Machine learning-based crashworthiness optimization for the square cone energy-absorbing structure of the subway vehicle

This paper presents a novel framework for predicting the crashworthiness of a square cone energy-absorbing (SCEA) structure using a machine-learning method. The structure consists of an anti-creep, a thin-walled structure with nonuniform thickness, diaphragms, two types of aluminum honeycombs and a guide rail. The finite element model of SCEA structure was established and validated by full-scale experimental test. Taking the thicknesses of thin walls (T-A and T-B) and diaphragms (T-gb), strengths of honeycombs (& delta;(A) and & delta;(B)) as parametric variables, the parameters of SCEA structure were changed based on a virtual design of experiments (DOE) to generate training data and test data. To improve the crashworthiness of SCEA structure, the structural parameters were employed as input data, four machine learning models were utilized to predict the energy-absorbing characteristic curve of the SCEA structure, and the prediction accuracy of different models was compared and analyzed. According to the results of comparison, the Gate Recurrent Unit (GRU) model was chosen to predict the structural energy-absorbing characteristics, also employed as the input of optimization. The energy absorption (EA) and initial peak crushing force (PCF) were adopted as objectives, and the global response surface method (GRSM) was employed as the optimization algorithm. The results showed that the optimal solution was obtained as PCF = 618.41 kN and EA = 297.99 kJ when T-A = 2.1 mm, T-B = 2.9 mm, T-gb = 2.4 mm, & delta;(A) = 5.99 MPa and & delta;(B) = 4.82 MPa. The machine learning method offers engineers and scientists a potential tool to accelerate the design and optimization of SCEA structures for rail vehicles.