Accelerating Perovskite Solar Cell Design with an Integrated Data and Machine Learning Platform
Basir Akbar a, Hilal Tayara a
a Department of Physical-AI Convergence Engineering, Jeonbuk National University, Jeonju 54896, Republic of Korea
Poster, Basir Akbar, 122
Publication date: 22nd October 2026

Perovskite solar cells (PSCs) have improved rapidly, yet designing better devices still means searching an enormous space of compositions, architectures and processing conditions. Progress is slowed because relevant data are scattered across many publications and repositories, are rarely standardized, and published machine learning (ML) models usually require coding skills to run. We present PerovskiteAI, an open, web-based platform that brings curated data, analysis tools and trained prediction models together in a single interface that needs no programming. The platform holds more than 2.5 million cleaned and standardized data points drawn from peer-reviewed experimental reports, density functional theory calculations and drift-diffusion device simulations. These data are arranged into 10 research modules: 3D cubic perovskites, 2D Ruddlesden-Popper and Dion-Jacobson phases, lead-free compositions, photovoltaic performance, optoelectronic properties, charge transport, recombination and loss mechanisms, voltage loss, defect engineering, and stability and degradation. Each module provides filtering, interactive plots, correlation analysis and automated feature generation from compositional descriptors, SMILES strings and numerical device parameters. On top of these datasets, the platform deploys specialized models ranging from gradient boosting and stacking ensembles to transformer networks. Across the regression tasks, which include bandgap, formation energy, HOMO level and device metrics such as open-circuit voltage and short-circuit current, test-set R² values range from 0.56 to 0.97, with most models above 0.85. Two classification models reach test accuracies of 0.93 for identifying the dominant recombination pathway (band-to-band, grain-boundary or interface) and 0.83 for separating Ruddlesden-Popper from Dion-Jacobson phases. As a worked example, analysis of the recombination module, built on 2.47 million simulated devices, shows that the ideality factor relates in opposite directions to band-to-band and grain-boundary recombination, giving users a physically interpretable basis for diagnosing losses. A conversational assistant guides users through data exploration, prediction and screening in plain language. By lowering the technical barrier to data-driven design, PerovskiteAI aims to serve as shared, extensible infrastructure that helps experimental groups prioritize compositions and device configurations before synthesis.

© FUNDACIO DE LA COMUNITAT VALENCIANA SCITO
We use our own and third party cookies for analysing and measuring usage of our website to improve our services. If you continue browsing, we consider accepting its use. You can check our Cookies Policy in which you will also find how to configure your web browser for the use of cookies. More info