AI-Assisted AM for dimensionally accurate ceramic vat-photopolymerisation

Ceramic fabrication via bottom-up vat-photopolymerisation additive manufacturing (AM) offers the inherent advantages of AM technologies, including high geometric precision and the capacity to realise complex architectures. Nonetheless, this approach introduces unique challenges that demand in-depth investigation. This PhD research addresses the critical barriers associated with ceramic vat-photopolymerisation and proposes strategies, enhanced by artificial intelligence (AI) algorithms, for the production of dimensionally accurate bulk ceramic components. The research is structured in three stages, each focusing on a key step in the ceramic vat-photopolymerisation process. The first stage is developing a binder formulation; the second examines suspension stability and 2D dimensional accuracy; while the third investigates post-processing shrinkage, with emphasis on the effects of geometry and print parameters. In each stage, experimental datasets were generated and supervised machine learning (ML) algorithms were employed to predict key material or process properties. Each phase of the research was systematically designed and executed with the aim of producing peer-reviewed journal publications. The primary technical challenge lies in the development of a photocurable resin compatible with ceramic suspensions that combines low viscosity with sufficient mechanical strength. As such, the careful selection and optimisation of resin constituents are critical. A full factorial design of experiments was employed to assess the influence of three different monomers/oligomers, each at varying concentrations, on selected binder properties. Simultaneously, supervised ML models like support vector machines, decision trees, and artificial neural networks were developed to predict essential binder characteristics, including binder viscosity and the ultimate tensile strength (UTS) of green bodies. The photocurable ceramic suspension comprises the optimised binder, ceramic powder, and various functional additives. Dispersants play a crucial role in covering ceramic particles to prevent sedimentation and agglomeration during printing. Effective particle dispersion also contributes to a reduction in overall slurry viscosity. Six dispersants were evaluated at different concentrations to determine their effects on suspension rheology. 3 mol% yttria-stabilised zirconia (YSZ) was selected as the model ceramic material, despite its high refractive index posing challenges for UV-based curing due to increased light scattering. This scattering often results in overcuring, thereby compromising dimensional accuracy. To address this, optimisation of ceramic particle morphology, formulation composition, and print parameters was undertaken, along with the addition of UV absorbers to minimise overcuring. Moreover, ML based models were developed to predict 2D dimensional errors in ceramic green parts, enhancing both understanding and mitigation of lateral inaccuracies. Green bodies produced via vat-photopolymerisation consist of a polymer matrix encapsulating ceramic particles. To obtain dense, functional ceramics, post-processing via debinding and sintering is required. These processes induce shrinkage, significantly affecting 3D geometric accuracy. The influence of printing and geometric parameters on shrinkage behaviour was systematically investigated. Variables such as infill type and density, two layer thicknesses, and three build orientations were studied to understand their role in shrinkage dynamics. While traditional approaches to shrinkage analysis and estimations are grounded in sintering theory and rely on material and process parameters, this study employed supervised ML models to predict shrinkage factors based on selected input variables, offering a data-driven alternative for enhanced dimensional control. In conclusion, this PhD research demonstrates the critical role of data-driven modelling in advancing ceramic additive manufacturing via vat-photopolymerisation. By integrating experimental data with supervised machine learning algorithms, the work successfully predicted key properties across the entire production chain, from binder formulation and slurry stability to dimensional accuracy and sintering shrinkage. Traditional empirical and theoretical approaches often fall short in capturing the complex, multivariable interactions inherent in AM processes. In contrast, data-driven models enable the estimation by material and process parameters with greater accuracy.