Data-driven bio-integrated design method encoded by biocomputational real-time feedback loop and deep semi-supervised learning (DSSL)

Created by W.Langdon from gp-bibliography.bib Revision:1.8414

@Article{Heidari:2024:jobe,

• author = "Farahbod Heidari and Mohammadjavad Mahdavinejad and Katia Zolotovsky and Mohammadreza Bemanian",
• title = "Data-driven bio-integrated design method encoded by biocomputational real-time feedback loop and deep semi-supervised learning ({DSSL)}",
• journal = "Journal of Building Engineering",
• year = "2024",
• volume = "98",
• pages = "110923",
• keywords = "genetic algorithms, genetic programming, Data-driven bio-integrated design, Deep semi-supervised learning, transfer learning, feature extraction, Programmable living building materials",
• ISSN = "2352-7102",
• URL = "https://www.sciencedirect.com/science/article/pii/S2352710224024914",
• DOI = "doi:10.1016/j.jobe.2024.110923",
• abstract = "Today, under the imperatives of synthetic biology (Synbio), material-based design strategies are conceived as biofactories that can recapitulate the functionalities of living systems for developing building materials with controllable structural features. These strategies include biofabrication techniques such as vitro-culturing, Fabrication Information Modeling (FIM), growth parametrization, machinery systems, and genetic programming articulated bioproducts as the construction material for low-carbon buildings. However, these biophysical systems need interactive development by data-oriented models and decision-maker tools that can mine, measure, and re-configure the complexity in biological systems through monitoring and controlling in a higher integration toward systemic decision-making solutions. This research presents a real-time feedback loop as the bio-integrated design method for design and fabrication of Gluconacebacter Xylinus (GX) Bacterial Cellulose (BC) with Transfer Learning (TL) and Deep Semi-Supervised Learning (DSSL) through TensorFlow and Keras Python libraries. The loop started by mining datasets from BC growth via the VGG16 pre-trained model. In this stage, the loop can learn the material behaviour over the growth process and attribute features to a measured BC thickness. Based on this attribution, the loop can drive real-time classifying of input datasets to thickness-based classes and guide the Region of Interests (ROIs) to a user-designed thickness till the captured features are fitted to the user-input class features. The training results over 8640 images show a 0.001 learning rate, an average loss of 0.18 (540 total steps), an accuracy of 0.706, and a precision of 0.752 for the feature extractor model. For the classifier, a 0.15 average loss (3375 total steps), an accuracy of 0.829, and a precision of 0.712 demonstrate the models' efficiency in pattern recognition and classification. Also, the loop operation persistently exhibits a more than 50 percent precision rate, displaying the loop's high performance in achieving user-designed 3D shapes. This research aims to a transformative shift in bio-integrated design based on feedback loops as the foundational step towards intelligent bio-digital design platforms. Also, this ground can break designing with biological systems based on desired tasks for the new generation of designers called biodesigners",

}

Genetic Programming entries for Farahbod Heidari Mohammadjavad Mahdavinejad Katia Zolotovsky Mohammadreza Bemanian

Citations