Empowering Reactivity Predictions through Noise-Based Data Augmentation.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Data scarcity is a major challenge for Artificial Intelligence (AI) in molecular science.
• Data augmentation is a common technique in other fields but its applicability to molecular reactivity is unknown.

Purpose of the Study:

• To evaluate the effectiveness of data augmentation for molecular reactivity prediction.
• To determine if data augmentation can improve AI model performance in low-data scenarios for chemical reactions.

Main Methods:

• Systematic evaluation of data augmentation on diverse reactivity problems.
• Application of Gaussian noise to existing data points for data augmentation.
• Training AI models with augmented and original datasets.

Main Results:

• Data augmentation significantly enhances predictive performance for molecular reactivity.
• Models trained with augmented data achieve accuracy comparable to models trained on full datasets.
• Data augmentation enables meaningful model training in low-data regimes.

Conclusions:

• Data augmentation is a powerful strategy for overcoming data scarcity in AI for molecular reactivity.
• This approach reduces the need for extensive experimental data, saving time and resources.
• Data augmentation accelerates the integration of machine learning in chemical research.