"Intelligent Mining of Infrared Spectral Data based on Multimodal Information Extraction of Catalytic Reaction in Scientific Lite"

"Accurate interpretation of infrared spectral features of the molecular intermediate states is essential for understanding the mechanisms in catalytic reactions. Spectral convolution caused by overlapping vibrational signatures from reaction intermediates poses critical challenges for conventional data analysis, creating an urgent demand for intelligent analytical solutions to accelerate analysis processing. However, building a sufficient experimental dataset for training deep learning models is a critical issue because experiments are very time-consuming. Here, we propose a multimodal information extraction framework that integrates and modifies target detection, character recognition, constructed color recognition, named entity recognition and data processing algorithms, to systematically retrieve key experimental data from the infrared spectroscopy plots and textual descriptions of extensive catalysis scientific literature. This work establishes a structured foundation for constructing an infrared spectroscopy knowledge base, which supports general deep learning model training of \\textit{in-situ} infrared spectral interpretation."

精准解析分子中间体状态的红外光谱特征，对于理解催化反应的反应机理至关重要。由反应中间体振动特征峰重叠引发的光谱卷积问题，为常规数据分析带来了严峻挑战，因此亟需智能分析方案以加速分析处理流程。然而，由于实验过程极为耗时，构建足够规模的实验数据集用于深度学习模型训练是一项核心难题。本文提出一种整合并改进目标检测、字符识别、定制化颜色识别、命名实体识别（Named Entity Recognition）与数据处理算法的多模态信息提取框架，可从海量催化科学文献的红外光谱图与文本描述中系统性提取关键实验数据。本研究为构建红外光谱知识库奠定了结构化基础，可为原位（in-situ）红外光谱解析的通用深度学习模型训练提供支撑。