How to construct and use delta-learning models
Delta-learning is a powerful concept that inspired many works of ML in computational chemistry.
The idea is simple: add ML correction to the predictions of the baseline low-level QM method to approximate the target high-level QM method.
Think about correcting semi-empirical methods to the quality of coupled cluster: this is exactly what we have done in the AIQM1 method!
Delta-learning is slower than pure ML models but the benefit is typically an increased robustness and accuracy. Here is an example where delta-learning model AIQM1 gives correct bond lengths in fullerene while the pure ML model ANI-1ccx fails. The price to pay is that delta-learning models are slower than pure ones.
You can learn how to construct such robust delta-learning models in our online tutorial. Please check out the hands-on online mini-course Modern AI and computational chemistry by Prof. Pavlo O. Dral.
Leave a Reply