Efficient injury risk predictions for a diverse population using parametric human modeling and inducing points in...

The objective of this study is to use parametric human modeling and machine learning methods to identify representative occupants that can account for injury variations among a more diverse population with a limited simulation budget.

A maximal projection method was used to sample 100 occupants, considering the variations in stature, weight, and sitting height. An automated mesh morphing method was used to morph the THUMS v4.1 midsize male model into the target geometries. US-NCAP frontal crash simulations were conducted with morphed human models and validated vehicle/restraint models. Surrogate models based on the Gaussian Process (GP) method were trained to find inducing points (IP), here defined as a small number of representative occupants whose outcomes could be used to accurately estimate the variations in the injury risks and patterns throughout the population. Statistical analysis was conducted to validate the IPs' coverage of total variation by illustrating the IP distribution. Restraint optimization was performed at IPs to yield an enhanced restraint system. The method was validated through comparisons among the predicted injury risks under the optimal and baseline designs.

Only 20 IPs were needed to sufficient to properly represent the variations in the injury risks and patterns in the whole population with acceptable accuracy. Compared to the surrogate model built from 100 crash simulations, the IP-based surrogate models incurred only 0.4% and 1.8% errors in head injury risks for males and females, respectively. Regarding the injury risks on the chest and lower extremities, the IP-based surrogate models resulted in less than 0.1% and 0.5% errors for males and females, respectively. The FE simulations indicated that the optimal restraint system design reduced the injury risk by relatively 16% and 13% for male and female respectively when delta-V = 25 (mph), and 47% and 27% for male and female when delta-V = 35 (mph).

The study proposed a method to generate more accurate injury risk predictions for a more diverse population under a limited simulation budget. Simulations using IPs may enable restraint system optimization to be conducted more efficiently while reducing injury risks across a more diverse population.