aiqbench analysis
This research introduces a Co-optimized Variational Autoencoder (Co-VAE) framework to solve the inverse molecular design problem for high-octane fuels by directly optimizing latent representations against specific property targets. By integrating quantitative structure-property relationships with differential evolution, the method efficiently navigates vast chemical spaces to generate novel fuel candidates that meet performance thresholds before synthesis attempts are made. The approach demonstrates a practical application of generative deep learning in accelerating materials discovery through automated exploration of complex molecular structures rather than relying on traditional trial-and-error or purely predictive modeling.
A Generative Deep Learning Workflow for Inverse Molecular Design of Fuels
arXiv:2504.12075v4 Announce Type: replace Abstract: In the present work, a generative deep learning framework combining a Co-optimized Variational Autoencoder (Co-VAE) with quantitative structure-property relationship (QSPR) techniques is developed to enable inverse molecular design of fuels. The Co-VAE approach integrates an auxiliary fuel property prediction regression head with the VAE latent space, enhancing molecular reconstruction and accurate property estimation (Research Octane Number (RON) chosen as the fuel property of interest for demonstration studies). A subset of the GDB-13 database, combined with a curated RON database, is used for the Co-VAE training. Hyperparameter tuning is further utilized to optimize the balance among reconstruction fidelity, chemical validity, and RON prediction. Subsequently, an independent regression model is trained to further improve RON prediction accuracy, and a differential evolution algorithm is employed to efficiently navigate the Co-VAE latent space and identify promising fuel molecule candidates with RON greater than a chosen threshold. The overall generative deep learning framework captures complex structure-property relationships within a latent representation, and can be readily extended to different or multiple fuel properties, allowing exploration of large chemical spaces relevant to fuel design. Furthermore, the framework can be further augmented by incorporating additional synthesizability criteria to improve applicability and reliability for de novo design of novel high-performance fuels.
Abstract:In the present work, a generative deep learning framework combining a Co-optimized Variational Autoencoder (Co-VAE) with quantitative structure-property relationship (QSPR) techniques is developed to enable inverse molecular design of fuels. The Co-VAE approach integrates an auxiliary fuel property prediction regression head with the VAE latent space, enhancing molecular reconstruction and accurate property estimation (Research Octane Number (RON) chosen as the fuel property of interest for demonstration studies). A subset of the GDB-13 database, combined with a curated RON database, is used for the Co-VAE training. Hyperparameter tuning is further utilized to optimize the balance among reconstruction fidelity, chemical validity, and RON prediction. Subsequently, an independent regression model is trained to further improve RON prediction accuracy, and a differential evolution algorithm is employed to efficiently navigate the Co-VAE latent space and identify promising fuel molecule candidates with RON greater than a chosen threshold. The overall generative deep learning framework captures complex structure-property relationships within a latent representation, and can be readily extended to different or multiple fuel properties, allowing exploration of large chemical spaces relevant to fuel design. Furthermore, the framework can be further augmented by incorporating additional synthesizability criteria to improve applicability and reliability for de novo design of novel high-performance fuels.
Machine Learning (cs.LG); Chemical Physics (physics.chem-ph)
arXiv:2504.12075 [cs.LG]
(or arXiv:2504.12075v4 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2504.12075 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 297 words
