Masked GraphTransformer Multi-Source Fusion Model for Ceramic Sintering Shrinkage Prediction

Accurate prediction of ceramic sintering shrinkage is still challenging because powder descriptors, green-body data, thermal schedules, atmosphere measurements and dilatometry traces are recorded at different resolutions and often have missing fields. Develop a Masked Graph Transformer multi-source fusion model for estimating the final linear shrinkage and intermediate densification states from heterogeneous process evidence in this paper. Each specimen is represented as a process graph whose nodes describe powder morphology, binder and additive conditions, compaction variables, furnace segments, atmosphere observations, and sensor-derived curve tokens. A masked reconstruction problem is added, and graph attention and a physics-informed shrinkage head are used to learn relational dependencies while maintaining usability in the presence of one or more missing sources. Experiments are carried out on a curated dataset of 1,284 ceramic specimens that include alumina, zirconia-toughened alumina, yttria-stabilized zirconia and silicon nitride systems. The proposed model has a final-shrinkage MAE of 0.36 %, an RMSE of 0.51 %, and an R² of 0.941, and has reduced the MAE by 31.0% compared with the best gradient boosting baseline. Under 30% random feature masking, the error increased by only 0.09 %, and the best non-masked neural baseline increased by 0.28 %. Attention and ablation analysis show that green density, peak temperature, dwell time, powder size distribution and mid-stage dilatometry slope jointly determine the prediction. Based on the above results, graph-based masked fusion can be used to achieve robust shrinkage estimation for data-fragmented ceramic manufacturing.