An autonomous living database for perovskite photovoltaics

TL;DR

This work presents PERLA, an autonomous living database that converts continuously growing perovskite literature into FAIR, NOMAD-hosted device data using physics-constrained LLM extraction. The pipeline achieves human-level precision (>90%) with minimal annotator variance, publishes data to a FAIR backbone, and provides open-source tools for other domains. Post-2021 data reveal real-time trends such as the rise of inverted architectures, SAM-based HTLs, and compositional diversification toward FA-rich absorbers, alongside sustained voltage-loss reductions. Collectively, PERLA demonstrates that data-velocity challenges in materials science can be mitigated by integrating LLMs with physics-validation and existing data infrastructure, enabling rapid, data-driven discovery at the speed of publication.

Abstract

Scientific discovery is severely bottlenecked by the inability of manual curation to keep pace with exponential publication rates. This creates a widening knowledge gap. This is especially stark in photovoltaics, where the leading database for perovskite solar cells has been stagnant since 2021 despite massive ongoing research output. Here, we resolve this challenge by establishing an autonomous, self-updating living database (PERLA). Our pipeline integrates large language models with physics-aware validation to extract complex device data from the continuous literature stream, achieving human-level precision (>90%) and eliminating annotator variance. By employing this system on the previously inaccessible post-2021 literature, we uncover critical evolutionary trends hidden by data lag: the field has decisively shifted toward inverted architectures employing self-assembled monolayers and formamidinium-rich compositions, driving a clear trajectory of sustained voltage loss reduction. PERLA transforms static publications into dynamic knowledge resources that enable data-driven discovery to operate at the speed of publication.

Figures (22)

• Figure 1: Overview of the autonomous living database ecosystem PERLA. While the number of papers published in a field continuously increases, the number of papers in a manually curated database remains constant after curation ends. Living databases continue to grow as the field expands and new results are added automatically (indicated by the red star on the efficiency versus year plot). For this, new scientific literature is ingested via journal RSS feeds (or manual uploads) and processed by a Large Language Model (LLM). Extracted data undergoes physics-based validation filtering before being serialized into a structured JSON format. The PERLA ecosystem enables diverse downstream use cases via an application programming interface (API). The API can distribute the continuously updated data to diverse workflows, including retraining machine learning (ML) models to prevent concept drift, facilitating human meta-analysis, and guiding experimental planning in self-driving laboratories. It also enables interactive exploration without programming skills in a graphical user interface.
• Figure 2: Performance of the extraction pipeline. All metrics shown are micro averages, i.e., averaged over all cells in the ground truth. Lines show different models. "Consensus" represents an aggregation of human labelers. a. Precision indicates how many of the extracted entries were correct, i.e., matching the ground truth. For numeric fields, this involves checking if a value is in a given tolerance range. A PCE precision of 0.9, for instance, means that 90% of extracted PCE values agree within the tolerance with the ground truth. In the development of the system, we prioritized precision. b. Recall measures how many of the entries that are in the ground truth have been extracted. A lower recall means that fewer entries have been extracted. Typically, there is a tradeoff between precision and recall and it is difficult to optimize both at the same time.
• Figure 3: PERLA reveals the continued improvement of perovskite solar cells.a Comparison of reported perovskite solar cell efficiencies over time between the curated Perovskite Solar Cell Database (blue) and values automatically extracted from the literature using an LLM (purple). b Yearly distribution of the voltage loss, defined as the difference between the Shockley--Queisser (SQ) limit $V_\text{OC}^\text{SQ}$ and the reported open-circuit voltage $V_\text{OC}$. The red line represents a linear fit ($R^2 = 0.97$), showing a steady reduction in voltage loss of approximately 25mV per year.
• Figure 4: Evolution of bandgap values and perovskite compositions over time.a Kernel density estimation (KDE) of reported bandgap values as a function of publication year, with a histogram of publication counts. The data shown are limited to bandgap values between 1.2 and 2.2eV ($n = 36\,788$). b Distribution of bandgap values represented as violin plots for pure iodide perovskites in the selected compositions ($n = 27\,593$). c Relative frequency of reported short-form compositions until 2021, dominated by MAPbI (96.3%), based on 26 342 entries. d Relative frequency of compositions from 2022 onwards, showing diversification with increased prevalence of FAPbI (32.0%), FAMAPbI (8.8%), and CsFAPbI (13.8%) ($n = 1\,777$). Panels b--d are based on a subset of pure iodide perovskites for which both the absorber composition and an explicit bandgap value could be retrieved from the literature. The analysis was performed for selected compositions (MAPbI, CsFAMAPbI, CsFAPbI, FAMAPbI, FAPbI).
• Figure 5: Temporal evolution of materials and device architectures.a, b, Sankey diagrams illustrating the shifting landscape of material combinations for p-i-n architectures for the a, pre-2022 and b, 2022+ eras. The flow width represents the relative frequency of material pairings between the perovskite absorber, hole transport layer (HTL), and electron transport layer (ETL). Notable trends include the transition from MAPbI toward mixed-cation absorbers and the emergence of self-assembled monolayers (SAMs) in the HTL position. The evolution of n-i-p devices is shown in \ref{['fig:nip_evolution']}. c, Adoption rates of device architectures from 2013 to 2025. The data highlights the historical dominance of the n-i-p (conventional) structure and the steady rise in p-i-n (inverted) configurations, which reached over 40% adoption by 2025. d, Evolution of ETL material preferences, showing the displacement of TiO2 by SnO2 as the primary electron-selective contact.