(156f) Neural Network-Based Automated Detection of Functional Groups in Spectroscopic Data- Bridging the Gap in Online Monitoring of Complex Reaction Systems

The problem of reliable monitoring of complex reacting systems is largely dependent on accurate enumeration of the reacting species and the reactions governing their conversion. However, the absence of information, as it is common in the case of distributed feedstocks such as bitumen and biomass, makes monitoring reliant on (i) molecular-level data from process analytical tools like spectroscopic sensors, and (ii) high physical fidelity statistical models that use the process data to infer the reaction mechanism structure. Understanding the actual chemistry in the reactions of such complex feedstocks without a priori knowledge of the system becomes intractable due to the expanse of chemical reaction space. In the case of system where both reactions and species are known, the development of the reaction network becomes a simple task of ‘applying’ the reactions to the species based on expert knowledge. In contrast, systems for which both reactions and species are unknown, require mathematical manipulations of the spectroscopic data to decipher the reaction network topology. Structure preserving non-negative tensorial factorization of the time evolving spectroscopic data results in latent factors that can be projected on to the spectral mode to be interpreted as the spectra of the reactive pseudo-components. Domain knowledge can then be employed to map the spectral peaks to signatures of active functional groups in each component [1].

We present an automated technique to infer information about functional groups in a spectrum without prior knowledge of the species. The identification problem is cast as a classification problem and is solved using an artificial neural network. The network architecture incorporates 1-dimensional convolution blocks to extract features corresponding to a specific functional group. The convolution block slides filters of predetermined size along the wavenumber axis. The convolution operation between the filter and the input spectra generates localized regions of ‘hot spots’ indicating a match between the two. The usage of a 1-D convolution block restricts this operation to just one axis, therefore making the classifier absorbance invariant. According to Beer’s Law, the height of a peak in spectrum is dependent on the concentration, path length and the absorptivity of the compound in that particular region. In the case of complex reactive mixtures, the concentration profiles of the species vary over time thus engendering a need of absorbance invariance in an automated detection scheme. Previous works have established a similar routine for automated functional group detection using Convolutional Neural Networks (CNNs) but incorporate a 2-D convolution operation, thus making the system dependent on the peak height [2]. The convolution blocks are followed by a fully connected layer to analyse the features from the convolutional layers. The output of the network is a bit vector of size 13 indicating the presence or absence of a particular functional group in the species with a sigmoidal activation layer being used to calculate the probabilities of each functional group. The network is trained to simultaneously predict the presence of different functional groups and hence training the neural network becomes a task of learning the joint probability distribution for each functional group in the presence of others [3]. This task is commonly referred to as multi-label classification, where an input can belong to multiple classes simultaneously. In this regard, the training is performed using backpropagation with the binary cross entropy loss for each bit in the output. Hamming loss is used as a test metric to quantify the degree of accuracy of the classifier. The Hamming loss can be defined as the fraction of incorrect predictions over the label size. In order to improve the confidence on the predictions, a reconstruction metric is also considered. The reconstruction metric computes the difference between the input spectrum and a theoretical spectrum based on the predictions of the classifier. This allows for the analysis of the feature space, improving the fidelity of the model.

The network hypothesis generation is then carried by learning a graphical structure from the resolved pseudo-components using Bayesian structure learning algorithms. The graph indicates the transition of one species to another, thereby providing the skeletal structure of the network where the nodes of the graph correspond to each pseudo- component whose spectral signature have been identified by the classifier. The generation of the reaction graph, also provides a validation metric for the classification. This protocol has been applied to describe the reaction network for the hydrous pyrolysis of biomass and results indicate a high level of correspondence with the reactions described in the literature thus providing the ability to perform online monitoring with interpretability without any prior knowledge of the system.

Keywords: Convolutional neural networks, reaction hypothesis, latent factor projection, Bayesian networks, reaction monitoring