This thesis principally addresses the stability issues of organic photovoltaics (OPVs) by adopting machine learning approaches to identify the causes of degradation and identify the factors which contribute most significantly.
To achieve this, multiple datasets of OPV stability and performance parameters have been collated and studied. The first of these datasets contains data sourced from the literature containing detailed device information, including the structure, materials and testing conditions. The sequential minimal optimisation regression (SMOreg) machine learning algorithm was applied to the dataset, allowing for the ten most beneficial, and ten most detrimental, attributes to be identified for OPVs tested under ISOS-L and ISOS-D protocols. In doing so, the optimum configuration of OPVs can be identified, although further experimental verification is needed to confirm the results.
Machine learning was also applied to an OPV dataset derived from modules tested outdoors. The aim of this work was to allow the performance and stability of the modules to be correlated with the weather conditions. Principal component analysis illustrated that the most influential weather conditions were UV, irradiance, temperature and wind speed. A novel approach was taken for degradation forecasting, whereby the cumulative contribution of each weather condition was used as the predictive attributes, such that the total “dose” of each condition was accounted for. The degradation was forecasted by combining two machine learning algorithms. The method allowed the energy yield of OPVs tested outdoors to be predicted within 5% of the actual value using unseen data.
Finally, the database of OPV structures, performance and stability was extended to include the embodied energy of each material and processing step. From the initial performance, the stability and the embodied energy, the ‘net energy output’ was calculated which assesses the predicted energy yield of the OPV and subtracts the embodied energy. The SMOreg algorithm was used to determine the attributes which most significantly govern the embodied and net energies. A genetic search clustering algorithm (GenClust++) was used to identify the optimum combinations of the materials to maximise the net energy. This provided a means of finding new and improved combinations of materials and structures which have a lower environmental footprint whilst maximising the energy generation potential.