A multi-geometry space self-adaptive visual language modeling system for astronomical simulation

Abstract

The present disclosure provides a multi-geometry space adaptive visual language modeling system for astronomical simulation. A two-stage training strategy is adopted to receive external images and user queries. First, the external images are input into a multi-geometry space prompt generation module, and the spatial features P E , H , S of three spatial geometry priors of Euclidean, hyperbolic and spherical are output. The spatial features are input into a multi-modal astronomical data input and alignment module together with the user query. After vector splicing, the input is input into a geometry adaptive backbone module, and finally the corresponding numerical value and reply are output through an astronomical task intelligent inference and output module. By integrating the representations of multiple geometry spaces such as Euclidean, spherical and hyperbolic into the visual language model architecture, the defects of existing visual language models that cannot handle complex astronomical physical phenomena due to their limitation to Euclidean space are effectively solved.

Images