| As chemical research progresses,a myriad of new chemical substances and compounds are continually discovered and synthesized,unlocking immense potential across various industries and scientific research fields.Ho Iver,with the daily production of numerous new chemical substances,understanding their structural knowledge is essential for comprehending their properties and functions.In traditional chemical research,molecular structure information is typically stored as images in literature and patents,making the extraction,organization,and utilization of this information from such resmyces exceedingly challenging.Manual identification and generation of molecular images are both labor-intensive and prone to errors.Chemistry professionals must invest significant time and effort to identify molecular structures and convert them into machine-readable formats.Moreover,due to the complex nature of molecular structures—comprising various atoms,chemical bonds,and stereochemistry—manual identification methods may have limitations in terms of accuracy.Consequently,it is crucial to develop a more efficient method for identifying and generating molecular images.Deep learning,a potent machine learning technology,has exhibited exceptional performance in image recognition,classification,and generation tasks.This thesis applies deep learning to molecular image recognition and generation,fostering further advancements in chemical research and related fields.In molecular image recognition,this thesis propose a model that integrates Efficient Net V2 with a hybrid attention mechanism LSTM(Eff Net V2+Atten＿LSTM).This model employs Efficient Net V2 for image feature extraction,while incorporating the hybrid attention mechanism LSTM for text generation,resulting in accurate and efficient molecular image recognition.Experimental data substantiates that the model achieves a93.1% accuracy rate on the validation set,significantly outperforming other methods and validating its superiority in molecular image recognition.For molecular image generation,this thesis introduces a new strategy that combines bidirectional LSTM,with autoregressive graph neural networks,and flow-based generative models.This thesis approach captures intermolecular relationships using bidirectional LSTM,learns the topological structure features of molecules through autoregressive graph neural networks,and employs flow-based generative models to enhance the efficiency of the generation process,ultimately achieving precise and efficient molecular image generation.Experimental results demonstrate that this strategy performs perfectly in generating valid,unique,and novel molecules,showing a substantial improvement compared to the prevously known generative models.Additionally,the generated molecules closely align with the original molecular properties and target attributes.Furthermore,this thesis provides a development and implementation process of a molecular translation system,including requirement analysis,system design,and functional demonstrations.Based on the proposed molecular image recognition and generation methods,this system efficiently recognizes and generates molecular images.This innovation is poised to advance research and applications in the chemistry field and enhance the efficiency of molecular recognition and conversion.By introducing novel concepts and technologies to the field of molecular image recognition and generation,this thesis contributes to the improvement of molecular recognition and generation quality,thereby supporting the ongoing development of chemical research and related industries. |
