2D transient thermal response predictions in fins using convolutional neural networks.
Variable parameters:
- Fin thickness
- Fin taper ratio
- Temperature boundary condition
- Convection coefficient
Input images (3×40×80):
- Channel 1: Fin shape
- Channel 2: Convection coefficient
- Channel 3: Temperature boundary condition
Output images (10×8×32 or 1×8×32):
- 2D temperature distribution over 10 instances in time (10 channels), or
- 2D thermal gradient distribution over 10 instances in time (10 channels), or
- 2D thermal stress distribution at final time only (1 channel)
Ansys (Mechanical APDL) is used to run simulations and generate response data, which are then processed in Python.
Define the simulation parameters in the script ansys_script.lgw. Constants are defined near the top of the file, and simulation parameters are defined in the for loops. The script loops over each combination of simulation parameters, and it simulates responses and saves the response data to text files. Refer to this reference on Ansys APDL commands.
Open Ansys and run the script ansys_script.lgw.
- File > Change Directory... > [Select the folder where the script exists]
- File > Read Input from... > [Select the script]
- To maximize performance, close the File Explorer window of the folder in which the program is running while simulations are running.
Each line in a text file contains the response value at a different node (not an element) along with the X and Y coordinates of that node. The coordinates of the node are used later to insert each response value into the proper location in a matrix. For transient simulations, a separate text file is written for each time step rather than having a single text file be written for all time steps. For static simulations, one text file is written for each simulation. Because of this difference in format between transient and static data, move all transient response files into one folder, and move all static response files into another folder.
Run preprocessing.py to process and cache the text files for easier loading when training models. Specifically, this code reads the response data contained in the text files and converts them into tensors and saves them as .pickle files. Caching the data in this way avoids having to read and convert the text files directly every time a model is trained, which takes longer.
Define which .pickle files should be used for which response data in the FinDataset class in datasets.py.
Call the main() function in main.py to train and test the model. The model architecture is defined in models.py. Results and model predictions are plotted after testing.
Set train to True to train the model, and set test to True to test the model. When training a model from scratch, train_existing must be set to False. Manually specify training hyperparameters, such as number of epochs, learning rate, batch size, etc.
During training, the best model weights (determined by lowest validation loss) are tracked and saved separately.