Quantifying Perovskite Solar Cell Degradation via Machine Learning from Spatially Resolved Multimodal Luminescence Time Series

This paper introduces LumPerNet, a deep-learning framework that accurately quantifies the efficiency retention of perovskite solar cells by analyzing spatially resolved multimodal luminescence images, thereby enabling rapid, non-invasive degradation monitoring and accelerated stability testing.

The Gist

The Big Picture: The "Aging" Problem

Imagine Perovskite Solar Cells (PSCs) as high-performance race cars. They are incredibly fast and efficient (they convert sunlight to electricity better than almost anything else), but they have a reputation for being fragile. Unlike a sturdy Toyota Camry (Silicon solar cells) that can run for 20 years with minimal issues, these race cars tend to break down or lose speed if the weather gets too humid, too hot, or if they are driven too hard.

The big question for scientists is: "How do we know when a race car is about to break down without taking it apart and running expensive, time-consuming diagnostic tests?"

Currently, to check a solar cell's health, engineers have to hook it up to wires and run complex electrical tests (like a doctor running a full blood panel). This is slow, expensive, and doesn't tell you where on the car the engine is sputtering.

The Solution: A "Smart Camera" Doctor

This paper introduces a new way to check the health of these solar cells using AI and special cameras. Instead of wires, they use light.

Think of the solar cell as a glowing jellyfish. When you poke it or shine light on it, it glows.

• Electroluminescence (EL): You give it a tiny electric shock, and it glows from the inside out.
• Photoluminescence (PL): You shine a blue light on it, and it glows back.

The researchers built a custom robot lab that takes photos of these glowing cells over time (5 to 70 hours) as they age. They take three different types of "glow photos" at every step:

1. EL: The "electric glow."
2. PLoc: The "open-circuit glow" (like the jellyfish resting).
3. PLsc: The "short-circuit glow" (like the jellyfish working hard).

The AI: "LumPerNet"

The team trained a computer brain (a Deep Learning model they named LumPerNet) to look at these glow photos and guess how much power the solar cell has left.

The Analogy:
Imagine you are trying to guess how tired a marathon runner is.

• The Old Way (Intensity-Only): You just look at the runner's overall brightness. "They look a bit dim, so they must be tired." This is okay, but it's not very accurate.
• The New Way (LumPerNet): You look at the pattern of their sweat, the way their muscles are shaking, and the specific spots where they are stumbling. You compare their current "glow" to a photo of them when they were fresh at the starting line.

The AI learned that spatial patterns (where the light is dimming or changing shape) tell a much richer story than just the total brightness. It's like noticing that a runner is favoring their left leg (a specific defect) rather than just saying "they look slow."

The Key Findings (The "Aha!" Moments)

1. Location Matters More Than Volume
The researchers found that simply measuring how bright the whole cell is (like measuring the volume of a song) wasn't enough. The AI needed to see the map of the light. By looking at the specific patterns of degradation (like a crack forming in a specific corner), the AI could predict the cell's efficiency with much higher accuracy.

• Result: The AI reduced its prediction errors by about 25% compared to the old "just measure the brightness" method.

2. You Need a Trio of Views
The AI works best when it sees all three types of glow photos (EL, PLoc, and PLsc) together.

• Analogy: It's like diagnosing a patient. If you only listen to their heart (one modality), you might miss a lung issue. If you only look at their skin (another modality), you miss the fever. But if you do an EKG, check their skin, and listen to their lungs all at once, you get the full picture.
• Surprise: Some combinations of views worked better than others. For example, combining the "electric glow" with the "working hard glow" was a winning team, while other combinations were confusing for the AI.

3. It Can Track the "Story" of Aging
The AI doesn't just give a single grade; it can watch the cell age in real-time. The paper shows graphs where the AI's prediction (the blue line) closely follows the actual performance drop (the pink line) as the cell gets older. It successfully predicted that some cells were stable while others were crashing fast, just by looking at the changing glow patterns.

Why This Matters

This is a game-changer for the future of solar energy.

• Speed: Instead of waiting days for electrical tests, you can snap a few photos and get an instant health report.
• Safety: It's non-invasive. You don't have to touch the cell with wires, which prevents accidentally damaging it.
• Scalability: This method can be put on a conveyor belt in a factory. As solar panels roll off the line, a camera can scan them, and the AI can instantly flag the ones that are already starting to degrade or have hidden defects.

The Bottom Line

The researchers built a "smart camera" that learns to read the fading glow of solar cells. By using AI to look at where and how the light changes (rather than just how bright it is), they can predict exactly how much power a solar cell has left. This paves the way for cheaper, faster, and more reliable solar panels that we can trust to last longer.

Technical Summary

1. Problem Statement

Perovskite Solar Cells (PSCs) have achieved remarkable power conversion efficiencies (PCE >27%), but their commercial viability is hindered by operational stability issues. Unlike silicon PV, PSCs are highly sensitive to environmental stressors (moisture, oxygen, UV, heat), leading to complex degradation mechanisms involving ion migration, phase transitions, and interfacial reactions.

Current monitoring methods face two main limitations:

• Electrical Characterization (J-V sweeps): Accurate but time-consuming, destructive to continuous monitoring, and lack spatial resolution (cannot identify localized degradation hotspots).
• Existing Imaging: While Electroluminescence (EL) and Photoluminescence (PL) imaging are used for silicon, their application to PSCs is underexplored. Previous studies often rely on single time-points or global intensity averages, failing to capture the spatially resolved patterns that correlate with specific degradation mechanisms.

The Core Challenge: How to develop a fast, non-invasive, and spatially resolved method to predict the Efficiency Retention ($R_{PCE}$) of PSCs during aging, moving beyond simple global intensity decay to leverage complex spatial degradation signatures.

---

2. Methodology

The authors propose a deep-learning framework named LumPerNet that predicts $R_{PCE}$ directly from multimodal luminescence imaging time series.

A. Experimental Setup & Data Acquisition

• Devices: 59 PSCs across three batches (two batches of Architecture A: CsFAMA; one batch of Architecture B: FACs).
• Aging Protocol: Devices were aged in a custom-built automated platform at 30°C in ambient air (unencapsulated to accelerate variation).
• Multimodal Imaging: At specific time intervals, the system acquired:
  1. EL: Electroluminescence under forward bias (1.5 V).
  2. $PL_{oc}$: Photoluminescence under open-circuit conditions.
  3. $PL_{sc}$: Photoluminescence under short-circuit conditions.
  4. Reference: All modalities were acquired at $t=0$ (initial state) and paired with aged states ($t>0$).
• Ground Truth: $R_{PCE}$ was calculated as $PCE(t) / PCE(0)$ derived from indoor J-V sweeps.
• Dataset: 12,258 paired samples (aged image + reference image) restricted to the operational window $R_{PCE} \in [0.8, 1.2]$ to focus on stable degradation rather than transient spikes.

B. Data Preprocessing

• Correction: Dark-field subtraction, hot-pixel removal, and spatial cropping of Regions of Interest (ROIs) to $72 \times 72$ pixels.
• Input Tensor Construction: For each time step $t$, the input tensor $x(t)$ includes:
  • Absolute images: $EL(t), PLoc(t), PLsc(t)$ and their references $EL(0), PLoc(0), PLsc(0)$.
  • Ratio channels: Pixel-wise ratios $I(t)/I(0)$ to normalize intensity changes and highlight relative spatial degradation.

C. Machine Learning Framework

• Model Architecture (LumPerNet): A compact Convolutional Neural Network (CNN) backbone followed by a Multilayer Perceptron (MLP) regressor.
  • Input: Stacked multimodal image tensors.
  • Mechanism: The CNN extracts spatial features from the reference-current image pairs to learn degradation signatures.
• Baseline Model: A shallow MLP trained only on space-averaged (global) intensities of the same modalities, serving as a control to test the value of spatial resolution.
• Evaluation Protocol:
  • Leakage-Aware: Strict device-level splitting (training/validation on 80% of devices, held-out test on 20% of different devices) to prevent temporal data leakage.
  • Cross-Validation: 4-fold CV on the training set.
  • Metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Coefficient of Determination ($R^2$).

---

3. Key Contributions

1. First Multimodal Time-Series Framework for PSC Degradation: The study introduces a pipeline that links automated, spatially resolved EL/PL time series to quantitative electrical degradation metrics ($R_{PCE}$).
2. Demonstration of Spatial Advantage: It proves that spatially resolved patterns contain predictive information that global intensity averages miss. The CNN significantly outperforms the intensity-only baseline.
3. Comprehensive Modality Ablation: The authors conducted an exhaustive study on single- and paired-modality inputs, revealing that:
   • Multimodality is essential for robust generalization.
   • Complementary Physics matters more than channel count: The combination of EL + $PL_{sc}$ or $PL_{oc}$ + $PL_{sc}$ performs best, while EL + $PL_{oc}$ is unstable. This suggests that combining injection-based (EL) and extraction-based ($PL_{sc}$) signals provides complementary views of recombination and ion migration.
4. Reproducible Pipeline: The work provides a fully open-source workflow (code on GitHub, data on Zenodo) including automated aging, image processing, and model training.

---

4. Key Results

• Performance Superiority:
  • LumPerNet (Multimodal CNN): Achieved MAE = 0.049, RMSE = 0.061, and $R^2 = 0.553$.
  • Baseline (Intensity-only MLP): Achieved MAE = 0.064, RMSE = 0.082, and $R^2 = 0.136$.
  • Improvement: LumPerNet reduced MAE by 23.4% and RMSE by 25.6%, while increasing $R^2$ by 0.417 (approx. 4.1x improvement).
• Robustness: The baseline model showed high variability across folds (negative $R^2$ in 2 out of 4 runs), whereas LumPerNet remained stable, confirming that spatial features are critical for generalizing to unseen devices.
• Modality Ablation Insights:
  • Single modalities (EL-only, $PL_{oc}$-only, $PL_{sc}$-only) all performed significantly worse than the full model.
  • EL + $PL_{sc}$ was the strongest bimodal pair (MAE 0.052), closely matching the full model.
  • EL + $PL_{oc}$ performed poorly (MAE 0.060) and was unstable, indicating that these two modalities probe similar physical states and lack complementary contrast.
• Qualitative Tracking: The model successfully reconstructed time-evolving degradation trajectories for individual devices, capturing slow decay and moderate instability, though it smoothed out transient electrical spikes (which is desirable for stability monitoring).

---

5. Significance and Impact

• Accelerated Stability Testing: This framework offers a non-invasive, high-throughput alternative to repeated J-V sweeps, enabling rapid screening of PSC stability during development and manufacturing.
• Localized Diagnostics: By leveraging spatial resolution, the method can potentially identify where degradation initiates (e.g., edges, contacts, shunts), guiding targeted improvements in device architecture and encapsulation.
• Physics-Informed AI: The results highlight that successful AI for materials science requires understanding the physical complementarity of data modalities (e.g., combining injection and extraction modes) rather than simply stacking more channels.
• Scalability: The approach is designed to be compatible with field deployment (e.g., drone-based imaging) and automated production lines, bridging the gap between laboratory research and commercial utility-scale PV deployment.

In summary, this work establishes that spatially resolved multimodal luminescence imaging, when processed through a leakage-aware deep learning framework, is a powerful tool for quantifying and monitoring the degradation of perovskite solar cells, outperforming traditional global intensity metrics.