An automatic visible explainer of geometric knowledge for aeroshape design optimization based on SHAP

Abstract

Aerodynamic shape design may involve complicated geometric optimization problems. When the number of geometric decision variables increases, the optimization algorithms can still search for solutions automatically, but it would work more like a black-box, for which the explanation will be of great help. This paper focuses on further mining the explanatory information of the shape optimization. A multi-objective optimization of a spacecraft is performed as a simplified case study. The LightGBM meta-model is built for the response from the geometric variables to the objectives, followed the SHAP explainer to quantify the contributions of variables. To visually interpret the deformation on shape, we propose the SHAP-Shape Map method to translate the attribution from the deformation features to the shape and make the attribution visible on the mesh surface. Much explanatory knowledge is easily obtained, for example, without seeing any flow field or pressure distribution image, the explainer successfully inferred that the significant influence of the drag coefficient occurs at the tip of the shape's nose. Furthermore, it has almost no barrier to learn that the trade-offs (conflicts) between the objectives are mainly concentrated in the nose as well. More importantly, with the proposed method, the above analysis results can be reached by anyone without specialized expertise of aerodynamics or geometric parameterization.