Computational design of surface grafted molecules that maximise the hydrophobicity of cellulose surface, as quantified via the water contact angle (CA), is a high-dimensional optimisation problem. Although substantial experimental water CA data are available for grafted surfaces in general, there is currently limited understanding of how the structures of these grafted molecules translates into a specific CA value. Molecular dynamics (MD) simulations can provide insights into these structure-CA relationships, but the corresponding dataset is relatively lean. In this instance, exploration of graftant design using MD simulations alone is too expensive for broad screening, and for which application of machine learning (ML) alone lacks the physical grounding needed to reliably guide molecular discovery. To address this, here we introduce a human-machine teaming framework that integrates expert intuition, MD-validated contact angles and MD-generated insights, alongside latent-space Bayesian optimisation within a closed iterative loop. Operating effectively with a small dataset and only a limited number of iterations, the approach rapidly converges toward high-performance candidates. When applied to cellulose, the framework consistently identifies compact, densely branched aliphatic motifs, particularly those with limited heavy-atom counts, as strong hydrophobes with water contact angles exceeding 110°. Additional targeted perturbation studies further confirm the sensitivity of hydrophobicity to local branching architecture. This work demonstrates that human-machine teaming enables fast, data-efficient molecular optimisation and yields interpretable insights unattainable through MD or ML alone.