Leveraging machine learning for predicting human body model response in restraint design simulations

A First Step Toward a Family of Morphed Human Body Models Enabling Prediction of Population Injury Outcomes

The injury risk in a vehicle crash can depend on occupant specific factors. Virtual crash testing using finite element human body models (HBMs) to represent occupant variability can enable the development of vehicles with improved safety for all occupants. In this study, it was investigated how many HBMs of different sizes that are needed to represent a population crash outcome through a metamodel. Rib fracture risk was used as an example occupant injury outcome. Morphed HBMs representing variability in sex, height, and weight within defined population ranges were used to calculate population variability in rib fracture risk in a frontal and a side crash. Two regression methods, regularized linear regression with second-order terms and Gaussian process regression (GPR), were used to metamodel rib fracture risk due to occupant variability. By studying metamodel predictive performance as a function of training data, it was found that constructing GPR metamodels using 25 individuals of each sex appears sufficient to model the population rib fracture risk outcome in a general crash scenario. Further, by utilizing the known outcomes in the two crashes, an optimization method selected individuals representative for population outcomes across both crash scenarios. The optimization results showed that 5-7 individuals of each sex were sufficient to create predictive GPR metamodels. The optimization method can be extended for more crashes and vehicles, which can be used to identify a family of HBMs that are generally representative of population injury outcomes in future work.

Evaluation of AIS3+ car occupant injuries using deterministic and probabilistic methods in frontal crashes

Computational modelling was used to assess the capability of a deterministic and a probabilistic method to predict the incidence of AIS3+ injuries in passenger car occupants by comparing the predictions of the methods to the actual injuries observed in real-world crashes. The likelihood of sustaining an injury was first calculated using a computer model for a selected set of injury criteria in different impact conditions based on real-world crashes; AIS3+ injuries were then predicted using each method separately. Regardless of the method, the number of serious injuries was over-predicted. It was also noted that the used injury criteria suggested the occurrence of specific injuries that were not observed in the real world. Although both methods are susceptible to be adapted to improve their predictions, the question of the suitability of using some of the most commonly accepted injury criteria used with crash test dummies for injury assessment with human body models deserves further research.