A modified conditional GAN(generative adversarial networks) architecture to generate pulses of DAB(driver side airbag) in drop tower impact tests has been newly developed and then a graph of GANs(generative adversarial networks) has also been developed to generate driver’s injury in crash scenarios. These neural networks are being directly used by designers through web-interface services. Pulses of airbag are obtained by drop tower impact test. Airbag is impacted with a freely dropped mass which is guided by a tower structure. Leakage parameter of numerical simulation model can be calibrated based on these pulses. Firstly, drop tower impact tests are done to calibrate leakage parameters of simulation models. Using calibrated simulation models, training time series data of 1,590 cases had been obtained. For injury AI model, driver’s injury test data in frontal and side impact crash had been used in calibrating corresponding model parameters. Using the calibrated injury simulation models, training data of more than 5,000 crash cases had been obtained. For approximating airbag pulses, at first, RNN(Recurrent Neural Network) with embedding table were used. The embedding table is obtained based on knowhow of domain expert. But, as the number of considering cases increases, the table should be replaced by a fully connected embedding layer. Furthermore, to consider latent variational random properties of airbag components, GAN architecture had been introduced. To meet the design specification, for example, the type of inflator, fabric, etc. Conditional GAN is used. Design specifications are applied as conditions of probability distributions of generator and discriminator. In addition to that, multi-head attention mechanism is also adopted to increase accuracy. Mean squared error of the trained model are decreased from 12.9 to 0.084 and correlation coefficient is increased by 5.4%. To improve GAN results, mean square error loss term is adopted in addition to the adversarial loss of standard GAN. With the modified loss function, enough accuracy is obtained. Driver’s injury in US-NCAP(New Car Assessment Programme) frontal crash, FMVSS-208 unbelted crash, IIHS-Offset Deformable Barrier and side impact NCAP have been considered. To transform the drop tower impact GAN architecture to injury model, conditions for safety restraints systems, vehicle information and crash pulse are added as conditions. Especially, to consider crash pulse more accurately, one more multi-head attention mechanism layer is added between the crash pulse input and the final layer output of the drop tower impact GAN model. GAN models are organized as a directed edge graph to consider injuries at the interested parts of dummy. The graph is defined by structural scenarios based on the injury researcher’s knowledge. At first, nodes of GAN unit are defined for all the required dummy responses respectively. According to the model validation procedure of injury simulation, each node is connected to the other relevant nodes with directed edge. Developed AI models are used directly by designers through web-interface services in the company. Because learned AI model requires small amount of computing power for generating new results, a general personal computer can be used as a server. A web application program is used to call AI models and the server system is maintained by a python networking library. Our developed networks are very useful for predicting simulation results in the limited design space. In addition to more data, applying some useful architectures of cycleGAN and diffusion model, our developed models will continuously be improved.