes (NS) equations governing the environment and boundary conditions. This is usually done through off-the-shelf software such as Mechsys or Ansys*. However, this is an iterative, time-consuming, and compute- and memory-intensive job. These factors are a deterrent for rapid design and development of infrastructure where aerodynamics plays a critical role in efficient operation, for example, design of a wind turbine blade, the spoiler of a Formula 1* car, or even the stacking of server chips in a large data center where the wind flow will affect cooling patterns or lead to hot spots.
Solution
In collaboration with Accenture*, Intel developed this computational fluid dynamics AI reference kit. Paired with Intel® software, this kit may help customers in developing a deep learning model for calculating a fluid flow profile.
An AI-based (deep learning) solution can serve as a replacement for numerical simulations and can provide quick approximate solutions.1
Although highly compute intensive, computational fluid dynamics codes are largely written and optimized to run on CPU architecture. The deep learning model, through a simple inference job, allows for faster design tests, increasing the throughput and enabling quicker design updates.
End-to-End Flow Using Intel® AI Software Products
The deep learning model for calculating a fluid flow profile is built using TensorFlow*. The input for training is images with random geometric shapes around which the fluid flow profile is calculated. This is called the boundary. The output is a 2D velocity vector (Vx, Vy) at each pixel that denotes the velocity value at each location. This can then be converted into a fluid profile image for visual representation. The model used was trained on 2,560 images for 50 epochs, with a batch size of 8. The final loss (mean-squared error) value was 4.82e-6.
For fluid flow profiling, the accuracy of the results is a critical aspect for design considerations. For an AI solution that approximates the results, the trade-off between the loss in accuracy and the speedup in model training and inference should be examined carefully before adoption.
This reference kit includes:
- Training data
- An open source, trained model
- Libraries
- User guides
- Intel® AI software products
At a Glance
- Industry: Automotive, aerospace and defense, energy, manufacturing, high tech
- Task: Help detect the anomalies in data collected from IoT Devices to monitor equipment condition and prevent any issue from being cascaded in the entire operation flow.
- Dataset: TFRecord file of 3001 images with random geometric shapes and the profile of a fluid flowing around it (boundary).
- Type of Learning: Deep learning
- Models: U-Net architecture
- Output: 2D velocity vector (Vx, Vy) at each pixel that denotes the velocity value at each location. It's converted to a fluid profile image when combined with boundary information.
- Intel AI Software Products:
- Intel® Optimization for TensorFlow* with oneAPI Deep Neural Network Library (oneDNN)
- Intel® Neural Compressor
Technology
Optimized with Intel® oneAPI for Better Performance
The computational fluid dynamics model was optimized by oneDNN and Intel Neural Compressor for better performance across heterogeneous XPU and FPGA architectures.
Intel Neural Compressor and oneDNN allow you to reuse your model development code with minimal code changes for training and inferencing.
Performance benchmark tests were run on Microsoft Azure* Standard_D8_v5 using 3rd generation Intel® Xeon® processors to optimize the solution.
Benefits
An AI-based model can provide a faster alternative to numerically solving NS equations to calculate fluid flow profiles. This approach can accelerate component design and development. Additionally, using technologies optimized by Intel (for example, TensorFlow with oneDNN optimizations and int8 quantization using Intel Neural Compressor for inference) can accelerate pipelines even further.
This is critical for workloads deployed on modest CPU architectures such as on-premise servers that cannot be customized or local workstations. These benefits outweigh the negligible accuracy losses seen in an AI-based solution. With Intel® oneAPI toolkits, little to no code change is required to attain the performance boost.