Dataset for Disentangled Generative AI-Guided Closed-Loop Optimization of Deposition Morphology for 3D Bioprinting Applications

The dataset supports the findings presented in the paper "Generative AI-guided in silico closed-loop optimisation of deposition morphology for 3D bioprinting applications." Reference Colin Zhang, Kelum Elvitigala, and Shinji Sakai. Generative AI-guided in silico closed-loop optimisation of deposition morphology for 3D bioprinting applications. Virtual and Physical Prototyping 21, e2671497 (2026). © 2026 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group, under CC BY 4.0. DOI: https://doi.org/10.1080/17452759.2026.2671497 Abstract Recent advances in novel bioinks have dramatically increased the feasibility and applicability of 3D bioprinting for tissue engineering and regenerative medicine. However, developing new bioinks still requires extensive trial-and-error testing due to bioink rheology, crosslinking reactions, printing parameters, and limited resources. Previous classification- or regression-based AI models for bioink optimisation are typically black-box and cannot provide visual results. To address these challenges, a state-of-the-art disentangled and explainable generative AI framework was developed. The framework comprises a beta-conditional variational autoencoder (β-CVAE) for generating novel, variational images of printed constructs based on ink properties and printing parameters. Furthermore, an in silico closed-loop Bayesian optimisation (BO) system coupled with a convolutional neural network (CNN) was employed to quantitatively predict pre-printing performance and classify defects. The trained β-CVAE model can generate realistic and condition-dependent images of printed constructs. Visualisation of the latent space revealed an interpretable organisation of the learned features, supporting the model’s explainability and controllability. Moreover, transfer learning was employed to rapidly adapt to new blueprint designs with limited data. Although this study focuses on acellular hydrogel printing, the bioink formulations and crosslinking conditions are cytocompatible and extensible to bioprinting applications. The proposed framework accelerates bioprinting optimisation through interpretable generative AI modelling. Keywords: 3D bioprinting; deposition morphology; machine learning; generative artificial intelligence; variational autoencoder File description "ALG-Ph_HA-Ph_rheology_data.zip" contains shear-rate-dependent viscosity, storage modulus, and loss modulus for alginate-phenol (ALG-Ph) and hyaluronic acid-phenol (HA-Ph) conjugated inks. "image_index.csv" contains tabular information on image ID, date, ink type with concentration, shape, irradiation intensity, and whether sodium persulfate (SPS) is present. Note: "completeness_sam" entry determines whether the construct is considered complete or incomplete; however, this entry contains missing data and is not used in this study. "images.zip" contains image files (resized to 800×800 pixels) for training β-CVAE models. The file names are formatted as "resized_IMG_{image ID}.png." The {image ID} can be located in "image_index.csv." The image files include four types of shapes: grid, ichou, star, and nose. "models.zip" contains trained SVR, CNN, and β-CVAE scalers and models used in this study. "csv_data_files_generation.zip" contains the necessary files (fitted rheological parameters and phenol contents) to run "generate_beta_CVAE_images.py" in the GitHub repository. The details on the fittings are presented in Section 5.2 of the manuscript. Code availability statementThe scripts used for data analysis and machine learning models in this study are available on GitHub at: https://github.com/KORINZ/generative-ai-bioprinting-framework.