🔬 physics

AI-Optimized Graded CsPbBr₃/MoS₂ Photodetectors for Broadband UV-Visible Detection

This study introduces an XGBoost-multi-objective Bayesian Optimization framework to design functionally graded CsPbBr₃₋ₓClₓ/MoS₂ heterostructure photodetectors, achieving record-breaking broadband UV-Visible performance metrics while reducing computational costs by 96.8% compared to traditional simulation methods.

Original authors: V Rama Krishna, Mamata Choudhury, M B S Sreekara reddy, K. S. Raghuram, S. N. Padhi

Published 2026-08-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: V Rama Krishna, Mamata Choudhury, M B S Sreekara reddy, K. S. Raghuram, S. N. Padhi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the world of light as a massive, colorful orchestra. Some instruments play deep, warm notes (infrared), others play bright, high-pitched squeaks (ultraviolet), and the middle range is the visible spectrum we see with our eyes. For decades, scientists have been trying to build a single "super-microphone" (a photodetector) that can hear every single note in this orchestra with perfect clarity. The problem is that most microphones are tuned to only one section; a microphone great at hearing the deep bass notes often misses the high squeaks, and vice versa.

Enter the world of "perovskites." Think of these as a special type of Lego brick made of atoms that can be snapped together in different ways to catch different colors of light. When you mix these bricks with a super-thin, atomically flat sheet of material called MoS₂ (which acts like a super-fast highway for electrons), you get a device that could hear the whole orchestra. But here's the catch: building these devices is like trying to bake the perfect cake by guessing the amount of sugar and flour. If you get the recipe slightly wrong, the cake is either too dry or too sweet, and the device stops working well. Scientists have been stuck in a cycle of "guess and check," trying to find the perfect mix of ingredients to catch both UV and visible light at the same time. This is where the story of this new research begins: using a digital "smart chef" to solve the recipe problem.

The Digital Chef and the Perfect Recipe

In this study, a team of researchers decided to stop guessing and start using Artificial Intelligence (AI) to design the perfect light-catching device. They focused on a specific type of photodetector made from a mix of two materials: a perovskite called CsPbBr₃ (which is great at catching visible light) and a 2D material called MoS₂ (which helps move the electricity around quickly).

The big idea they wanted to test was something called a "Functionally Graded Material." Imagine a chocolate bar where the flavor slowly changes from milk chocolate on one end to dark chocolate on the other, rather than being the same flavor throughout. The researchers wanted to create a perovskite layer where the chemical "flavor" (specifically the amount of Chlorine mixed in) slowly changes from one side of the device to the other. They hoped this "gradient" would create an invisible electric push that helps catch more light and move it faster.

However, figuring out exactly how to change that flavor (how steep the gradient should be, how thick the layers should be, and what temperature to bake them at) is incredibly complex. There are too many variables to test one by one in a real lab. So, the team built a digital twin of the device using a simulation program called SCAPS-1D. They ran this simulation 2,200 times, each time tweaking the ingredients slightly, to create a massive library of "what-if" scenarios.

The AI Super-Brain

Once they had this library of 2,200 simulated devices, they trained a powerful AI brain called XGBoost. Think of XGBoost as a super-smart student who read all 2,200 simulation results and learned the patterns. Instead of needing to run a slow, heavy simulation every time, the AI could instantly predict how a new design would perform just by looking at the recipe.

But the researchers didn't just want any good device; they wanted the best possible device that balanced several competing goals: catching as much light as possible, being very sensitive, and responding quickly. To find this "perfect balance," they used a second AI tool called MOBO (Multi-Objective Bayesian Optimization). You can think of MOBO as a treasure hunter with a map that shows where the gold is. It didn't just look for one type of treasure; it looked for the spot where you get the most gold, the most silver, and the most jewels all at once, without sacrificing one for the others.

The Winning Design

After the AI did its work, it pointed to a specific "Pareto-optimal" design—a recipe that was unbeatable. Here is what the AI found:

The Gradient: The AI determined that the Chlorine content should change in a specific, curved pattern (an exponential curve) rather than a straight line. This created a built-in "quasi-electric field" of 4.2 × 10⁴ V/cm. Imagine this as a gentle but constant wind blowing inside the device, pushing the light-generated electrons toward the exit so they don't get lost or stuck.
The Performance: The resulting device was a powerhouse. It could detect light at 365 nm (UV) with a responsivity of 1.2 × 10⁴ A/W and at 520 nm (Visible) with 8.4 × 10³ A/W.
The Speed: It was incredibly fast, with a response time of just 0.82 ms.
The Sensitivity: It had a specific detectivity (D*) of 6.8 × 10¹³ Jones, which is a measure of how well it can hear a whisper in a noisy room.

To put this in perspective, the researchers compared their AI-designed device to the best "uniform" devices (where the material is the same throughout) found in previous studies. Their new design was 15 times better at catching UV light, 84 times better at sensitivity, and 78% faster in response time. It also expanded the range of light it could see, covering everything from 300 nm to 600 nm, whereas older devices were limited to a narrower slice of the spectrum.

Why It Matters (And What It Doesn't Do)

The researchers used a tool called SHAP to peek inside the AI's "brain" and understand why this design worked so well. They found that the steepness of the Chlorine gradient was the most important factor for catching UV light, while the thickness of the MoS₂ layer was the key to keeping the "dark current" (unwanted noise) low.

It is important to note that these results come from simulations. The paper presents a highly detailed, validated computer model that matches real-world experimental data from other studies very closely (with deviations of less than 7%). The authors have not yet built this specific AI-optimized device in a physical lab, but the simulation suggests it is physically possible and would outperform everything currently published.

The study concludes that this "AI-Optimized" approach is a game-changer. By using AI to navigate the complex design space, they reduced the computational cost of finding the best design by 96.8% compared to traditional methods. This means that in the future, scientists won't need to spend years baking and testing different cake recipes; they can let the AI find the perfect one in a fraction of the time, paving the way for next-generation light detectors that can see the full spectrum of light with incredible speed and clarity.

Technical Summary: AI-Optimized Graded CsPbBr₃/MoS₂ Photodetectors for Broadband UV-Visible Detection

Problem Statement
All-inorganic cesium lead halide perovskites (CsPbX₃) combined with 2D MoS₂ offer high potential for photodetection, yet existing CsPbBr₃/MoS₂ heterostructures are limited by uniform halide composition. This uniformity restricts the absorption edge to ~520 nm, preventing effective broadband UV-Visible operation. While functionally graded materials (FGM) with varying halide compositions could theoretically extend UV absorption and enhance carrier extraction via built-in quasi-electric fields, no systematic strategy exists to optimize these graded profiles. Furthermore, the multi-dimensional design space involving halide gradients, layer thicknesses, and defect densities is too complex for traditional parametric sweeps, and no AI-driven framework has been applied to this specific heterostructure system.

Methodology
The study introduces a novel XGBoost-multi-objective Bayesian Optimization (MOBO)-SHAP framework to optimize functionally graded CsPbBr₃₋ₓClₓ/MoS₂ photodetectors. The methodology proceeds in three stages:

High-Fidelity Dataset Generation: A dataset of 2,200 device simulations was generated using SCAPS-1D (Solar Cell Capacitance Simulator in 1 Dimension) via Latin Hypercube Sampling. The simulations modeled a planar heterojunction (ITO/Spiro-OMeTAD/FGM Perovskite/MoS₂/Au) across six design variables:
- Cl gradient steepness ( $\alpha$ )
- Perovskite thickness ( $d$ )
- MoS₂ thickness (1–3 monolayers)
- HTL thickness
- Defect density ( $N_{def}$ )
- Annealing temperature
  The model incorporated an exponential Cl composition profile $x(z)$ to create a graded bandgap from 2.36 eV (CsPbBr₃) to 3.03 eV (CsPbCl₃).
Surrogate Modeling (XGBoost): Six separate XGBoost (eXtreme Gradient Boosting) models were trained to predict key performance metrics: UV responsivity ( $R_{UV}$ ), Visible responsivity ( $R_{Vis}$ ), specific detectivity ( $D^*$ ), external quantum efficiency (EQE), response time ( $\tau$ ), and dark current density ( $J_{dark}$ ). The models achieved high fidelity with $R^2 > 0.995$ on independent test sets.
Optimization and Interpretability:
- MOBO: Multi-objective Bayesian Optimization using Expected Hypervolume Improvement (EHVI) was applied to the XGBoost surrogate to identify Pareto-optimal designs that simultaneously maximize responsivity and detectivity while minimizing response time and dark current.
- SHAP: SHapley Additive exPlanations (SHAP) TreeExplainer was used to interpret the XGBoost model, quantifying the contribution of each design variable to the output performance.

Key Contributions

First AI Framework for this System: This study presents the first application of an XGBoost-MOBO-SHAP framework specifically for optimizing functionally graded CsPbBr₃₋ₓClₓ/MoS₂ heterostructures.
Computational Efficiency: The framework reduces the computational cost of optimization by 96.8% compared to full SCAPS-1D parametric sweeps, requiring only 42 XGBoost evaluations to converge versus 2,200 full simulations.
Mechanistic Insight: The study provides a physics-based explanation for performance enhancements, linking specific design parameters to device physics (e.g., quasi-electric fields and band alignment) through SHAP analysis.

Results
The AI-optimized design yielded a Pareto-optimal device with the following specifications:

Composition Profile: Exponential Cl gradient with steepness $\alpha = 3.2$ .
Structural Parameters: Perovskite thickness of 380 nm, MoS₂ thickness of 1.3 nm (2 monolayers), and HTL thickness of 75 nm.
Performance Metrics:
- UV Responsivity ( $R_{UV}$ at 365 nm): $1.2 \times 10^4$ A/W
- Visible Responsivity ( $R_{Vis}$ at 520 nm): $8.4 \times 10^3$ A/W
- Specific Detectivity ( $D^*$ ): $6.8 \times 10^{13}$ Jones
- External Quantum Efficiency (EQE): 91.4%
- Response Time ( $\tau$ ): 0.82 ms
- Detection Range: 300–600 nm (Broadband UV-Visible)

Comparative Performance:
Compared to uniform CsPbBr₃/MoS₂ devices, the AI-optimized FGM device demonstrated:

15× improvement in UV responsivity.
2.7× improvement in Visible responsivity.
84× improvement in specific detectivity.
78% reduction in response time.
13× improvement in noise equivalent power (NEP).

Significance and Claims
The paper claims that this work establishes a new benchmark for CsPbBr₃-based photodetectors, surpassing all previously published experimental results in terms of responsivity, detectivity, and spectral range. The authors assert that the Cl gradient steepness is the dominant parameter for UV responsivity (mean |SHAP| = 0.44), as it creates a built-in quasi-electric field of $4.2 \times 10^4$ V/cm that drives photogenerated carriers toward the MoS₂ electron transport layer. Conversely, MoS₂ thickness is identified as the primary driver for dark current suppression (0.42) by enhancing band alignment.

The study concludes that the XGBoost-MOBO-SHAP framework provides a practical, interpretable platform for designing next-generation broadband perovskite-2D photodetector arrays, offering specific process guidelines (e.g., annealing at 220°C, specific layer thicknesses) for fabrication. Future work is proposed to include experimental validation of these AI-optimized profiles and extension to NIR photodetectors.

The Digital Chef and the Perfect Recipe

The AI Super-Brain

The Winning Design

Why It Matters (And What It Doesn't Do)

Technical Summary: AI-Optimized Graded CsPbBr₃/MoS₂ Photodetectors for Broadband UV-Visible Detection

More like this