Electrolyte additives are crucial for accelerating the commercialization of lithium metal batteries (LMBs), yet designing effective additives is challenging due to the need to balance conflicting properties, such as eectrochemical performance and nonflammability. To address this challenge, a deep learning-assisted generative model is developed for multiobjective optimization of electrolyte additives. Overcoming data scarcity, the dataset is expanded using a molecular categorization derivation method, increasing single-property data points to 70 095 multiproperty data points. Coupled with an asynchronous limited decoder and adversarial regulation strategy for latent distribution, this approach achieved 100% generative efficiency for structurally complex and diverse molecules in vast chemical space. The method is validated by discovering 2,4-bis(2-fluoroethoxy) tetrafluorocyclotriphosphazene (DFEPN), a novel additive with excellent flame resistance and stable dual electrode/electrolyte interphases. In a Li||LiFePO4 full cell with a commercial electrolyte, DFEPN enables an order of magnitude increase in capacity retention, outperforming the state-of-the-art flame-retardant additive ethoxy(pentafluoro)cyclotriphosphazene by 33%. This study offers a pathway for developing safe and reliable lithium battery electrolytes, particularly under severe data constraints, and has broader implications for advanced battery design.
Keywords: artificial intelligence; electrolyte; generative model; lithium battery; molecular design.
© 2025 The Author(s). Advanced Materials published by Wiley‐VCH GmbH.