DrivAerNet replication issues

[brooks-Miles]

Hi,

Thank you very much for open-sourcing this dataset. While reproducing the DrivAerNet paper, I used the 4,000 .stl models with the DrivAer_ prefix previously downloaded from Dropbox, along with the AeroCoefficients_DrivAerNet_FilteredCorrected.csv file that records drag coefficients.

First, following the hyperparameter values in the Implementation Details section of the DrivAerNet paper, I modified the patience value in train.py from the DrivAerNet_v1 folder to 10, while keeping all other parameters consistent with the paper, and conducted a training session. Here are the results I obtained:
Testing the final model:
Test MSE: 0.000723,Test MAE: 0.023060,Max MAE: 0.027240,Test R²: 0.0067
Total inference time: 0.31s for 576 samples
Testing the best model:
Test MSE: 0.000717,Test MAE: 0.022918,Max MAE: 0.026814,Test R²: 0.0167
Total inference time: 0.31s for 576 samples

Second, I adjusted the patience value back to 20 and conducted another training session using the same parameters as provided in your train.py. Here are the results I obtained:
Testing the final model:
Test MSE: 0.000994,Test MAE: 0.027736,Max MAE: 0.032676,Test R²: -0.4203
Total inference time: 0.34s for 576 samples
Testing the best model:
Test MSE: 0.000172,Test MAE: 0.010169,Max MAE: 0.013186,Test R²: 0.7665
Total inference time: 0.58s for 576 samples

In both training sessions, the R² value only reached a maximum of 0.7665, which falls short of the 0.901 reported in the paper. I would like to ask where this issue might lie. This problem has been troubling me for a long time, and I sincerely hope you can provide an explanation. Thank you very much in advance!