Olfactory language model system and cross-modal conversion method

Abstract

The present application provides an olfactory language model system and a cross-modal conversion method, comprising: a molecule encoder module for converting an input chemical molecule structure into a molecule embedding vector representing the odor characteristics; an odor atmosphere encoder module for processing the odor representation of a multi-component mixture composed of a plurality of molecules mixed in a specific proportion, and generating an odor atmosphere embedding vector; a multi-modal alignment module for constructing a unified vector space, coupling a multi-fingerprint fusion strategy with a nonlinear mixing effect encoder, so that the system can capture the microscopic synergy and antagonism between molecules, solve the technical problem of inaccurate representation of complex odor atmosphere, and realize strong association between molecules, atmosphere and text in the unified vector space through a contrast learning mapping mechanism, eliminate the cross-modal semantic gap, so that the odor description not only stays at the label level, but also can generate time evolution characteristics including pre, middle and post modulation through the Transformer architecture.

Images