Machine Learning Assisted Selection of Catalyst for\nγ‑Valerolactone Hydrogenation from Levulinic Acid

Abstract

The\npreparation of γ-valerolactone by levulinic acid (LA)\nhydrogenation is green, efficient, and economical, but the traditional\ncatalyst selection and optimization methods have low efficiency and\nhigh cost, which cannot meet the development needs of the chemical\nindustry. To this end, we conducted new research to develop a machine\nlearning framework to predict LA conversion and γ-valerolactone\nyield, accelerating catalyst selection and optimization. Through <i>K</i>-means clustering preliminary classification data sets,\ncomposite minority oversampling technology and adaptive composite\nsampling resampling unbalanced data sets were used to solve the problem\nof small sample size and improve classification performance. The performance\nof four machine learning optimization algorithms was evaluated, and\nthe superior performance of the support vector machine was found to\nbe the core of the model. We not only pursue prediction accuracy but\nalso find that reaction temperature was the main influencing factor\nthrough the shapley additive explanation. The most potential catalyst\nRu/N@CNTs was selected based on the feature importance analysis results\ncombined with the genetic algorithm multiobjective optimization model.