Automated Experimentation and Machine Learning for Perovskite Photovoltaics

As anthropogenic global warming and climate change continues to intensify, it is more important than ever to curb our use of fossil fuels for heat and electricity. The clean energy transition requires investment in photovoltaic (PV) technology to replace unsustainable generation sources and to meet increasing electrification demand. Most currently available PV modules use Si as the light absorber material. Advancements in the manufacturing process of Si solar cells has led to a reduction in cost over the past decade, yet there are still inherent difficulties in the refinement process. Next-generation PV must improve upon Si by maintaining or increasing device efficiency while simplifying manufacturing. Metal halide perovskite solar cells (PSC) have proven to meet these criteria, with >25% efficiency and cost-effective fabrication options such as blade coating and ink jet printing. However, degradation in perovskite materials under several environmental stressors (light, humidity, temperature, bias, and oxygen) precludes commercial adoption. Stability testing of PSC is time-consuming, particularly due to the large compositional space available. As a result, methods to quickly vet the stability of various perovskite compositions are critically needed. My thesis addresses this open problem by applying automated, in situ characterization and machine learning (ML) forecasting to PSC degradation studies.First, I present a generalized ML roadmap for perovskite PV. I delineate three levels of PSC design and provide examples of ML projects which could accelerate the development process. Next, I design and build a high-throughput setup for in situ, environmental photoluminescence (PL) spectroscopy, a technique which requires <2 seconds to acquire data. The system uses a custom chamber containing up to 14 samples and an x-y translation stage to automatically move from one sample to the next. We select ten CsyFA1-yPb(BrxI1-x)3 perovskite thin films of varied composition and track changes in radiative recombination under repeated 6-hour temperature and rH cycles. Using the high-throughput setup, I obtain 240 PL spectra every hour and 14,000 spectra over the course of a single experiment. The temperature cycling results show increased non-radiative carrier recombination as samples are heated above 23°C. We show that FA-rich perovskites with 10-30% Cs+ have minimal lattice strain which promotes high structural and thermal stability. During rH cycling, all compositions displayed a PL enhancement with increasing rH as H2O passivates band gap trap states and suppresses non-radiative recombination. FA-rich films show the greatest PL increases over the course of the rH cycling while Cs-rich films reach a plateau in maximum PL value after 5-10 cycles. Finally, I apply ML models to the datasets and generate forecasts of environment-dependent PL responses. I use linear regression, Echo State Network (ESN), and Auto-Regressive Integrated Moving Average with eXogenous regressors (ARIMAX) algorithms. For the temperature cycling, I attain an average normalized root mean square error (NRMSE) over all compositions of 24.4% (linear regression), 16.6% (ESN), and 7.3% (ARIMAX) for prediction windows extending 70 hours into the future. For the rH cycling, NRMSE values of 72.5% (linear regression) and 44.0% (ESN) indicate difficulty in tracking long-term changes over a 50-hour window. Using ARIMAX with seasonality components, I achieve an error of only 10.3%, demonstrating the algorithm’s capability to model complex, non-linear data from varied perovskite compositions. My high-throughput characterization results and accurate time series forecasts illustrate the potential of data-centric approaches for perovskite stability investigations and showcase the promise of automation, data science, and ML as tools to drive PSC commercialization