AI-supported Degradation Study of Carbon-based Perovskite Solar Cells: Learning the Device Physics of Perovskite Solar Cells: A Drift-Diffusion Guided Autoencoder Approach

This paper presents an AI-supported degradation study of carbon-based perovskite solar cells that utilizes a drift-diffusion guided autoencoder to extract physical parameters from scan-speed-dependent J-V curves, enabling in situ tracking of device aging and the creation of a digital twin to bridge the gap between classical interpretation and machine learning predictions.

The Gist

The Big Picture: Teaching a Computer to "X-Ray" a Solar Cell

Imagine you have a brand-new, high-tech solar cell. It's like a tiny, complex city where electricity is generated. Over time, this city starts to crumble (degrade) due to heat and light, just like a real city might suffer from potholes, traffic jams, or crumbling bridges.

Usually, to see what's broken inside, you have to tear the city apart (take the device apart) or stop it from working to run a deep diagnostic. But the researchers in this paper wanted to watch the city degrade while it was still running, without stopping it.

They used Artificial Intelligence (AI) as a "super-spy" to look at the solar cell's output and guess exactly what was happening inside, down to the microscopic level.

---

The Setup: Two Solar Cells in a Stress Test

The researchers took two carbon-based solar cells and put them under a "heat lamp" (simulated sunlight) for 23 days.

• Cell A (The MPP Tracker): This cell was forced to work hard, constantly trying to produce maximum power. Think of this as a marathon runner sprinting at full speed.
• Cell B (The VOC Tracker): This cell was held at a standstill, generating voltage but no current. Think of this as a runner holding a heavy weight in place, straining but not moving.

Every day, they took a quick "snapshot" of how the electricity flowed (a J-V curve) at different speeds.

The Problem: The "Black Box" Mystery

When you look at a solar cell's output graph, it looks like a squiggly line. If the line changes shape, you know something is wrong. But what is wrong?

• Did the "roads" get clogged (charge transport issues)?
• Did the "traffic lights" break (recombination issues)?
• Did the "construction crew" get lost (ion migration)?

Traditionally, figuring this out is like trying to guess the contents of a sealed box just by shaking it. You have to make a lot of guesses.

The Solution: The AI "Decoder Ring"

The researchers built a special AI tool called an Autoencoder. Here is how it works, using a metaphor:

1. The Training Phase (The Simulator): Before looking at the real cells, the AI was trained on a massive library of simulated solar cells. They created 50,000 fake solar cells, each with slightly different "broken" parts (some had clogged roads, some had broken traffic lights). The AI learned to look at the output graph of these fake cells and say, "Ah, this graph means the roads are clogged," or "This graph means the traffic lights are broken."
2. The Real Test: Once the AI was a master detective, they fed it the daily snapshots from the real, aging solar cells.
3. The Result: The AI didn't just say "it's broken." It gave a specific diagnosis: "Surface recombination is increasing," or "The mobility of electrons is dropping."

What Did They Discover?

By watching the AI's daily predictions, they found some fascinating stories about how these solar cells die:

1. The "Surface" Problem (The Rusting Fence)
For both cells, the AI detected that the surface recombination velocity was increasing.

• Analogy: Imagine the solar cell is a house. The "surface" is the outside walls. As the cell ages, the walls start to rust and leak. Electrons (the energy) are trying to get out, but they hit the rusty walls and get lost before they can do any work. The AI saw this "rust" getting worse every day.

2. The "Runner" vs. The "Strainer"

• The Hard Worker (MPP Cell): This cell held up surprisingly well. Its performance stayed relatively stable. The AI showed that while the "roads" (mobility) got a bit bumpy, the cell compensated.
• The Strainer (VOC Cell): This cell degraded much faster. The AI noticed a massive drop in performance. It turned out that holding the cell at a standstill (Voltage tracking) actually caused ions (charged particles) to pile up at the interface, creating a traffic jam that blocked the flow of electricity.
  • Analogy: It's like holding a hose nozzle closed. The pressure builds up inside until the hose bursts or the water leaks out the wrong way.

3. The "Digital Twin"
The researchers took the AI's guesses and built a Digital Twin—a perfect virtual copy of the real solar cell.

• They ran the simulation using the AI's predicted numbers.
• The Surprise: The simulation didn't perfectly match the real measurements.
• Why? The AI's model was a 1D (flat) model, but the real solar cell is a 3D sponge (mesoporous). The AI realized that the "surface" it was measuring wasn't just a flat wall; it was a massive, complex sponge where the "rust" was happening on millions of tiny internal surfaces. The AI had to invent a "super-rust" number to explain the data because the real world was more complex than the simple model.

The Takeaway

This paper is a breakthrough because it moves AI from just predicting the future (e.g., "This cell will die in 2 years") to understanding the present (e.g., "The cell is dying because the surface is rusting").

By using AI to act as a real-time X-ray, scientists can now:

1. See exactly how a solar cell is failing without destroying it.
2. Understand that different stress conditions (working hard vs. standing still) cause different types of damage.
3. Use this knowledge to build better, longer-lasting solar cells in the future.

In short: They taught a computer to listen to the "heartbeat" of a solar cell and diagnose its specific illness, helping us cure the disease before the patient (the solar panel) dies.

Technical Summary

1. Problem Statement

Perovskite Solar Cells (PSCs) have achieved remarkable efficiency gains (approaching 27%), yet degradation and stability remain the primary barriers to commercialization. While Artificial Intelligence (AI) has been applied to PSCs, most existing studies focus on predicting overall performance metrics (e.g., Power Conversion Efficiency or T80 lifetime) rather than identifying the underlying physical mechanisms of degradation.

• The Gap: Traditional analysis of Current-Voltage (J-V) curves is often insufficient to isolate specific physical parameter changes (e.g., surface recombination vs. ion migration) during continuous stress testing without interrupting the device.
• The Goal: To develop a method for in situ tracking of fundamental physical parameters during device aging to understand the specific degradation pathways of carbon-electrode-based PSCs.

2. Methodology

The study employs a hybrid approach combining 1D Drift-Diffusion (DD) device simulation with a Machine Learning (ML) Autoencoder (AE) architecture.

A. Device Simulation & Data Generation

• Device Structure: A mesoporous carbon-based PSC structure (FTO/TiO₂/m-TiO₂-MAPI/m-ZrO₂-MAPI/Carbon).
• Parameter Variation: Six key physical parameters were varied randomly across 50,000 simulated devices to create a training dataset:
  1. Surface recombination velocity ($S$)
  2. Charge carrier mobility ($\mu$)
  3. Recombination lifetime ($\tau$)
  4. TiO₂ electron mobility ($\mu_{TiO2}$)
  5. Cation mobility ($\mu_C$)
  6. Ion density ($\rho_{Ion}$)
• J-V Curve Generation: Simulations were run at various scan speeds (0.005 to 100 V/s) in both forward and reverse directions to capture hysteresis and ion migration effects.

B. Machine Learning Architecture (Drift-Diffusion-Guided Autoencoder)

• Model: A Convolutional Autoencoder (Encoder-Decoder).
• Input: 1D arrays of J-V curve current values (interpolated to 0.02 V steps) across multiple scan speeds.
• Loss Function: A dual-term loss function was crucial:
  1. Reconstruction Loss: Minimizes the difference between input and reconstructed J-V curves.
  2. Parameter Constraint: Minimizes the difference between the AE's latent output and the known physical parameters from the simulation. This ensures the latent space maps directly to physical quantities, avoiding "black box" behavior.
• Output: The Encoder predicts the six physical parameters from experimental J-V data.

C. Experimental Setup

• Samples: Two carbon-based PSCs from Solaronix were aged for 23 days (approx. 550 hours) under 1 Sun equivalent illumination.
  • Device 1: Tracked at Maximum Power Point (MPP).
  • Device 2: Tracked at Open Circuit Voltage ($V_{OC}$).
• Measurement: Daily J-V curves were recorded at multiple scan speeds. Light-intensity dependent measurements and Impedance Spectroscopy (EIS) were also performed for validation.

3. Key Results

A. Macroscopic Performance Degradation

• MPP Device: Retained ~90% of initial efficiency. $V_{OC}$ decreased slightly; $J_{SC}$ was relatively stable.
• VOC Device: Showed significant degradation. $V_{OC}$ dropped from 0.88 V to 0.84 V. $J_{SC}$ decreased steeply initially before saturating.
• Hysteresis: Both devices exhibited inverted hysteresis at slow scan speeds, which increased in magnitude over time, indicating ionic redistribution.

B. AI-Driven Parameter Evolution

The Autoencoder successfully tracked the evolution of physical parameters, revealing distinct degradation mechanisms:

1. Surface Recombination ($S$): Increased significantly for both devices.
   • MPP device: Approached ~50 cm/s.
   • VOC device: Approached ~100 cm/s.
   • Interpretation: This suggests severe interface degradation, likely exacerbated by the VOC bias.
2. Carrier Mobility ($\mu$) & Lifetime ($\tau$):
   • MPP Device: Showed a decrease in mobility ($\mu$) and an increase in lifetime ($\tau$), resulting in a relatively constant diffusion length ($L = \sqrt{\mu\tau}$). This correlated with the stable $J_{SC}$.
   • VOC Device: Showed a drastic drop in mobility and a massive increase in estimated lifetime (orders of magnitude higher than TrPL measurements), suggesting the device entered a regime where the 1D model assumptions (or the training data range) were exceeded.
3. Ionic Parameters ($\mu_C, \rho_{Ion}$):
   • The AI estimated minimal changes in ion density ($\rho_{Ion}$), consistent with EIS data.
   • However, the series resistance increased significantly, particularly for the VOC device. The authors attribute this to ion accumulation at the interface (specifically cations near the carbon-perovskite interface) causing field screening and recombination, a phenomenon difficult to capture fully in a 1D model.

C. Model Limitations & Discrepancies

• Simulation vs. Reality: While the AE-predicted parameters could simulate J-V curves that matched trends, absolute values often deviated.
• Cause: The 1D Drift-Diffusion model cannot fully capture the complex geometry of the mesoporous scaffold (which acts as a massive effective interface) or the specific ionic species interactions in real devices. The model assumes uniform layers, whereas the real device has a porous structure where ions move through TiO₂ and ZrO₂ layers.

4. Key Contributions

1. In Situ Physical Tracking: Demonstrated the ability to extract time-resolved physical parameters from J-V curves during continuous stress testing without device interruption.
2. Drift-Diffusion-Guided ML: Validated a specific ML architecture where the loss function is constrained by physical simulation data, ensuring the AI learns physics rather than just curve fitting.
3. Degradation Mechanism Identification: Differentiated between degradation pathways driven by MPP (mild, bulk/interface recombination changes) and VOC (severe interface degradation and ion accumulation).
4. Digital Twin Concept: Created a "digital twin" of the aging devices, allowing researchers to simulate the impact of specific parameter changes on device performance.

5. Significance

• Beyond Efficiency Prediction: This work shifts the AI application in photovoltaics from predicting when a device will fail to explaining why it fails.
• Optimization of Stability: By identifying that VOC tracking accelerates surface recombination and ion accumulation more than MPP tracking, the study provides actionable insights for stability protocols and device design.
• Methodological Advancement: The approach offers a robust framework for "physics-informed machine learning" in materials science, bridging the gap between complex experimental data and simplified physical models. It highlights that while 1D models have limitations, AI can effectively map experimental data to these models to reveal trends that classical analysis misses.

Conclusion: The paper establishes that AI-driven parameter estimation is a powerful tool for decoding the complex physics of perovskite degradation, specifically revealing that surface recombination and interfacial ion accumulation are the dominant failure modes in carbon-based PSCs under different operating biases.