紅外線光譜常被使用來研究化學物質結構，但是使用傳統的計算方法取得計算紅外線光譜將是不準確的或是需要耗費大量計算資源。在這項研究中，深度學習將被研究用來生成有機化合物的計算光譜。我們設計並訓練了一個卷積神經網路模型以加快及提高分子力學生成的光譜品質，這個模型學習了計算與實驗光譜之間的差異，可用來生成比原本只使用分子力學更高品質的光譜。我們設計了一個採用採用變分自動編碼器架構的模型，以研究生成式模型學習光譜生成的能力。這個變分自編碼器模型不僅在重建光譜方面性能非常接近先前的卷積模型，同時也能夠使資料的分布在潛在空間（latent space）中維持高斯分佈。收集和預處理光譜資料的方法也有被研究以為神經網路模型準備訓練資料集。;Infrared spectroscopy serves as a common tool for the analysis of chemical structure. The conventional computational methods for infrared spectral simulating are either time-consuming or inaccurate. In this work, the use of deep learning models was studied for the task of generating infrared spectra of organic compounds. A convolutional neural network model was designed and trained to fast generate experimental-like spectra with molecular mechanicsspectra as input.The model which learned the differences between experimental and molecular mechanics generated spectra can be applied to produce spectra that are better than the original spectra generated by molecular mechanics calculations. A model that adopts variational autoencoder architecture was designed to investigate the power of generative models for learning the generation of spectra. The variational autoencoder model is able to not only reconstruct spectra with a performance very close to the previous convolutional model but also maintain the distribution of data in the latent space Gaussian. The methods of collecting and preprocessing spectral data were also investigated to prepare a training dataset for the neural network models.
