Polyurethane (PU) foam is a highly versatile plastic with excellent thermal insulation properties. Conventional PU relies on polyols and polyisocyanates from non-renewable sources, raising environmental and sustainability concerns. Emerging PU formulations with biobased content aim to reduce fossil dependency and advance global decarbonization and sustainability goals. Owing to its chemical structure and functionality, lignin, the second most abundant biopolymer, stands as a promising alternative to polyols in PU foams. In this work, we present a methodology for predicting mechanical properties of lignin-containing PU foams from imaging, and inception of new foams through desired mechanical properties, by solving the inverse problem to design formulations with target properties. To ensure generation soundness, the used Rank Reduction Autoencoder (RRAE) combines a projection-based model order reduction technique with machine learning technology. A truncated latent space enables high regularization and avoids overfitting [1,2]. The proposed RRAE is a therefore a kernel-free kPCA approach combined with machine learning autoencoders. The method enables identification of mechanical properties from microstructure, generation of microstructures from desired properties, and design of bio-based formulations, addressing the complexity of non-unique and mixed-component materials. The results showcase the ability of the RRAE neural network to integrate material knowledge from microstructure imaging, along with material inputs to predict mechanical properties. Moreover, it is able to infer the material components from the desired functional properties. Training and validation analysis compared to ablation studies testify of the RRAE ability to outperform classical technologies.