Stable and Reproducible Halide Perovskites: A Machine Learning Approach
- Hering, Abigail Rae
- Advisor(s): Leite, Marina S
Abstract
As global energy consumption surges, it is becoming crucial to improve renewable energy sources and alternative functional materials to offset these demands. Computational energy costs associated with data centers and artificial intelligence (AI) are especially projected to grow in the coming decades. Halide perovskites (HPs), an emerging semiconducting material, can be used as both solar absorber and data storage materials, but their deployment in commercial devices must be achieved on accelerated timescales to meet this urgent global need. HPs were discovered at a unique period, where advancements in data science, autonomous labs, and machine learning (ML) tools can be harnessed to accelerate their development.HPs have remarkable optoelectronic properties, yet two key challenges in the field remain: material irreproducibility and instability. In this dissertation, I discuss the latest achievements in HP materials research accomplished through the implementation of automated experimentation and ML-driven data analysis. Automated synthesis and characterization tackle problems with material irreproducibility by systematically controlling parameters with very high precision, creating massive datasets, and allowing methodical comparisons from which unbiased conclusions can be drawn. ML can reveal otherwise unnoticed trends, inform future experiments with the highest potential information gain, and forecast future performance.HPs exhibit dynamic properties in response to environmental stressors due to several composition-dependent degradation mechanisms. It is difficult to quantify their behavior upon exposure to each stressor by exclusively using trial-and-error methods due to the high-dimensional parameter space involved. I combine high-throughput experiments, data visualization, and ML techniques to quantify correlations between stressors and material properties by analyzing high-throughput, in situ environmental photoluminescence (PL) experiments. Dimensionality reduction visualization methods uncover clear composition-based clusters despite overlapping datasets. A robust screening of 10 ML algorithms effectively forecasts PL features with over 90% accuracy. Using a multi-output composition-generalized Extra-Trees and Ridge Regression stacked model, a full PL spectrum can be predicted for any time, temperature, and composition input.Next, I implement automated robotic synthesis for high-throughput fabrication of multiple batches of reproducible samples. High-throughput optical and structural characterization of these robot-synthesized materials allow for comparisons of metrics before and after in situ degradation experiments. I discuss composition-based predictions of HP features over time in response to temperature and relative humidity interrogation. High-accuracy predictions (>85%) are made with five distinct ML algorithms, which are effective for both classification and regression tasks. Initial optical characterization methods, such as PL imaging, can effectively predict whether an HP sample will experience a phase transition upon environmental exposure. These results show that despite vast differences in long-term trends and irregular behavior, some generalizations about the effects of stressors, composition, and initial state can be used in predicting long-term stability.This dissertation concludes with a forward viewpoint of how human-assisted closed-loop laboratories and shared databases allow halide perovskite materials’ processing, properties, and performance to be optimized with ML, thus accelerating scientific breakthroughs related to this material.