Dispersed Generative Inverse Design: Localized Uniformity Accelerating Polymer Property-to-Processing Discovery

Generative inverse design is key to accelerating polymer processing discovery with prescribed properties. However, when trained on small experimental data, the generative model generally exhibits limited prediction accuracy. Here, relying on iterative learning from experimental validations, we introduce a dispersed strategy to minimize the experimental cost toward high-precision generative inverse design, as exemplified by prescribing the toughness and strength behaviors of injection-molded polypropylene. The dispersed model accelerates polymer inverse design via a combined strategy of ensemble prediction, design space localization, and sampling uniformity. Notably, compared to a standard ensemble model requiring 30 iterative validations, the dispersed model shows a significantly reduced experimental cost of 1/3 validations to achieve the same high precision. This acceleration arises from the localized uniformity pattern of experimental data, i.e., the dispersed distribution in a local design space, which maximizes the reconstruction accuracy of the polymer property landscape in the local target area. Overall, this work establishes the state-of-the-art efficient generative inverse design platform for polymer processing and holds the promise of accelerating various materials’ inverse design.