Answers are grounded in this paper and cite their sources.
Additive manufacturing (AM) offers significant advantages for producing complex geometries with high efficiency. However, inconsistent dimensional accuracy across different units of the same printer model remains a major barrier to AM’s adoption in industrial-scale production. While prior studies have successfully optimized print parameters for individual machines, relatively little work has examined reproducibility across multiple printer units. This study proposes a Deep Auto-Encoded Multi-Domain Gaussian Process framework to predict dimensional deviations in stereolithography-based AM, with a focus on cross-printer reproducibility. The method integrates an autoencoder with inspired by Gaussian Process regression to learn a data-driven, nonstationary kernel that captures nonlinear interactions among process parameters, dimensional deviations, and machine-specific behavior. Unlike traditional fixed-kernel or single-domain approaches, the learned kernel enables adaptive modeling of similarity across multiple printer units. The framework is evaluated through a case study involving thirty prints of a geometrically complex part produced across three SLA printer units of the same model type under varying process parameters. An additional case study was conducted using simulated printer data across ten printers to assess the model’s consistency when applied to a larger dataset. Results show that the proposed DAE-MDGP achieves the lowest prediction error across multiple metrics, outperforming both its individual components and conventional machine learning models. Even small improvements are consequential in this setting, where data scarcity and process noise constrain predictive performance. Overall, this work demonstrates that learned multi-domain kernels can improve predictions across printer units, supporting better parameter selection and more reliable AM deployment in production environments.
No takes yet. Share an insight, caveat, or question.
Heck et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: