A GPU Accelerated Algorithm For Solving Inverse Problems In Atmospheric Dispersion

With the development of the economics, the chemical industries and related fields are blossom. At the same time, the frequency of poisonous gas dispersion increases. As the disasters of this kind always happen unexpected and suddenly, it would cause a great loss of life and property, if not handle in time. So the related department should take a quick response to the emergency to minimize the lost. First of all, people in the threatened areas should be evacuated according to the principles of Humanitarian Relief Efforts. And the demarcation of the threatened area becomes very important. If the area is too large, a lot of unaffected people would be dispersed. While if it is too small, people in the unsafe area would be injured. As a consequence of that, find out the leak source strength is a very import issue. And as the poisonous gas density and the atmospheric related data can be obtained by sensors, it offers us a way to calculated the leak source strength with inverse method. But due to large scale data and the computation complexity, it computing time is often too long to meet the real time requirement. It can be concluded that, find out the leak source strength accurate and timely is of great importance.We make the prediction of gas density with the Gaussian plume diffusion model, and compare the result with the data obtained by sensors. And the aim of the optimization is the variance of them. So the inverse problem is reversed to be a combinatorial optimization problem. Considering convergence speed traditional serial Genetic Algorithm base on CPU for solving the large-scale problem is too slow, we immigrate the algorithm into GPU and made some adjustment, owning to its floating-point calculations ability and high parallel architecture. In our algorithm, a coarse-grain parallel model is employed, Each thread is associated with the genetic operation for each chromosome and each block deals with an island, In addition, exchange of information takes place among individuals and among populations to make sure the quality of whole population is steadily increasing.According to the data test, the GPU accelerated genetic algorithm can find the solution in a very short time, the speed-up ratio can be 10 or more, and the accuracy is also improved. It can be concluded that the modified algorithm performs well.The GPU accelerated genetic algorithm proposed in our paper can solve the inverse problems in atmospheric dispersion efficiently and effectively. By comparing the obtained dispersion density data and predicted the gas dispersion density, we can find the leak source strength. And it can work as a good decision support for the emergency management.