Bridging Experimental Materials Science and Machine Learning: A Data-Centric Study of Hybrid Metal Halides

Abstract

The discovery and optimization of functional materials face fundamental challenges because the relationships among structure, properties, and performance are intrinsically high-dimensional, nonlinear, and are associated with significant uncertainty. These challenges are particularly pronounced in hybrid metal halide materials, where subtle variations in organic cation chemistry, electronic structure, and synthesis conditions can induce substantial changes in crystal structure, optoelectronic properties, and device performance. Experimental materials data in this field are typically limited in size, heterogeneous in nature, and subject to significant inter-experimental variability, which complicates the extraction of generalizable design principles.This dissertation explores how experimental studies and machine learning can be integrated to enable rational materials design in data-scarce and uncertainty-rich material systems. Using hybrid perovskites and hybrid metal halides as representative research platforms, this work presents an integrated research framework that begins with experiment-driven physical understanding and progressively extends toward data-driven modeling and uncertainty-aware design strategies.In Chapter 2, a systematic experimental investigation of p-type charge transport doping in mixed Sn-Pb halide perovskites establishes the physical foundations of charge carrier modulation. By employing molecular dopants introduced via a sequential deposition strategy, this work investigates how dopant chemistry, interfacial interactions, and dopant incorporation behavior within thin films influence carrier concentration, mobility, and electrical conductivity, demonstrating the pronounced sensitivity of electronic properties to subtle chemical variations.In Chapter 3, this work develops an interpretable machine learning model based on a curated dataset of two-dimensional perovskite solar cells. By integrating literature-derived device data with chemical descriptors of organic spacer cations, a gradient boosting model capable of predicting power conversion efficiency under small-sample and data-imbalanced conditions is constructed. Feature importance analysis confirms that intrinsic chemical factors - such as the organic spacer-to-metal ratio - dominate device performance, rather than the device architecture itself, highlighting the central role of chemical design in perovskite solar cells.Finally, Chapter 4 proposes an uncertainty-aware machine learning framework using Bayesian Additive Regression Trees to predict the structural dimensionality of low-dimensional hybrid metal halides. By quantifying predictive uncertainty, this approach identifies key molecular features governing crystal dimensionality and suggests that information-gain-driven active learning strategies can more efficiently guide experimental exploration. Compared to conventional models that rely on point predictions, this framework enables more reliable design decisions in small, heterogeneous data regimes via uncertainty quantification.Through this work, it is established that understanding and designing complex materials systems such as hybrid metal halides require strategies that go beyond reliance on individual experiments or intuition, instead integrating experimental investigation with data-driven modeling. Detailed experimental analyses of molecular doping establish a physical foundation for electronic structure modulation, while interpretable machine learning demonstrates that chemically meaningful descriptors can govern performance even in limited experimental datasets. Furthermore, by employing Bayesian models to quantify uncertainty and incorporating active learning strategies, this work shows that rational experimental design can be achieved by accounting for predictive reliability. Collectively, these findings indicate that materials design can transition from empirical exploration toward a data-centric decision-making process that inherently incorporates uncertainty.