AVS 72 Session AIML3-FrM: Connecting Datasets, Models, and Characterization

We present a hybrid machine learning (ML) and physics-based computational modeling framework that informs experimentally realizable thin charge-transport-layers for perovskite solar cells (PSCs). Using supervised ML models, 12294 previously published curated datapoints predicted device efficiency. Ten high-efficiency descriptor combinations with greater than 21% power conversion efficiency were mapped into SCAPS-1D (a physics-based simulation software) employing a FASnI₃ absorber to assess their physical plausibility. In contrast with the ML predictions, the SCAPS configurations yielded PCEs in the range of 0.03–22.7%. Discrepancies between ML and SCAPS predictions arose from interfacial recombination, optical losses, and physical constraints not explicitly encoded in the ML dataset, describing a tightly focused set of fabrication parameters. Thus, the complementary use of both machine learning and physical simulations offers a time- and material-saving scalable pathway for the development and fabrication of low-cost photovoltaic devices.

In this project, we are developing an AI-assisted thin-film materials discovery workflow for halide perovskite solar cells. The goal is to make materials discovery more connected, adaptive, and efficient by bringing together steps that are often handled separately: literature knowledge, predictive modeling, robotic synthesis, and high-throughput characterization. Instead of treating these as independent parts of the research process, the workflow is being designed so that information from one stage can help guide the next. A large part of this work has focused on building a literature-mining agent for perovskite solar-cell papers. The perovskite literature contains a large amount of useful information, including device structures, perovskite compositions, processing conditions, interface layers, power conversion efficiencies, and stability measurements. However, this information is often reported in different ways across papers, which makes it difficult to use directly for machine-learning studies. The literature agent is designed to read papers, extract the relevant information, and organize it into a structured format compatible with the Perovskite Database. The workflow also aims to preserve where the information came from, separate evidence from model-ready features, flag incomplete records, and identify possible duplicate device entries. This makes the dataset update process more transparent and easier to check. The updated dataset can then be used to train sequential machine-learning models for device performance and stability. First, power conversion efficiency is predicted from design, processing, interface, and device-context features while avoiding target leakage from already measured performance values. Stability is then modeled using a similar feature space. We compare design-only stability prediction, stability prediction using model-predicted PCE, and post-fabrication prediction using measured initial device performance. This structure helps separate early design guidance from later experimental decision-making. This work is an important step toward a closed-loop discovery platform for perovskite and related semiconductor thin films. Literature knowledge can update the database, models can suggest promising directions, robotic synthesis can test those ideas, and high-throughput characterization can provide new information that feeds back into the system. While the workflow is still being developed, it provides a practical foundation for more efficient and autonomous thin-film experimentation.