Enhancing Molecular Dipole Moment Prediction with Multitask Machine Learning

Area of Science:

• Computational chemistry
• Machine learning applications in chemistry

Background:

• Predicting molecular dipole moments is crucial for understanding molecular properties.
• Traditional methods may lack efficiency or accuracy.
• Machine learning offers potential for improved predictions.

Purpose of the Study:

• To develop a multitask machine learning strategy for enhancing molecular dipole moment prediction accuracy.
• To investigate the utility of incorporating low-quality auxiliary data (Mulliken charges) into the training process.
• To assess the impact of auxiliary data on learning physically grounded molecular representations.

Main Methods:

• Implemented a multitask learning framework.
• Trained a model simultaneously on quantum dipole magnitudes (primary target) and Mulliken atomic charges (auxiliary task).
• Utilized a weighted loss function to balance the contribution of each task.

Main Results:

• Incorporating Mulliken atomic charges improved dipole prediction accuracy by up to 30%.
• The multitask approach led to more physically grounded representations of molecular charge distributions.
• Enhanced accuracy and consistency in predicting dipole magnitudes were observed.

Conclusions:

• Multitask learning with auxiliary data, even if quantitatively unreliable, can significantly improve predictive models.
• Qualitative physical insights from low-quality data enhance machine learning model performance.
• This strategy offers a pathway to more accurate and robust predictions of molecular properties.