| name | delta_ml_transfer_learning |
| description | Implementing Delta-ML (residual learning between low and high levels of theory) and model transfer learning. |
Delta-ML & Transfer Learning
Use this skill when attempting to achieve "chemical accuracy" (~1 kcal/mol) for 3D molecular properties (e.g. system energies, dipole moments, solvation free energy) with limited high-fidelity quantum mechanical or experimental data.
Conceptual Framework
Delta-ML ($\Delta$-ML)
Instead of predicting the high-level quantum mechanical (HL) property directly from structure, Delta-ML predicts the residual difference between a cheaply calculated low-level (LL) value (e.g., semi-empirical methods like PM7 or low-basis DFT) and the HL value:
$$\Delta = \text{Property}{\text{HL}} - \text{Property}{\text{LL}}$$
The ML model is trained to predict $\Delta$. The final prediction is:
$$\text{Property}{\text{HL_Predicted}} = \text{Property}{\text{LL}} + \Delta_{\text{Predicted}}$$
This greatly reduces the learning complexity as the low-level calculation already captures the primary physical interactions.
Transfer Learning
- Pre-training: Train a neural network model on a large database of cheap calculations (e.g., COSMO-RS or low-level DFT coordinates).
- Fine-tuning: Keep representation backbone weights or adjust with a small learning rate, and fine-tune the output layers on high-quality experimental data.
Execution
Run the script delta_transfer_ml.py to compare standard regression, Delta-ML, and transfer learning:
python scripts/delta_transfer_ml.py --epochs 30 --lr 0.01
This showcases predictions of quantum residuals and fine-tuning transitions.