Dataset's chemical diversity limits the generalizability of machine learning predictions

Area of Science:

• Computational chemistry
• Materials science
• Chemical informatics

Background:

• The QM9 dataset is a benchmark for machine learning (ML) in predicting chemical properties, derived from GDB's chemical space exploration.
• Recent ML models achieve accuracy comparable to Density Functional Theory (DFT) calculations for molecular predictions.
• Generalizing ML models requires testing on diverse, real-world datasets.

Purpose of the Study:

• Introduce PC9, a new dataset equivalent to QM9 but with enhanced chemical diversity.
• Evaluate the performance of ML models on both QM9 and PC9 datasets.
• Assess the generalization capabilities of models trained on PC9.

Main Methods:

• Utilized Kernel Ridge Regression, Elastic Net, and the SchNet neural network model.
• Performed statistical analysis of bonding distances and chemical functions within PC9.
• Trained and tested ML models on both the QM9 and PC9 datasets.

Main Results:

• The PC9 dataset exhibits greater chemical diversity compared to QM9.
• While QM9 subset yielded higher overall energy prediction accuracy, PC9-trained models demonstrated superior generalization.
• Models trained on PC9 showed a stronger ability to predict energies for the other dataset.

Conclusions:

• PC9 represents a valuable, more diverse alternative to QM9 for ML model development.
• ML models trained on PC9 exhibit improved generalization, crucial for real-world chemical property prediction.
• The findings support the use of PC9 for robust testing and validation of ML models in chemistry.