Design of fuel molecules based on variational autoencoder. (15th May 2022)

• Record Type: Journal Article
• Authors: Liu, Ruichen; Liu, Runze; Liu, Yifan; Wang, Li; Zhang, Xiangwen; Li, Guozhu
• Is Part Of: Fuel. Volume 316(2022)
• Publication Date: 2022-05-15
• Subjects: Variational autoencode -- Molecular design; Fuel -- Periodicals; Coal -- Periodicals; Coal; Fuel; Periodicals; 662.6
• DOI: 10.1016/j.fuel.2022.123426
• Languages: English
• ISSNs: 0016-2361

Abstract

Graphical abstract: Highlights: A variational autoencoder is established to generate new hydrocarbon molecules. A multilayer perceptron is jointly trained to predict fuel properties quickly. A fuel database containing 11, 291, 051 hydrocarbon molecules has been established. Multi-objective optimization of fuel molecules is realized based on a new method. Abstract: A variational autoencoders (VAE) has been developed to reversibly represent hydrocarbon molecules in the mathematic form of continuous multidimensional vector for fuel design. The characteristics of a hydrocarbon molecule can be efficiently captured to establish the latent space of hydrocarbon fuels by the VAE. The physicochemical properties of a hydrocarbon fuel were automatically calculated by a piece of python code based on the group contribution method, and the energy properties were quickly and accurately predicted by a multilayer perceptron jointly trained with the VAE. Handily manipulating of the fuel molecules have been realized in the latent space, such as sampling new structures, decoding random vectors, perturbing a given fuel structure, and interpolating between fuel molecules. A large database (CH-02) containing 11, 291, 051 hydrocarbon molecules and their fuel properties have been established. Moreover, an approximate-concentric-sphere method was developed to realize multi-objective optimization of a given fuel molecule. This work opens new avenues in the de novo design of new hydrocarbon fuels to meet the requirements of next generation engine.