A Comparative Study Between a Finetuned Conditional Generative Adversarial Network (cGAN) and the A* Algorithm for BIM Clash Coordination

Building Information Modelling (BIM) is a widely used technology in the Architecture, Engineering, and Construction (AEC) industry for integrating non-geometrical and geometrical information of physical buildings and assets. However, the current BIM technology poses a challenge when detecting and resolving clashes between models where it requires manual labor, and it is time-consuming. Hence, this project investigates the use of Generative Artificial Intelligence (Gen AI) to automate the BIM clash resolution process and benchmark its performance against the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> with dynamic cost function algorithm. To implement GenAI in BIM clash resolution, we fine-tune conditional Generative Adversarial Network (cGAN) model using a synthetic image dataset of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2, 0 0 0}$</tex> images. The performance of the cGAN model is compared with a benchmark <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> algorithm that is implemented with a dynamic cost function that can set varying cost. A suite of test cases based on real-world BIM scenarios are used to evaluate the performance of the cGAN model and the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> algorithm with dynamic cost. The results demonstrated that the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> algorithm outperforms the cGAN model with 32.91 % similarity to the optimal solution while the cGAN model performs significantly worse than <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> algorithm to determine a viable clash-free path, achieving only 4.78 % similarity to the optimal solution, providing unclear and impractical paths within the solution. In contrast, the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> algorithm with dynamic cost function consistently determined a viable clash-free path. Moreover, the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$A^{*}$</tex> algorithm achieved the highest similarity of 61.7 % recorded in low obstacle density scenarios at the expense of high computation load.