WGAN-based molecular clumps generation

Abstract

In this paper, a new molecular clumps generation algorithm based on the Wasserstein Generative Adversarial Network (WGAN) is investigated for the problem that the simulated molecular clumps generated by the traditional Gaussian model have a large difference from the measured clumps. The method first generates position–position intensity map of the clump using WGAN and then performs 3D expansion in the direction of the velocity axis using a Gaussian function containing weak noise, which ultimately generates 3D simulated molecular clumps in position–position–velocity (PPV) space. Semi-supervised molecular clumps validation algorithm (SS-3D-Clump) is utilized to validate the 3D molecular clumps data generated by the method, and the confidence level of them are above 0.999. Experimentally, it is proved that the new method can provide simulated clumps with morphology and intensity closer to the measured clumps. The simulated molecular clumps are used on the one hand to generate PPV data in combination with measured backgrounds for testing the performance of molecular clump detection algorithms; on the other hand, they can also be used to supplement the 3D molecular clump samples. Experiments have demonstrated that, if 30%–50% of the measured positive samples are retained in the VoxNet training set, and the rest are simulated clumps generated, the classification accuracy can also be guaranteed.