Training Strategy

Three Training Regimes

To evaluate the impact of transfer learning, we trained three models:

1. DFT Scratch Model

• Trained only on DFT data (80% split).

• No pretraining.

• Evaluated on 20% held-out DFT test set.
  
  

This serves as the baseline.

2. BVSE Pretraining

• Model trained on 122,421 BVSE barriers.

• Learns general structural diffusion features.

• No exposure to DFT barriers during this stage.

3. BVSE → DFT Fine-Tuning

• Initialize model with pretrained BVSE weights.

• Fine-tune on 80% DFT training data.

• Evaluate on same held-out DFT test set.
  
  

This isolates the effect of pretraining.

Why Transfer Learning?

The BVSE dataset is large but approximate. The DFT dataset is small but accurate.

The hypothesis:

Pretraining on approximate physics allows the model to learn structural representations that transfer to high-fidelity DFT prediction.