USING MACHINE LEARNING TO PREDICT OPTICAL PROPERTIES OF MOLECULES

Abstract

Understanding and predicting optical properties at the molecular scale is essential for the development of functional materials in fields such as photovoltaics, sensing, and molecular electronics. Several approaches have been developed to model these properties, ranging from traditional quantum mechanical simulations to emerging datadriven techniques. Traditional quantum chemical methods, such as time-dependent density functional theory (TD-DFT), are known for their high computational demands. This dissertation focuses on using machine learning (ML) to predict optical spectra for molecular systems and potentially reduce the computational cost for calculations. Two sets of molecules were used as testbeds for the machine learning workflow: 1) Organic Molecules from the QM8 database and 2) Metalloporphyrins. Initially, a dataset of small organic molecules was used to train and evaluate machine learning models for the prediction of UV-Vis profile. Two regression algorithms, Kernel Ridge Regression (KRR) and Random Forest (RF), were applied using molecular descriptors generated xv with RDKit. These models were trained and validated on optical property data obtained from quantum chemical calculations using TD-DFT. To further validate the ML models, additional DFT-data was collected for metalloporphyrins. This included information about the geometry, electronic properties, and optical spectra of metalloporphyrins that included first and second row transition metals with varying anchoring groups. The relatively small dataset for the optical properties of metalloporphyrins introduced challenges to the ML model. This research highlights importance of structure and composition on optical properties and how machine learning can provide insight into the optical properties and ultimately molecular design principles for specific applications.