Thanks for your nice work! I tried to reproduce the 'MAE on SPICE / TorchMD-Net' from Figure 2 d in your preprint using your config.yaml and your provided checkpoint with torchmdnet. The only two parameters I change are batch_size=64 and num_workers=16.
In your preprint, the MAE is 9.8 or 9.0, without and with pretrain (50 epochs). But I got 17.26 and 18.50 in my two runs (10 epochs), which means the result got worse after loading the pretrained weights. What might be the problem? Besides, I wonder in your preprint, did you use a different config of torchmdnet?