Deep Learning for Raman Spectroscopy: Bridging the Gap between Experimental Data and Molecular Analysis

Raman spectroscopy is an analytical method frequently used in the fields of materials science and general chemistry, which measures the characteristic responses of molecules to light. Being a non-destructive technique, it has many scientific and industrial applications, such as material identification, drug production and airport security. By combining Raman spectroscopy with machine learning, powerful tools can be developed to make accurate predictions on unknown substances or quantities, without explicit programming.

This thesis focuses on two main areas. Firstly, through the application of machine learning and image processing, a novel tool was developed to study single molecule interactions with metal surfaces using surface-enhanced Raman spectroscopy (SERS). A convolutional autoencoder (CAE) architecture was utilised in a processing pipeline alongside various image processing techniques to extract and isolate complex, transient Raman features from SERS data. This was followed by a clustering process to obtain representative events pertaining to atomic-scale metal-molecule interactions on multiple catalyst surfaces, which provided a unique insight into the formation dynamics of atomic-scale features. The process was extended through the use of a Siamese convolutional neural network (Siamese-CNN) to incorporate spatiotemporal information relating to interactions between individual vibrational modes. This foundational research paves the way for tailoring metal-molecule interactions and assists in rational heterogeneous catalyst design. It introduces an analytical tool capable of studying metal-molecule interactions under the influence of strong local field gradients. This is a scenario that cannot be efficiently modelled with the conventional quantum mechanical method, density functional theory (DFT), which assumes a homogeneous field when analysing electronic structures of molecules.

Secondly, machine learning analysis has been applied to Raman data obtained in both nuclear and biopharmaceutical industrial applications. A key focus of this work is on the practical challenges faced in the design of data processing tasks and machine learning architectures due to real-world limitations in data collection. A fully connected (FC) autoencoder is employed as part of a regression task, which generates predictions on analyte concentrations in mixed substances. The method was shown to outperform industry standard regression tools, principal component regression (PCR) and partial least squares (PLS) regression, each used as comparative benchmarks, by over 50% in a test of model precision across various datasets in the investigated industrial applications. Advancements in the precision, speed and effectiveness of such tools are of critical importance in an industrial environment. This is driven by compelling motivations to reduce not only the costs associated with these procedures, but also to increase the quality of resulting products, or to reduce the risks within industrial operations, where applicable.