(375t) Neural Network Analysis of NMR and IR Spectra

Nuclear magnetic resonance (NMR) and infrared (IR) spectroscopy are useful analytical tools for determining chemical composition based on functional group behavior. However, the manual evaluation of these spectra is time-consuming and requires significant expertise. Furthermore, expert review incurs a time delay, making the implementation of these techniques within a process control scheme difficult. An autonomous computational technique to quickly and accurately read these spectra would allow the information to be used to make real-time process control decisions. In this work, machine learning algorithms are applied on ¹H NMR and IR spectra to gauge their effectiveness in accurately identifying functional groups within small organic compounds. Both simulated and experimental spectra are extracted from multiple online databases and reformatted according to a consistent standard. The presence of thirty-seven functional groups for each of 5,000 spectra was determined within our dataset, and then this set is split into training and testing subsets. Using the Keras library within Python, convolutional neural networks (CNNs) are then implemented to take the input spectra from the training set and return the predicted functional groups. The predictions for both the training and testing sets are evaluated with strong interest in metrics measuring overfitting, and the accuracy is compared to that obtained via the traditional route of human-led identification, confirming the effectiveness of the model. Continued work aims to optimize parameters of the neural networks to improve their accuracy and allow prediction for components in mixtures, and in a variety of solvents.