Fire prediction and simulation for VIP with Prof. Bo Zhu and mentor Duowen.
Data generated using GPU-GEMS-3D-Fluid-Simulation GPU based simulation in Unity. We use a simple CNN with L2 regularization to predict the Temperature Amount parameter from image, based on the flame shape, smoke amount, and color.
The "Temperature Amount" parameter range is [5, 200]. The Mean Absolute Error of the model's predictions on the test dataset is 10.53 after 150 epochs, with LR=0.0001, batch_size=16. The MAE expressed as a percentage of the output range is 5.4%.
We take two images of the fire per "Temperature Amount," so we have 390 synthetic images. We use 80% as training data and 20% as test data.
Below are some example temperature predictions on 3 images. Images can be found in the unitySim/UnitySim5-200 folder.
Image: Temp1_022.png
Predicted temperature: 32.17
Actual temperature: 22.00
Absolute error: 10.17
Image: Temp1_102.png
Predicted temperature: 107.19
Actual temperature: 102.00
Absolute error: 5.19
Image: Temp_197.png
Predicted temperature: 175.44
Actual temperature: 197.00
Absolute error: 21.56
Below is a graph of the MAE per each image in the test set, with the X axis being test set images. Images are ordered from lowest temperature (6) to highest temperature(199).
Synthetic data generated in Embergen of a thin candle flame. We use a CNN to predict temperature from image, due to flame shape and color.
The Mean Absolute Error of the model's predictions on the test dataset is 39.92 after 50 epochs. The MAE expressed as a percentage of the output range is 1.996%.
Below are some example temperature predictions on 3 images. Images can be found in the genericFire/GenericFire3000-5000K folder.
Image: generic_fire16_3089.1K.png
Predicted temperature: 3076.68K
Actual temperature: 3089.10K
Absolute error: 12.42K
Image: generic_fire301_4676.9K.png
Predicted temperature: 4734.52K
Actual temperature: 4676.90K
Absolute error: 57.62K
Image: generic_fire346_4927.6K.png
Predicted temperature: 4821.37K
Actual temperature: 4927.60K
Absolute error: 106.23K
In the folder genericFire, run python3 genericFire_train.py.
This script splits the images in GenericFire3000-5000K folder into a training set and a test set. It will use the training set to train a model, which is saved as genericFire_temperature_model.h5. The script also generates a text file genericFire_test_set_filenames.txt that lists the images in the test set for our reference.
In the folder genericFire, run python3 genericFire_inf.py.
This script runs inference on a list of images from GenericFire3000-5000K which were part of the test set. At the bottom of the script, you can modify the selected_images list to pick which images you want to run inference on. Make sure they are images in the test set (see genericFire_test_set_filenames.txt, generated by genericFire_train.py).
You can also uncomment the first few lines in __main__ to run inference on a randomly selected image from the test set.
Synthetic data generated in Embergen of a thin candle flame. We use a CNN to predict temperature from image, due to candle shape and color.
The Mean Absolute Error of the model's predictions on the test dataset is 297.64 after 50 epochs. The MAE expressed as a percentage of the output range is 14.882%.
Below are some example temperature predictions on 7 images. Images can be found in the candle/CandleFire3000-10000K folder.
Image: candle3_3087.5K.png
Predicted temperature: 4249.16K
Actual temperature: 3087.50K
Absolute error: 1161.66K
Image: candle46_4341.7K.png
Predicted temperature: 4028.54K
Actual temperature: 4341.70K
Absolute error: 313.16K
Image: candle381_5858.3K.png
Predicted temperature: 5891.45K
Actual temperature: 5858.30K
Absolute error: 33.15K
Image: candle361_6441.7K.png
Predicted temperature: 6589.43K
Actual temperature: 6441.70K
Absolute error: 147.73K
Image: candle338_7112.5K.png
Predicted temperature: 7567.24K
Actual temperature: 7112.50K
Absolute error: 454.74K
Image: candle185_8395.8K.png
Predicted temperature: 8769.85K
Actual temperature: 8395.80K
Absolute error: 374.05K
Image: candle231_9737.5K.png
Predicted temperature: 9553.99K
Actual temperature: 9737.50K
Absolute error: 183.51K
In the folder candle, run python3 candle_train.py.
This script splits the images in CandleFire3000-10000K folder into a training set and a test set. It will use the training set to train a model, which is saved as candle_temperature_model.h5. The script also generates a text file test_set_filenames.txt that lists the images in the test set for our reference.
In the folder candle, run python3 candle_inf.py.
This script runs inference on a list of images from CandleFire3000-10000K which were part of the test set. At the bottom of the script, you can modify the selected_images list to pick which images you want to run inference on. Make sure they are images in the test set (see candle_test_set_filenames.txt, generated by candle_train.py).
You can also uncomment the first few lines in __main__ to run inference on a randomly selected image from the test set.
The candle_combined_train_inf.py will train and run inference in one script.