Surrogate modeling of heat transfer in 3D-printed facades with active learning

Recently, large-scale 3D printing has emerged as an alternative manufacturing technique for novel facade components aiming at high operational efficiency and low environmental impact. Previous research has demonstrated that cavity geometry, size, and arrangement significantly influence the insulation properties of 3D-printed (3DP) components. Estimating heat transfer effects within cavities typically requires computationally expensive computational fluid dynamics (CFD) and finite element (FE) simulations or experimental campaigns.

To leverage computational design and digital fabrication, efficient methods must be established to guide the performance-informed design of this new facade type. Embedding thermal performance evaluations in the design workflow is essential to enable efficient iterative design without resorting to extensive simulations or fabricating multiple designs.

This study aims to develop an AI-based surrogate model to accurately describe heat transfer in polymer 3DP structures, considering heat conduction, cavity convection, and radiation. The model relates geometrical and material features of 3DP facades, as well as boundary conditions, to their thermal performance. This approach enables rapid comparisons of different design alternatives and exploration of site-specific designs under various climatic conditions.

The model is trained on a hybrid dataset comprising experimental data from hot-box measurement campaigns on 3DP samples and simulation results from validated CFD-FE models. The experimental data includes temperature and heat flux measurements from physical prototypes under different geometrical configurations, while the simulation data includes additional geometrical variations and wider boundary condition variations. Various AI techniques, from traditional machine learning (ML) to neural networks (NN), are implemented and compared for prediction accuracy, computational efficiency, robustness, and generalization potential.

We expect both ML techniques and NN models to provide accurate predictions of thermal performance. NNs are anticipated to capture complex, non-linear relationships within the data, with ML offering a good balance between accuracy and computational cost, provided features are well-engineered. Models trained on hybrid datasets are expected to demonstrate robustness to varying boundary conditions and generalization to unseen geometric configurations. Interpretability techniques will provide valuable insights into the impact of different design parameters.

This modeling approach offers a quick and efficient alternative to traditional iterative design, significantly reducing the time and computational resources required for thermal performance evaluation and providing immediate feedback. This is particularly important in early-stage design for fast exploration of the design space prior to refining and increasing the resolution of evaluations in detailed design phases.