Extracting hidden features in chemically measured data using deep learning

Abstract

Machine learning is an approach to infer knowledge based on the data using various learning modes (supervised, unsupervised, reinforcement learning, etc). With Machine learning, system can make an own decision from extracted rules. Recently, because of high availability of tremendous data, state-of-the-art machine learning algorithms and fast computational resources, machine learning have brought innovations to various traditional industries and fields by giving new problem-solving paradigm. Unlike conventional machine learning which requires exhaustive a feature engineering, deep learning, which is composed of many stacked layers, have been shown remarkable ability to extract useful features automatically in by optimizing the parameters of artificial neurons with backpropagation algorithm. Amazingly, for recent many years, deep learning has outperformed conventional various approaches in computer vision, natural language processing, drug discovery and so on. In this work, we introduced the deep learning based end-to-end strategy to solve problems in chemical engineering by extracting hidden features in chemically measured data using deep learning. At first, we enhanced both sensitivity and selectivity of chemical gas sensor which are representative performance index of gas sensing by applying deep learning. For an example of sensitivity, we applied deep learning-based anomaly detection to gas sensing and described that our neural network can easily extract hidden signals under conventional limit of detection, which lead to increased sensing abilities for $H_2$ concentration. For an example of selectivity, we showed that deep learning gives metal oxides sensors selectivity of various gases under room temperature. Secondly, we introduced the characterization of porous materials, especially metal organic frameworks. The characterization of structural and adsorption properties of MOFs requires time-consuming experimental characterization processes or computationally expensive molecular simulations. We showed the end-to-end strategy to reduce the time by the deep learning based characterization. We described that artificial intelligence can classify the pristine and linker-defected MOFs after learning with x-ray diffraction data of crystalline MOFs. On the other hands, we simulated the adsorption potential of MOFs by deep learning and described the possibility of calculation of adsorption properties using the learned potential. Our approaches are end-to-end (no need for feature engineering) and easy-and-fast (need simple architecture and only few minutes for training). Furthermore, this approaches have a possibility to be applied to other applications in chemical engineering.