Conditional Variational AutoEncoder to Predict Suitable Conditions for Hydrogenation Reactions.

Area of Science:

• * Computational Chemistry
• * Machine Learning in Chemistry
• * Synthetic Chemistry

Background:

• * Accurate prediction of reaction conditions (RCs) is vital for chemical synthesis planning.
• * Multiple RC combinations can yield desired outcomes, complicating condition recommendation.
• * Existing methods often require extensive enumeration or combinatorial searches.

Purpose of the Study:

• * To develop a generative model for predicting suitable reaction conditions.
• * To enhance flexibility and accuracy in reaction condition recommendation.
• * To circumvent the need for combinatorial searching of potential RCs.

Main Methods:

• * A conditional variational autoencoder (CVAE) generative model was employed.
• * The CVAE was customized to generate diverse sets of valid RCs.
• * Three latent distribution variants (Gaussian, Riemannian Normalizing Flow, Hyperspherical Uniform) were tested to optimize accuracy.
• * Models were evaluated on datasets of varying sizes and complexity.

Main Results:

• * The CVAE model successfully predicted catalysts, additives, temperature, and pressure for hydrogenation and H2-mediated reactions.
• * The Hyperspherical Uniform CVAE (h-CVAE) demonstrated superior overall performance and accuracy for top-k predictions.
• * Benchmarking confirmed the high performance of CVAE models against state-of-the-art approaches.
• * The model handled datasets with up to ~7 × 10^42 potential condition combinations.

Conclusions:

• * CVAE models offer a flexible and accurate method for predicting chemical reaction conditions.
• * The h-CVAE variant is particularly effective for applications requiring high accuracy.
• * This approach significantly advances automated chemical synthesis planning.