<p>Cellulose-based materials are promising sustainable alternatives to petroleum-derived plastics, but their development is hindered by inherent flammability, brittleness, and inefficient trial-and-error design. Herein, an expert-guided active learning framework integrating machine learning and experiments was developed to accelerate the design of flame-retardant cellulose nanofiber (CNF)/polyvinyl alcohol (PVA) nanocomposite films. Leveraging literatures, experiments, and data augmentation, a 628-sample database was constructed, wherein synergistic modification of flame-retardant elements boron (B), nitrogen (N), phosphorus (P), and silicon (Si) on CNF were effectively realized via experiments. After 11 expert-guided active learning iterations, the Extreme Gradient Boosting (XGBoost) model demonstrated superior predictive accuracy due to its strong capability in handling heterogeneous datasets and capturing nonlinear relationships. The test and validation sets of XGBoost model achieved coefficients of determination (R²) of 0.96 and 0.97 for limiting oxygen index (LOI), and 0.90 and 0.93 for tensile strength (<i>T</i>_<i>s</i>), respectively. Guided by model predictions, the optimized candidate (SCNF₂₃/PVA) achieved an LOI of 33.6%, <i>T</i>_S of 83&#xa0;MPa, confirming satisfied flame retardancy and mechanical properties. The film also showed UV-shielding capability, biodegradability of 10 days and potential as a flame-retardant coating for wood and polyurethane foams. Combining domain knowledge, the flame-retardant mechanism of SCNF₂₃/PVA film from the synergizing among B, N, P, Si has been understood assisted by Shapley Additive Explanations (SHAP) and Pearson feature correlation matrix (Pearson) analysis. This work establishes a data-driven, interpretable, and experimentally validated strategy for intelligent discoveries of sustainable nanocomposites.</p> Graphical Abstract <p></p>