Property-Oriented Reverse Design of Hydrocarbon Fuels Based on c-infoGAN.

Area of Science:

• Computational chemistry
• Materials science
• Chemical engineering

Background:

• Traditional fuel design relies on a forward approach, screening existing molecules.
• The vast chemical space of organic molecules necessitates a reverse design strategy.
• Predicting structure-property relationships for fuels remains a significant challenge.

Purpose of the Study:

• To develop a reverse design methodology for hydrocarbon fuels using deep generative models.
• To generate novel fuel molecules with specific target properties.
• To validate the efficacy of the developed models in discovering high-performance fuels.

Main Methods:

• Implementation of conditional generative adversarial networks (c-GANs), specifically c-GAN and c-infoGAN, for molecular generation.
• Training generative models on hydrocarbon molecules with target fuel properties as input.
• Analysis of generated molecules for validity, uniqueness, novelty, and property alignment.
• Experimental synthesis and testing of a designed fuel molecule.

Main Results:

• c-infoGAN demonstrated superior performance in generating valid, unique, and novel hydrocarbon molecules.
• The model successfully rediscovered JP-10 and generated 27 new fuel candidates with desirable properties.
• The latent space constructed by c-infoGAN exhibited an ordered structure, facilitating property-guided design.
• Experimental validation confirmed the robust design capability of the c-infoGAN model.

Conclusions:

• Conditional generative adversarial networks provide a powerful tool for the reverse design of hydrocarbon fuels.
• This approach enables the discovery of novel molecules tailored to specific performance requirements.
• The study opens new avenues for designing advanced fuels for next-generation engines.