Data-driven inverse parameter design for stable dimensional control in extrusion-based additive manufacturing

Accurate dimensional control is critical in extrusion-based additive manufacturing, especially with customized printing materials. Deviations from target geometries can violate tolerances, degrade functional performance, and complicate assembly. Achieving high precision is challenging due to nonlinear, many-to-one, and stability-constrained parameter-to-outcome relationships, particularly with high viscosity materials and customized formulations. Traditional optimization approaches such as response surface methods struggle with high dimensional mappings, often converging slowly and trading accuracy for reliability. This study applies established inverse-design components to a new Direct Ink Writing application. A β-Conditional Variational Autoencoder (β-CVAE) with latent-space gradient refinement and an integrated printability classifier is operationalized to predict printing parameters for target line widths on a customized platform. A forward neural network is trained on experimental data to map parameters to line width. Three inverse approaches are compared, namely a global-search baseline formed by Differential Evolution coupled with the forward model, a local-search baseline using forward-model gradient refinement, and a β-CVAE with integrated optimization. The proposed application achieves sub-second parameter proposals and tight dimensional control within typical ±0.02–0.05 mm qualification bands, with 95% confidence analysis supporting reliability across tested conditions. A structured latent space reduces evaluation cost relative to global search, and the classifier steers solutions toward feasible and stable operating zones. Physical validation on the printer platform showed close agreement with model predictions within the specified tolerance bands, demonstrating an effective and practical route to fast, reliable parameter selection in extrusion-based additive manufacturing.