Gallou, Fabrice (2025) Empowering Micellar Catalysis and Representation Learning with Limited Data Availability: Surfactant Design Principle Can Boost Yield Predictions. ACS Catalysis (15). pp. 14207-14214.

Abstract

Accurate prediction of chemical reaction yields is crucial for advancing synthetic chemistry, particularly in process de-velopment. Traditional trial-and-error methods for reaction optimization are increasingly inadequate due to high time and resource consumption. This study presents the development of an AI-driven model for predicting reaction yields in mi-cellar catalysis, leveraging representation learning and predictive analytics to reduce waste, and promote sustainable micellar methodologies. Despite the challenge of data scarcity, we trained the model based on limited available data for micellar catalysis and selected the closely related data from traditional organic solvents based on the design principle of surfactant PS-750-M and its intrinsic polarity match with organic solvents. The data set was manually curated from pa-tents and journals, ensuring robust model performance. The model employed a hybrid representation learning frame-work, integrating autoencoders for dimensionality reduction with a gradient-boosting regressor for prediction tasks. This approach demonstrated high predictive accuracy, with experimental validation showing yields closely resembling pre-dicted values. The findings highlight the potential of AI to transform synthetic micellar chemistry by enabling resource-efficient and environmentally sound amide couplings as a chemical transformation. This work lays a strong foundation for integrating advanced AI strategies into micellar catalysis, addressing current data limitations, and paving the way for future advancements in sustainable chemical design.

Item Type:	Article
Date Deposited:	19 Aug 2025 00:45
Last Modified:	19 Aug 2025 00:45
URI:	https://oak.novartis.com/id/eprint/57008