Stability follows efficiency based on the analysis of a large perovskite solar cells ageing dataset

Perovskite solar cells (PSCs) have achieved competitive efficiencies (up to 26.1%), indicating strong potential for large-scale implementation. However, their operational lifetime remains significantly below that of silicon devices (by roughly an order of magnitude), even when common environmental stressors are excluded. Recent efforts have applied machine learning to understand PSC stability, with the Perovskite Database Project aggregating over 42,400 data points; yet less than 20% include degradation data and the quality and homogeneity of these data limit the utility of supervised ML. Prior statistical work proposed normalizing stability metrics (e.g., projecting T_sso to a reference T_sso_m) using acceleration factors, but uncertainties from stressor co-dependencies and parameter ranges, as well as the neglect of degradation-curve shapes, reduce accuracy. To address these gaps, the present study analyzes a large, homogeneous, in-house MPPT ageing dataset collected under controlled conditions. The primary research questions are: (1) how the maximum power conversion efficiency (PCE) achieved during the first 150 hours relates to relative PCE loss after 150 hours, and (2) whether degradation-curve shapes, derived via unsupervised clustering, correlate with efficiency and stability. The goal is to clarify stability–efficiency relationships and to inform the search for more universal lifetime metrics for PSCs.

The study situates itself within a growing literature applying data-driven methods to PSC stability. The Perovskite Database Project compiled >42,400 PSC entries but with limited and heterogeneous degradation data (~7,500 points), which constrains robust supervised learning. Graniero et al. highlighted that incomplete and low-quality degradation records impede ML model training and recommended improving data completeness and quality over simply increasing quantity. Zhang et al. proposed a stability metric T_sso_m by projecting T_sso measurements to standardized reference conditions (300 K, 20% RH, 1 sun) via assumed acceleration factors accounting for temperature, humidity, and illumination; however, uncertainties in stressor interdependencies and parameter ranges, and the lack of consideration of curve-shape diversity, limit precision and generality. Collectively, this literature underscores the need for high-quality, homogeneous ageing datasets and analysis frameworks that account for diverse degradation trajectories when assessing PSC stability.

Dataset and conditions: The authors compiled 2,245 MPPT ageing curves from devices fabricated in one laboratory (HySPRINT, Helmholtz-Zentrum Berlin) between August 2019 and August 2022, spanning various perovskite absorbers (e.g., triple cation CsxMAyFAzPbImBrn, CsPbI3, FAPbI3), charge-selective layers (small molecules, polymers, inorganic), contact metals (Ag, Cu, Au), and architectures (502 n-i-p; 1,743 p-i-n). Ageing was performed under continuous 1-sun illumination in a custom High-throughput Ageing System with individual MPP tracking (perturb-and-observe, 1 s delay, 0.01 V step), measuring PCE at 2-minute intervals. Device temperatures were actively controlled at 25, 45, 65, or 85 °C via Peltier elements, typically under flowing nitrogen (some tests in air). Most devices were unencapsulated (2,220), with 25 encapsulated (glass-to-glass). A UV filter (<380 nm) was used in 475 tests. Light intensity was stabilized using a silicon irradiance sensor calibrated with a KG3-filtered, Fraunhofer-ISE-certified reference cell. Tests followed ISOS-L-1 or ISOS-L-2 protocols. Data selection: For comparability, only the first 150 hours of each MPPT experiment were analyzed. Curves reaching their maximum efficiency after 150 hours were excluded. Devices were grouped into five equal-size bins (449 cells each) by the maximum PCE reached within the first 150 hours: <10.4%, 10.4–14.2%, 14.2–16.8%, 16.8–19.2%, and >19.2%. Pre-processing: MPPT time series were resampled to 10-minute intervals, interpolated using Akima interpolation, normalized to the maximum PCE within the first 150 hours (MaxAbsScaler: x/max(x)), and smoothed with a Savitzky–Golay filter (window length = 71) using SciPy. Stability metric: For each cell, the relative change in PCE after 150 hours with respect to its maximum was computed as ΔPCE_rel = (Maximum PCE − PCE at 150 h) / (Maximum PCE). This metric captures the relative efficiency loss over 150 hours referenced to the device’s maximal observed capability, accommodating both initial burn-in decays and initial light-soaking gains. Statistical analysis: The means and interquartile ranges (IQRs) of ΔPCE_rel were evaluated across maximum-PCE groups. A linear regression (scikit-learn) related group-mean ΔPCE_rel to group-mean maximum PCE. Unsupervised clustering of degradation shapes: Self-organizing maps (SOM, MiniSOM package) clustered normalized PCE-vs-time curves into four shape clusters, selected via elbow analysis of quantization error: (1) initial gain, (2) slow exponential decay, (3) medium exponential decay, and (4) fast exponential decay. SOM parameters included sigma = 0.5 and learning rate = 0.1. Results were cross-checked with k-means clustering (scikit-learn). Cluster distributions were analyzed as a function of maximum-PCE group, and ΔPCE_rel distributions were examined within each cluster and PCE group.

- Across five maximum-PCE groups, the mean ΔPCE_rel decreased as maximum PCE increased, indicating that higher-efficiency devices typically experienced smaller relative losses over the first 150 hours. A linear regression on group means showed that for every 1% increase in maximum PCE, ΔPCE_rel decreased by approximately 1.5%. - SOM clustering identified four dominant degradation-curve shapes with counts: cluster 1 initial gain (n = 1,324), cluster 2 slow exponential decay (n = 722), cluster 3 medium exponential decay (n = 237), and cluster 4 fast exponential decay (n = 97). Four clusters were optimal per elbow analysis. - The fraction of initial gain (cluster 1) increased with higher maximum PCE and exceeded 50% in the highest-efficiency groups, whereas the fast exponential decay (cluster 4) fraction decreased with increasing maximum PCE and was absent in the >19.2% group. - Cluster 4 exhibited the largest ΔPCE_rel values, frequently near 100% (i.e., near-total degradation over 150 h). In contrast, initial gain curves generally showed lower ΔPCE_rel, suggesting improved short-term stability. - The observed distribution of ΔPCE_rel within maximum-PCE groups is a composite of contributions from different curve-shape clusters, explaining the broader spread in the lowest-efficiency group where all cluster types are represented. - The results imply that early detection of an initial gain curve shape during the first hours of ageing may serve as a practical indicator of improved stability over the first 150 hours.

The analysis addresses the central question of whether device efficiency correlates with operational stability in PSCs. The statistically significant trend—higher maximum PCE coinciding with lower relative loss after 150 hours—suggests that, at a fleet level across diverse stacks and materials, improving efficiency tends to co-occur with improved short-term stability. Two high-level causal hypotheses are proposed: (1) Energy conservation model: Lower-efficiency devices leave more of the absorbed solar energy unextracted, which is then dissipated as heat or retained as excess charges within the device. This surplus energy/charge can activate or accelerate degradation reactions (e.g., nonradiative recombination heat, ion migration, charge-induced phase separation). Quantitatively, under 1 sun and for bandgaps typical of triple-cation perovskites, a 10% efficient device may leave ~366 W/m^2 potentially available to drive degradation, versus ~266 W/m^2 for a 20% device. This view aligns with the empirical observation that MPP operation (maximal energy extraction) is a milder stress than J_SC or V_OC operation for PSC stability. (2) Common-cause defects: Poor initial device quality (e.g., pinholes, residual solvent, defects) can simultaneously depress initial PCE and predispose devices to faster degradation by providing initiation sites for failure. In this view, the same defects that lower efficiency also reduce stability. Importantly, the authors caution that the finding should not be interpreted as a universal design rule; specific layers or architectures may enhance efficiency at the expense of longevity and vice versa. Given the heterogeneity of device stacks and degradation mechanisms in the dataset, the analysis is intentionally high-level and does not assign specific physical mechanisms to individual trends. The clustering results further connect degradation-shape phenotypes to efficiency and stability, reinforcing that curve morphology carries prognostic information. The diversity of curve shapes indicates that commonly used single-parameter lifetime metrics (e.g., T_sso) are not universally applicable, motivating new, shape-aware stability metrics.

Using a large, homogeneous MPPT ageing dataset, the study demonstrates a clear statistical association between higher maximum PCE and lower relative efficiency loss over 150 hours, indicating that improvements in PSC efficiency can, at a fleet level, align with better short-term stability. The work also shows that degradation-curve shapes—identified via SOM clustering—are linked to both efficiency and stability: initial gain shapes are more prevalent in high-efficiency devices and exhibit lower ΔPCE_rel, while fast exponential decay shapes are largely absent in the highest-efficiency group and show near-total degradation. These insights encourage simultaneous optimization of efficiency and stability and suggest that early-time curve-shape features can serve as practical indicators of stability. The findings further highlight the need for new, possibly multi-parameter, shape-aware stability metrics that are robust across diverse ageing behaviors. Future work should investigate mechanistic underpinnings of the efficiency–stability relation, extend analyses beyond 150 hours and across broader operating conditions, and develop predictive models that integrate curve-shape features to forecast long-term stability.

- The analysis is intentionally high-level across a heterogeneous set of device stacks, implying multiple, intertwined degradation mechanisms; results cannot be directly tied to specific physical failure modes. - The observed efficiency–stability correlation is statistical and not a universal design rule; certain materials or architectures may trade efficiency against stability. - The study focuses on the first 150 hours of ageing for comparability, excluding devices that reach maximum efficiency after 150 hours; longer-term behavior and late-onset failure modes are not assessed here. - Tests were performed under controlled laboratory conditions (e.g., MPP tracking, controlled temperature, inert atmospheres for most tests), which may not capture field or encapsulated-device behavior comprehensively. - Most devices were unencapsulated, which can influence degradation pathways relative to commercial encapsulated modules.