Transfer Learning from Simulation to Experimental Data: NMR Chemical Shift Predictions.

Area of Science:

• Computational Chemistry
• Machine Learning
• Spectroscopy

Background:

• Accurate prediction of chemical shifts (δ) is crucial for molecular structure elucidation but remains challenging.
• Current machine learning (ML) methods show promise but are hindered by the need for extensive chemical databases.

Purpose of the Study:

• To investigate the transferability of prior knowledge from simulation databases to predict experimental chemical shifts (δ).
• To enhance the applicability of ML models for chemical shift prediction using limited experimental data.

Main Methods:

• Utilizing prior knowledge gained from a simulation database.
• Applying additional training with small, randomly sampled experimental data.
• Leveraging knowledge transfer to train models on focused chemical spaces.

Main Results:

• Reliable accuracy for predicting experimental chemical shifts (δ) was achieved.
• Successful training of ML models on sparsely covered experimental chemical spaces was demonstrated.
• The approach effectively transferred knowledge despite differences between simulation and experimental databases.

Conclusions:

• Knowledge transfer from simulation databases is a viable strategy for improving experimental chemical shift prediction.
• This method enhances the usability of local experimental databases for ML applications.
• The approach offers a solution to the database size limitations in ML-driven chemical structure elucidation.