AI Meets Solar Simulation to Track Moisture Damage in Perovskite Cells

Perovskite solar cells have largely solved their efficiency problem, with power conversion efficiencies exceeding 26 percent within roughly a decade of modern development. What now stands between the technology and commercialization, according to a new systematic review, is not how much power these devices can produce but how long they can keep producing it. The review argues that the intrinsic hydrophilic behavior of organic–inorganic halide perovskites drives time-dependent efficiency decay, and that the research community needs to shift its benchmarking from efficiency-centric metrics toward stability-normalized retention metrics.

The review, published in Discover Electronics, was conducted by Twinkle Gupta, Ankur Bhardwaj, and Ajay Kumar using the PRISMA guidelines for systematic reviews. The team searched Scopus, IEEE Xplore, Google Scholar, ACS Publications, Nature Energy, Nano Energy, the Royal Society of Chemistry, and Elsevier’s ScienceDirect for studies published within the last ten years, combining keyword strings covering perovskite solar cells, moisture, degradation mechanisms, SCAPS-1D, machine learning, and hydrophilic behavior. From an initial corpus of approximately 2,818 articles, duplicate removal left 1,756 records, and screening of titles and abstracts excluded 1,183. After full-text assessment of 573 articles against eligibility criteria, 105 studies were included in the final synthesis.

The central chemical problem the review identifies is structural. Organic cations such as methylammonium (MA+) and formamidinium (FA+), which are widely used in high-efficiency perovskite absorbers, contain polar amine functional groups like –NH2 and –NH3+ that readily form hydrogen bonds with water molecules. When devices are exposed to humidity, degradation proceeds through a multistep mechanism: water molecules first adsorb onto the perovskite surface, then intercalate into the crystal lattice, promoting hydration reactions that form intermediate hydrate phases such as MAPbI3·H2O. These phases alter the optoelectronic and light-harvesting properties of the absorber layer and ultimately cause irreversible loss of photovoltaic performance. Decomposition reactions can also split MAPbI3 into gaseous byproducts and solid lead iodide.

Once moisture initiates this chemistry, the electrical consequences cascade through the device. Water infiltrating via grain boundaries and defect-rich regions accelerates non-radiative recombination, increases trap density, and disrupts the interfacial energetics at the electron and hole transport layer junctions, all of which manifest as a progressive decline in efficiency over time. The review establishes a conceptual chain linking hydrophilicity to dipole dynamics, trap density, material degradation, and ultimately PCE retention, framing moisture sensitivity as a quantitative outcome measure that prior reviews have mostly described only qualitatively.

A key technical contribution of the review is its candid assessment of what SCAPS-1D, the widely used one-dimensional solar cell capacitance simulator developed at Ghent University, can and cannot do with moisture. The software numerically solves Poisson’s equation and the electron and hole continuity equations, models up to seven semiconductor layers, and computes quantities including open-circuit voltage, short-circuit current density, fill factor, quantum efficiency, and power conversion efficiency. But moisture-induced degradation cannot be simulated directly. Instead, researchers treat humidity exposure as a driver of assumed changes in trap density, interface recombination velocity, carrier mobility, or band-edge alignment, imposing these as static input perturbations. The review is explicit that SCAPS-1D-based moisture stability studies are therefore parametric sensitivity analyses under assumed degradation states, not first-principles simulations of the degradation process itself.

Where SCAPS-1D handles device physics, artificial intelligence frameworks handle the vast parameter space. Machine learning models predict continuous performance parameters such as PCE, voltage, current density, and fill factor as regression tasks, or discrete stability and degradation classes as classification tasks. The review catalogs several illustrative results. Mustafa and colleagues generated more than 56,000 device configurations for CsPbI3-based all-inorganic cells using SCAPS-1D and applied XGBoost, identifying absorber thickness near 2 micrometers and bulk defect density below 10^15 per cubic centimeter as critical parameters, and reporting a simulated efficiency improvement from 15.15 percent to 19.16 percent with a near-perfect correlation coefficient of 0.999 between simulation outputs and model predictions. Rahmani and colleagues simulated 2,565 configurations of CH3NH3SnX3 and reported a theoretical PCE of 39.68 percent for an optimized tin-based structure, with a random forest model achieving an R2 of 0.98 for efficiency prediction. Other studies reported roughly 89 percent accuracy with random forest approaches and root mean square errors as low as 1.13 percent for XGBoost compared with alternative algorithms.

The review urges caution about the headline numbers, however. Many of the highest reported efficiencies remain simulation-based, depend strongly on dataset quality, and often lack generalization to real device operating conditions. Single-junction tin-based absorbers with bandgaps of 1.2 to 1.4 electronvolts have been reported in some studies with efficiencies of 38 to 40 percent, figures that exceed the Shockley–Queisser detailed-balance limit by a significant margin. The authors treat such values as most plausibly reflecting idealized simulation inputs, such as near-zero defect densities, unrealistically long carrier lifetimes, or ohmic-contact assumptions, rather than physically attainable device performance, and recommend that future SCAPS-1D studies validate reported efficiencies against the Shockley–Queisser limit for the corresponding bandgap as a basic physical sanity check.

The shift toward lead-free absorbers adds another layer of complexity. Driven by toxicity concerns around lead, research over the past two years has moved toward materials such as CsGeI3, CsSnBr3, MASnI3, CsBi3I10, and FAGeI3. But tin-based devices suffer from oxidation of Sn2+ to Sn4+, which increases defect density and recombination losses and accelerates degradation, while non-toxic elements like tin, bismuth, antimony, and manganese bring high densities of bulk and interfacial defects that can act as preferential pathways for chemical decomposition and phase instability. Machine learning models such as XGBoost, random forest, and artificial neural networks largely rely on simulation-derived datasets, further limiting predictive reliability under real-world operating conditions.

Data availability emerges as a bottleneck in its own right. The review categorizes existing datasets by their capacity to support retention-based stability analysis and finds that most prioritize efficiency optimization or materials property prediction, such as bandgap and formation energy, rather than experimental aging or moisture-driven degradation. Datasets like PDP and NOMAD offer rich parametric information but limited retention analysis capability, while the Zenodo perovskite aging dataset, with explicit time-resolved PCE curves and aging metadata, stands out as directly supporting quantitative stability indicators and normalized efficiency decay analysis. The authors call for curating standardized aging datasets and developing humidity-dependent defect density models within SCAPS-1D.

The review’s proposed way forward is a unified conceptual hybrid SCAPS–AI roadmap for moisture-aware device design, combining physics-based simulation with machine learning for parameter sensitivity analysis, predictive optimization, and fast material screening, supplemented by encapsulation strategies, hydrophobic interface design, defect passivation, and compositional engineering. The authors also flag physics-informed AI combined with SCAPS as a route beyond black-box modeling. Their bottom line is that perovskite photovoltaics will become commercially viable not by chasing higher peak efficiencies in simulation, but by proving, with time-resolved data under realistic humidity, that devices can retain their efficiency over years of operation.

Subject of Research: AI-assisted SCAPS-1D modeling of hydrophilic degradation and efficiency retention in perovskite solar cells

Article Title: A systematic review of AI assisted SCAPS 1D modeling for hydrophilic degradation analysis and efficiency retention in perovskite solar cells

Article References: Gupta, T., Bhardwaj, A., & Kumar, A. (2026). A systematic review of AI assisted SCAPS 1D modeling for hydrophilic degradation analysis and efficiency retention in perovskite solar cells. Discover Electronics, 3(1), Article 136. https://doi.org/10.1007/s44291-026-00285-w

Image Credits: AI Generated

DOI: 10.1007/s44291-026-00285-w

Keywords: perovskite solar cells, SCAPS-1D, machine learning, moisture degradation, hydrophilicity, power conversion efficiency, lead-free perovskites, systematic review, PRISMA, device stability, XGBoost, defect density