Abstract:Infrared spectroscopy can effectively carry the information of compound structure and compound composition, which has been widely used in chemical research, purity detection and drug identification. However, in practical application scenarios, due to the lack of standard sample preparation conditions, the low accuracy and low recognition efficiency of infrared spectroscopy also make this technology limited. By using a tunable infrared Quantum Cascade Laser(QCL) as the source, an experimental system for recording the diffuse reflection spectra of powder samples is established. Taking the diffuse reflectance spectrum of a mixed powder of glucose and polyethylene as an example, the Kubelka-Munk(K-M) equation and the Kramers-Kronig(K-K) relations are utilized to synthesize the transmission spectrum of the sample from the experimentally measured diffuse reflectance spectrum of pure glucose, and the possibility of the transformation of the transmission spectrum from the diffuse reflectance spectrum is verified. On this basis, spectroscopic data are applied to two neural network models to predict the mass fraction of mixed powder. The results show that the Long Short-Term Memory(LSTM) predicts the best results under the K-K relations, significantly better than the BP neural network. Under the K-M equation, both neural networks are more accurate in predicting the glucose samples with high mass fraction and poorer in predicting the glucose samples with low mass fraction. Both diffuse reflectance spectral correction methods improve the accuracy of the training results, and the LSTM prefers to the BP neural network. This work contributes to the development of the identification of unknown mixed powder samples based on frequency-tunable or broad spectral infrared lasers.