CAFE-Net: predicting dimensional error in ceramic 3D printing using a novel ceramic-aware feature-enhanced network

Abstract

Ceramic 3D printing direct ink writing/robocasting has the promise of transformative capability to produce complex geometries in future structural, biomedical and energy applications; nevertheless, dimensional precision is a key challenge to industry implementation. Process-induced dimensional errors are a result of the nonlinear interaction between thermal gradients, rheological behaviour and material-specific shrinkage interactions which conventional empirical models do not sufficiently capture. This paper presents CAFE-Net (Ceramic-Aware Feature-Enhanced Network), a deep learning-based architecture that combines multi-head self-attention, material-conditioned gating, residual dense blocks with squeeze-and-excitation attention, and multi-scale feature fusion to estimate dimensional error in ceramic extrusion-based additive manufacturing. Experimental 3D printing datasets from Clay, Stoneware and Porcelain materials were assembled, and ten main process parameters were transformed into nine physics-inspired interaction features. CAFE-Net was compared with fourteen models including Ridge Regression, Lasso, ElasticNet, Support Vector Machine, Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, a standard four-layer MLP, and a tabular ResNet. Under a rigorous 10-fold cross-validation protocol, CAFE-Net achieved the highest and most consistent relative predictive performance among all sixteen models compared, despite the inherently high noise-to-signal ratio of the dataset limiting the absolute accuracy attainable by any model tested. Ablation studies validated the contribution of each architectural component, while 10-fold cross-validation demonstrated strong generalisation. SHapley Additive exPlanations analysis identified Thermal Gradient, Nozzle Temperature, and Extrusion Rate as the most influential features.