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

Fuel design is usually "forward": candidate molecular structures are designed first, and then their properties are predicted for screening. Owing to the large latent space of organic molecules (1060 order), reverse design by giving target fuel properties is urgently needed. However, it is hardly realized due to the unknown complex rule of the structure-property relationship. In this work, reverse design of hydrocarbon fuels is realized based on the conditional generative adversarial network of hydrocarbon molecules. Two deep generative models, c-GAN and c-infoGAN, are established and trained for generating new candidate fuel molecules when target fuel properties are input. c-infoGAN exhibited superior generation ability in terms of the validity, uniqueness, and novelty of the as-generated molecules. JP-10, a classical hydrocarbon fuel, was rediscovered by c-infoGAN. The latent space of fuels constructed by c-infoGAN is ordered, as proved by linear interpolation and linear algebra in this high-dimensional space. Given the target of high density, low freezing point, high heating value, and large specific impulse, 27 new fuel molecules with novel structures, high diversity, and expecting properties were designed. One of the as-designed fuels was experimentally synthesized and tested, which verifies the robust design ability of c-infoGAN. This work opens new avenues for the design of new hydrocarbon fuels to meet the strict requirements of next-generation engines.