CASES AI｜Vehicle Aero CFD Flow-Field Prediction for Automotive Air-Conditioning Piping｜physicsAI

案例實績

所有實績分類

CFD Flow-Field Prediction for Automotive Air-Conditioning Piping｜physicsAI

Apr.10,2024

In most cases, CAE engineers in industry work on very similar models, such as CFD of HVAC ducts or CAE work for automotive bumpers. Through Design of Experiments (DOE) iterations, they continuously optimize dimensional parameters and product shape to reduce flow resistance and improve flow uniformity in ducts, or to achieve lightweight structures while enhancing safety and durability.

Normally, after a project analysis is completed, the大量 simulation result files stored on the server side are deleted. Even if they are backed up, they are rarely reused. Whenever a new analysis project starts, CAE engineers still need to repeat the steps of geometry cleanup, meshing, and solving from scratch.

Can CAE engineers instantly predict results when designers propose a new design?

physicsAI uses historical CAE data as training samples for machine learning. Whenever new CAD data or surface meshes are provided as input, it can quickly predict physical field distributions such as pressure, temperature, stress, and deformation. In this way, CAE engineers can perform prediction and optimization much faster and significantly improve their productivity.

【 Value of physicsAI 】

Based on geometric deep learning, no manual parameterization is required. (Many CAE models are difficult to parameterize, and manually created parameter sets are often not optimal.)
Users do not need to write code or edit scripts.
Predictions can be made directly on surface meshes or CAD models, skipping the complex modeling process.
Prediction is much faster than a simulation solver and supports design exploration.
It can be used to predict arbitrary physical fields (structures, fluids, electromagnetics, etc.).

【 Using physicsAI 】

Install HyperWorks Desktop v2023.
Supports Nvidia GPU training with at least 8 GB of VRAM.
Training time depends on the number of samples and the mesh size of each sample. Small models may take about half an hour, while larger models can take more than a day.
Training samples can be run locally or on a remote HPC.
Training samples must be in .h3d, .fem, or .rad result formats, while prediction inputs can be Parasolid or surface meshes.

In the following, we will demonstrate how to use physicsAI with an HVAC duct model. The basic workflow consists of four steps:

Build the dataset
Train the model
Test the model
Run predictions with the trained model

Finally, solver-based validation is an optional step.

In the demo model, the geometry is created with Inspire. HyperStudy drives five CAD parameters to build a DOE study. The flow field is solved by AcuSolve, generating seven training samples and two test samples, plus two new designs for prediction.

Inspire design parameters

physicsAI runs on a laptop (Intel CPU i7-10850H, 32 GB RAM). For this flow case, the prediction speed is 7–8 times faster than the CFD solver:

Epochs = 300, training time 18 min 15 sec

Epochs = 100, training time 6 min 18 sec

When making predictions, physicsAI quantifies how similar the input design is to the training data using a confidence score, which is shown in the upper-right corner of the window.

A confidence score of 1.0 indicates that the input design is identical to one of the training points, which is the maximum possible value.
A confidence score of 0.0 indicates that the difference between the input design and its nearest training point is equal to the difference between the two most distant training points.
A negative confidence score means the input design is very different from the training data. Unless a new model is trained with similar designs, the prediction quality may be poor.

Determinism: approximate determinism can be achieved by setting a random seed in the configuration file. However, even after doing so, it is normal to observe subtle differences between two models trained with the same settings and data. Using more training samples and including more design concepts in the dataset can improve the generalization capability of predictions.
To ensure good demo performance, the example model was further trained with 280 sample points (7 duct types × 40 size combinations each), taking 3 hours to train on a GPU server.

Comparison of pressure distribution: AcuSolve (left) vs. physicsAI prediction (right)

Comparison of flow velocity: AcuSolve (left) vs. physicsAI prediction (right)
physicsAI can also be launched in Inspire to predict directly on CAD models.

Confidence score to evaluate prediction reliability

Bumper crash safety model prediction:

【 Typical Training Time for Common CAE Models 】

Explicit dynamics: Ball impact
Fluid dynamics: HVAC duct
Solid mechanics: Bracket static analysis

【 Some Basic Concepts of physicsAI 】

Q: What is geometric deep learning?
A: Geometric deep learning is a neural-network-based approach that learns from non-Euclidean data types.

Euclidean structure data includes images, text, audio, etc.:
Non-Euclidean structure data can represent more complex structures than 1D or 2D expressions, such as molecular structures, neural networks, Feynman diagrams, and manifold data, etc.:

【 Training Parameters in physicsAI 】

Epochs: number of training iterations.
Early Stopping / Patience: stops training early if the model does not improve after several iterations.
Learning Rate: step size for each training update (larger is faster, but may fail to converge).
Width: controls the complexity of patterns that can be learned.
Depth: controls the complexity of patterns and the distance over which information can propagate locally. (Has a strong impact on training time.)

【 Setting a Reasonable Learning Rate 】

A larger learning rate will move quickly toward the optimal direction, but if it is too large, the model may never reach a converged solution.
A smaller learning rate is more likely to converge but may take much longer to reach convergence.
A good learning rate should be large enough for fast convergence, but not so large that optimization stalls before reaching convergence. Noisy loss curves in the log indicate that the learning rate may be too high. Start with the default value 1e-3 and, if it is noisy, try 1e-4.

Comparison of different learning rates:

【 Width and Depth 】

Width controls how much detail physicsAI can see, while depth controls how far physicsAI can see at one time.
Training time for the model roughly follows the relation: Training Time = Width × Depth.

【 Overfitting and Underfitting 】

Overfitting refers to a phenomenon where the model performs extremely well on the training set but poorly on the test set or unseen data. The main reason is that the model is too complex and learns too much of the noise and fine details in the training set, resulting in weaker generalization to unseen data. A typical sign of overfitting is that every sample in the training set is well fitted, while performance degrades significantly on new data.

Underfitting refers to a phenomenon where the model performs poorly on both the training set and the test set. The main reason is that the model is too simple and cannot capture the complexity and features of the data. A typical sign of underfitting is that the model cannot fit the training samples well, leading to poor performance on both training and test data.

HVAC duct model operation demo

Training physicsAI

Testing a physicsAI model

Real-time physics predictions

From CAD to contour

Source of article: Altair official blog

Richin Tech is the "expert in CAE and AI data analytics", and we have completed many successful case studies.

▶ Contact us now to get more information.
▶ Subscribe to the Richin YouTube channel to explore more about CAE and data analytics.

Back to list