Machine learning assisted evaluations in structural design and construction

This paper proposes a new design approach based on an iterative machine learning algorithm to speed up the topological design exploration of compression-only shell structures with planar faces, considering both structural performance and construction constraints. In this paper, we show that building neural networks allows one to train a surrogate model to accelerate the structural performance assessment of various possible structural forms without going through a significantly slower process of geometric form-finding. The geometric form-finding methods of 3D graphic statics are used as the primary structural design tool to generate a single-layer, compression-only shell with planar faces. Subdividing the force diagram and its polyhedral cells using various rules results in a variety of topologically different compression-only structures with different load-bearing capacities for the same boundary conditions. The solution space for all possible compression-only forms for a given boundary condition is vast, which makes iterating through all forms to find the ideal solutions practically impossible. After training with an iterative active sampling method, the surrogate model can evaluate the input data, including the subdivision rules, and predict the value of the structural performance and the construction constraints of the planar faces within milliseconds. As a result, one can then evaluate the nonlinear relations among all the subdivision rules and the chosen structural performance measures, and then, visualize the entire solution space. Consequently, multiple solutions with customized thresholds of the evaluation criteria are found that show the strength of this method of form-finding in generating design solutions. Besides, considering the total training time of the neural network model, the proposed framework is still faster than a traditional optimization method, such as the genetic algorithm that can find only the optimum values. This process will result in interactive sampling methods in which the machine learning models assist the designer in choosing and controlling different design strategies by providing real-time feedback on the effects of the selected parameters on the design outputs.