🔬 materials science

An autonomous living database for perovskite photovoltaics

This paper introduces PERLA, an autonomous, self-updating living database that leverages large language models and physics-aware validation to extract high-precision data from post-2021 perovskite photovoltaics literature, revealing critical evolutionary trends toward inverted architectures and formamidinium-rich compositions that were previously obscured by data stagnation.

Original authors: Sherjeel Shabih, Hampus Näsström, Sharat Patil, Asmin Askin, Keely Dodd-Clements, Jessica Helisa Hautrive Rossato, Hugo Gajardoni de Lemos, Yuxin Liu, Florian Mathies, Natalia Maticiuc, Rico Meitzner

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sherjeel Shabih, Hampus Näsström, Sharat Patil, Asmin Askin, Keely Dodd-Clements, Jessica Helisa Hautrive Rossato, Hugo Gajardoni de Lemos, Yuxin Liu, Florian Mathies, Natalia Maticiuc, Rico Meitzner, Edgar Nandayapa, Juan José Patiño López, Yaru Wang, Lauri Himanen, Eva Unger, T. Jesper Jacobsson, José A. Márquez, Kevin Maik Jablonka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of solar cell research as a massive, chaotic library. Every day, hundreds of new books (scientific papers) are written about how to make better solar cells. The problem is that the librarians (human scientists) are drowning. They can't read, categorize, and file these new books fast enough. As a result, the "index" of the library—the database that tells us what we know—stopped updating in 2021. This left scientists working with an outdated map while the territory changed rapidly.

This paper introduces a solution: PERLA, an "autonomous living database." Think of PERLA not as a static filing cabinet, but as a self-updating, robotic librarian that never sleeps.

Here is how it works, broken down into simple concepts:

1. The Robotic Librarian (The Pipeline)

Instead of humans manually reading every new paper, PERLA uses a team of AI "readers" (Large Language Models).

The Watcher: Imagine a robot that constantly scans the newsstands of scientific journals. As soon as a new paper about solar cells is published, the robot grabs it.
The Reader: The robot reads the paper and extracts specific details, like "What materials did they use?" or "How efficient was the battery?"
The Physics Filter: This is the most critical part. AI can sometimes "hallucinate" (make up numbers). To prevent this, PERLA acts like a physics teacher. If the AI says a solar cell is 200% efficient (which is impossible), the teacher immediately rejects it. The system checks every number against the laws of physics to ensure it makes sense before saving it.

2. The "Living" Difference

The authors compare the old way of doing things to a photograph and their new way to a live video feed.

The Old Way (Manual Curation): Like taking a photo of a crowd in 2021. Even if you look at that photo in 2026, the people in it haven't changed, but the world outside the photo has. The data becomes stale the moment the photo is taken.
The New Way (PERLA): Like a live video stream. As soon as a new scientist publishes a breakthrough, it appears in the database instantly. The database "grows" automatically, just like a living organism.

3. What Did They Discover?

Because they finally had access to the "live feed" of research from 2021 to 2026, they uncovered trends that were hidden by the old, stagnant database:

The Shift in Design: The field has decisively moved away from one type of solar cell design (called "n-i-p") toward a newer, more efficient design (called "inverted" or "p-i-n").
The Magic Ingredients: Scientists have stopped using one specific chemical ingredient (methylammonium) and are now mixing in new, more stable ingredients (like formamidinium).
The "Voltage Loss" Mystery: They found that scientists are getting better and better at reducing "voltage loss" (energy wasted as heat) by about 25 millivolts every year. This is a steady, linear improvement that was hard to see before.

4. Why This Matters for Computers (AI)

The paper also explains a problem with training computer models (AI) to predict solar cell performance.

The "Outdated Map" Problem: If you train a GPS on a map from 2021, it will give you wrong directions for a city that has built new roads since then.
The Fix: By feeding the AI the current data from PERLA, the computer models learn the actual rules of the modern world, not the rules of the past. The paper shows that models trained on this fresh data are significantly more accurate.

Summary

The paper claims that PERLA solves the bottleneck of human speed by using AI to read, verify, and organize scientific papers automatically. It turns a static, outdated list of facts into a dynamic, self-updating resource that reflects the real-time speed of scientific discovery. This allows scientists to see trends immediately and trains computer models on the most current data available, rather than data that is years old.

What they did NOT claim:

They did not claim to have invented a new solar cell that powers a city.
They did not claim this technology is ready for commercial use in homes yet.
They did not claim this solves all problems in materials science, only that it solves the specific problem of "data lag" in perovskite solar cells.

The core message is simple: We built a robot librarian that keeps the map of solar energy research up-to-date, so scientists don't have to waste time reading old books.

Technical Summary: An Autonomous Living Database for Perovskite Photovoltaics

Problem Statement
Scientific discovery in materials science, particularly in perovskite photovoltaics, faces a critical bottleneck: the rate of publication has outpaced the capacity for manual data curation. While the Perovskite Database (PDB) successfully established FAIR data standards for approximately 42,400 devices across 15,000 papers up to 2021, it has remained stagnant since. The thousands of papers published post-2021, containing tens of thousands of additional device records, remain trapped in unstructured text. This creates a "knowledge blackout," preventing the scientific community from systematically assessing recent breakthroughs or training machine learning models on state-of-the-art architectures. Manual extraction is too slow to keep pace with the literature, and existing automated approaches often lack the flexibility to understand scientific context or the reliability to match human precision.

Methodology: The PERLA Pipeline
The authors introduce PERLA (Perovskite Living Archive), an autonomous, self-updating database that converts the continuous stream of scientific literature into structured, FAIR-compliant data. The system operates through a multi-stage pipeline:

Literature Ingestion: An automated "watcher" monitors RSS feeds from major journals. It uses permissive regular expressions to flag candidate papers, followed by strict pattern matching on abstracts to filter out reviews, computational studies, and non-solar applications. Full texts are retrieved via open-access channels (e.g., Unpaywall).
Constrained LLM Extraction: The core of the system utilizes Large Language Models (LLMs) to extract device parameters. Unlike previous regex-based methods, LLMs are employed for their ability to understand scientific context (e.g., recognizing that "Spiro-OMeTAD" and "Spiro-MeOTAD" are identical). The extraction is constrained by a strict device schema aligned with community practice (architecture, stack, processing, performance metrics, stability) and enforced via structured decoding (JSON output) to prevent schema hallucinations.
Physics-Aware Validation: To ensure reliability, extracted data undergoes rigorous filtering:
- Unit Normalization: All quantities are converted to canonical forms.
- Physical Consistency Checks: The system verifies that reported Power Conversion Efficiency (PCE) aligns with the Shockley-Queisser limit and the fundamental relation $PCE = J_{sc} \times V_{oc} \times FF / P_{in}$ . Entries violating these physical laws or showing dimensional mismatches are discarded.
- Hallucination Detection: The system cross-references extracted numerical values against the original PDF text to ensure numbers were actually reported.
Evaluation and Ground Truth: The system's performance was benchmarked against a ground truth dataset constructed by 13 domain experts. Due to significant inter-annotator variance (only 13% full consensus among humans), the study utilized arbitration to resolve discrepancies. Multiple state-of-the-art models were tested, with Claude Sonnet 4 selected for the final pipeline due to its superior precision.

Key Results

Human-Level Precision: The automated pipeline achieves a precision exceeding 90% across all field categories (including PCE, $V_{oc}$ , $J_{sc}$ , and material composition), matching or exceeding the performance of top-tier human experts while eliminating annotator-to-annotator variance.
Data Volume: The system has successfully processed post-2021 literature, contributing 2,582 publications and 7,816 devices to the database, extending the legacy PDB baseline.
Emerging Scientific Trends: By analyzing the newly structured post-2021 data, the authors identified critical evolutionary trends previously obscured by data lag:
- Voltage Loss Reduction: A sustained, linear reduction in voltage deficit (approx. 25 mV per year) indicates continuous progress in mitigating recombination losses.
- Compositional Shift: The field has decisively shifted from Methylammonium Lead Iodide (MAPbI $_3$ ) dominance (>96% pre-2021) to Formamidinium-rich compositions (FAPbI $_3$ and mixed-cation variants), driving bandgaps closer to the optimal 1.50–1.55 eV range.
- Architectural Transformation: There is a rapid adoption of inverted (p-i-n) architectures, which now account for over 43% of devices, alongside the explosive emergence of Self-Assembled Monolayers (SAMs) as hole-transport materials.
- Maturation of the Field: Information-theoretic analysis (Shannon entropy and Heaps' Law) suggests the field has moved from broad chemical exploration to the refinement of established high-performance systems, with interface engineering becoming the primary frontier of innovation.

Significance and Claims
The paper claims that PERLA resolves the "data velocity" problem in materials science by transforming static publications into dynamic, self-updating knowledge resources. The authors argue that:

Automation is Viable: Coupling LLMs with physics-based constraints allows for automated data extraction that is not only faster but potentially more consistent than manual curation.
Machine Learning Relevance: Continuously updated databases are essential for training machine learning models to avoid "concept drift." The paper demonstrates that models trained on historical data fail to generalize to modern compositional shifts, whereas models trained on mixed (historical + modern) data show significantly improved prediction accuracy for bandgap and PCE.
Scalability: The infrastructure serves as a blueprint for other rapidly evolving research domains, offering a scalable path to operate at the true pace of scientific discovery.

The authors release the complete system—including the extraction pipeline, accumulated data, code, and documentation—as an open-source community resource hosted on NOMAD, enabling programmatic access and interactive exploration without requiring programming skills.