Extended Data Fig. 5: Machine learning-driven optimization and experimental validation in three consecutive rounds.

(a) Adhesive strength (\({F}_{a}\)) of hydrogels fabricated in experiments according to the formulations proposed by GP_KB and RFR-GP models. (b) Variations in performance metrics, including: (i) successful rate (SR), defined as the fraction of the test set with higher true \({F}_{a}\) than the training set; (ii) ratio of maximum true \({F}_{a}\) between the test and training sets; and (iii) root mean squared errors (RMSEs) of the test sets. The success rate and \({F}_{a}\) ratio decrease from the first round to the second round and level off in the third round, implying convergence toward the global optimum via SMBO. Meanwhile, the RMSE decreases continuously over the three rounds, indicating that expanding the training dataset improves the accuracy of regression models. (c) Parity plots comparing ML predicted \({F}_{a}\) versus true \({F}_{a}\).