Machine learning-based prediction and optimization of performances of porous ZrO₂/Al₂O₃ ceramics: a few-shot learning approach

Porous ZrO₂/Al₂O₃ (ZTA) ceramic has shown great potential for fluoride removal from contaminated water owing to its tunable structure, high strength, and excellent chemical stability. The porosity, compressive strength, and fluoride adsorption rate of porous ZTA ceramics prepared by the direct foaming method were systematically studied. A machine learning framework integrating transfer learning, Bayesian optimization, and continuous model refinement was developed to predict and optimize performances of porous ceramics under small-sample conditions. It leverages prior knowledge and limited experimental data to simulate variable-performance relationships. Based on this model, it searches for optimal processing configurations and iteratively improves prediction accuracy. Under the optimal process parameters suggested by the model, the predicted performance (porosity of 91.51 %, compressive strength of 1.58 MPa, and fluoride adsorption rate of 71.71 %) shows strong agreement with experimental results (92.42 %, 1.49 MPa, and 72.49 %), confirming the predictive accuracy achieved through Bayesian optimization. Continuous learning further reduces the mean absolute percentage error of the three performances metrics from 1.02 %∼6.15 % to 0.87 %∼2.74 %, realizing more efficient and reliable material design under data-scarce conditions.