Reaction Profile Forecasting by Artificial Data Generation for Wittig-Type Geminal Bromofluoroolefination.

Area of Science:

• Organic Chemistry
• Computational Chemistry
• Chemical Informatics

Background:

• Machine learning (ML) is increasingly used in organic synthesis for reaction prediction and optimization.
• Large datasets are typically required for ML models, posing challenges for experimental scientists.
• Existing ML approaches often rely on extensive feature sets and data preparation.

Purpose of the Study:

• To develop simple ML models for predicting reaction profiles in geminal bromofluoroolefination.
• To address the challenge of limited experimental data in ML-driven reaction development.
• To explore the efficacy of data augmentation strategies with minimal feature sets.

Main Methods:

• Developed ML models using minimal, readily accessible features like 13C NMR chemical shifts and Sterimol values.
• Employed a tabular augmentation method by fitting sparse data points to sigmoidal curves.
• Combined data augmentation with a conditional tabular generative adversarial network (CTGAN).
• Utilized a feed-forward neural network (FNN) for prediction.

Main Results:

• Simple ML models achieved effective prediction of reaction profiles with minimal data.
• Tabular data augmentation significantly enhanced the predictive ability of the FNN.
• The combination of augmentation techniques with CTGAN further refined model performance.
• Demonstrated the utility of sigmoidal curve fitting for sparse data augmentation.

Conclusions:

• Tailored data augmentation strategies are effective solutions for small experimental datasets in ML.
• This approach simplifies ML model development for experimental scientists.
• The study provides a pathway for more accessible ML applications in organic synthesis.
• Highlights the potential of CTGAN and sigmoidal augmentation for enhancing ML model accuracy.