Autonomous design of new chemical reactions using a variational autoencoder

Area of Science:

• Chemistry
• Artificial Intelligence
• Machine Learning

Background:

• AI models for chemistry are limited by biased experimental synthesis datasets.
• Existing datasets often only report optimal reactions, hindering comprehensive model training.
• Bias in AI chemistry models can lead to inaccurate predictions for reaction design.

Purpose of the Study:

• To develop an AI model for synthetically generating continuous datasets of chemical reactions.
• To address inherited bias in AI chemistry models by creating diverse training data.
• To enable complete exploration of chemical reaction design spaces.

Main Methods:

• Developed a Variational AutoEncoder (VAE) based artificial intelligence model.
• Employed latent space sampling to generate new chemical reactions.
• Utilized a small experimental dataset to generate a large synthetic dataset.

Main Results:

• Successfully generated over 7,000,000 new chemical reactions.
• The synthetic dataset significantly exceeds the size of the original training set (7,000 reactions).
• Generated molecular species are larger and more diverse than those in the training data.

Conclusions:

• The VAE model effectively generates large, diverse datasets for AI chemistry.
• This method mitigates bias in AI models by expanding the training data solution space.
• Synthetic data generation offers a powerful approach for advancing chemical reaction design.