Question Regarding Reproducibility and Brittleness of MEGNet-MP-2019.4.1-BandGap-mfi Model

[akima6]

I am writing to report a potential reproducibility issue I've encountered with the pre-trained MEGNet-MP-2019.4.1-BandGap-mfi model. After a thorough investigation, I believe the model may be exhibiting "brittleness," where its predictions are highly sensitive to the modern software environment, leading to significant deviations from the expected results.
The Issue:
When using the model to predict band gaps for well-known materials from the Materials Project, the predictions for semiconductors and insulators are catastrophically incorrect, often collapsing to near-zero. The model only seems to perform correctly for metals.
This happens even when using the exact methodology outlined in the repository's example notebooks—specifically, by using the Structure2Graph converter that the model was originally trained with.

Material	MP ID	Model Prediction (eV)	MP Ground Truth (eV)	Absolute Error (eV)
Si	mp-149	-0.007	0.611	0.617
GaAs	mp-2534	0.063	0.186	0.123
MgO	mp-1265	1.443	4.429	2.987
Cu	mp-30	-0.010	0.000	0.010

As you can see, the errors for Si and MgO are extremely large, indicating the model is not generalizing as expected.

Environment Details:
OS: Windows 11
Python: 3.10
matgl: 1.0.1
pymatgen: 2024.5.21
dgl: 1.1.3+cu118 (or CPU equivalent)
torch: 2.1.2+cu118 (or CPU equivalent)

My Question:
Could you confirm if this is a known issue? The discrepancy between the model's reported MAE and its performance in a modern environment is significant. My hypothesis is that subtle changes in the dependency libraries (pymatgen, dgl) over the years are creating graph representations that the older, brittle model can no longer interpret correctly.
Would it be possible to share the original requirements.txt or environment.yml file used for the 2019 training? This would be invaluable for the community to be able to fully reproduce the model's original performance.
Thank you for your time and for all your work on this excellent library.
Best regards,
R Akhil Raj
predict_bandgap.py