Rapid Parameter Estimation from Photoluminescence Decays of Halide Perovskite Thin Films

Technical Summary: Rapid Parameter Estimation from Photoluminescence Decays of Halide Perovskite Thin Films

Problem Statement
Extracting material parameters from transient photoluminescence (tr-PL) data in halide perovskite solar cells is challenging because the relationship between experimental observables and material parameters (e.g., defect densities, trap depths, capture coefficients) is governed by systems of non-linear differential equations rather than invertible analytical expressions. Traditional fitting methods, such as Markov-Chain Monte-Carlo (MCMC) sampling, require solving these coupled rate equations repeatedly, leading to prohibitive computational costs (often hours) and difficulties in exploring high-dimensional parameter spaces (>5 dimensions). Furthermore, these methods often struggle with parameter degeneracy, where multiple distinct parameter sets yield equally good fits, making it difficult to quantify uncertainty or identify correlations between parameters. Simplified models that neglect physical processes like detrapping (thermal emission) are sometimes used to reduce complexity, but they fail to accurately describe the highly non-exponential kinetics observed in most perovskite films.

Methodology
The authors propose a workflow that combines Bayesian inference with machine learning to accelerate parameter estimation while retaining full physical fidelity. The core components of the methodology are:

1. Physical Model: The study utilizes a zero-dimensional rate equation model describing Shockley-Read-Hall (SRH) recombination via two trap states (acceptor-like) and radiative recombination. Crucially, the model includes detrapping terms, ensuring detailed balance between capture and emission coefficients. The model solves for electron ($n$), hole ($p$), and trapped carrier ($n_{t1}, n_{t2}$) concentrations over time.
2. Data Representation: Experimental tr-PL data is transformed into a plot of differential decay time ($\tau_{diff}$) versus Fermi-level splitting ($\Delta E_F$). This representation preserves carrier density information and allows for the immediate identification of dominant recombination pathways (deep vs. shallow traps) based on the slope of the decay.
3. Surrogate Modeling: To bypass the computational bottleneck of solving differential equations during optimization, the authors train deep neural networks (NN) on a dataset of over $10^4$ (for one trap) and $2.6 \times 10^5$ (for two traps) numerical simulations. These NNs act as surrogate models, mapping material parameters to $\tau_{diff}$ curves with high fidelity ($R^2 > 0.998$) and enabling rapid evaluation (approx. 100x speedup).
4. Optimization and Inference:
   • Initialization: Analytical approximations derived from simplified rate equations are used to generate physics-informed starting values for the optimizer, classifying the decay as shallow or deep trap dominated.
   • Optimization: A Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithm coupled with the NN surrogate model rapidly identifies the best-fit parameter set.
   • Uncertainty Quantification: Instead of full MCMC, the authors perform a local, sliced grid search around the optimum. This generates "corner plots" showing conditional posterior probability densities and pairwise correlations. The likelihood function incorporates uncertainties from measurement noise, numerical optimization, and model error (specifically the neglect of carrier diffusion in the 0D model).

Key Results

• Synthetic Data Validation: Applied to synthetic data for a single shallow trap, the workflow successfully recovered the best-fit parameters. However, the analysis revealed significant practical non-identifiability: while the optimizer found a solution, the posterior distributions for defect density ($N_t$) and electron SRH lifetime ($\tau_{SRH,n}$) were broad, spanning several orders of magnitude. This indicates that multiple parameter combinations can reproduce the same decay curve, even in the absence of noise.
• Experimental Application (85:15 Film): The method was applied to a triple-cation perovskite film ($Cs_{0.05}FA_{0.73}MA_{0.22}PbI_{2.56}Br_{0.44}$) treated with n-octylammonium iodide. The decay was best described by two shallow traps. The workflow achieved a high-fidelity fit ($\epsilon_{rms} \approx 3.0 \times 10^{-2}$) in tens of seconds.
• Parameter Correlations: The sliced corner plots revealed strong correlations between parameters. For instance, the defect density and SRH hole lifetime are coupled such that a wide range of densities can be compensated by adjusting lifetimes to maintain the same decay shape.
• Steady-State vs. Transient Comparison: The authors demonstrated that while both high and low defect density scenarios could fit the transient data equally well, they predicted different steady-state photoluminescence (ss-PL) behaviors. Specifically, the ratio of trap density to the effective density of states ($N_t/n_1$) influences the charge neutrality condition (photodoping vs. no photodoping), leading to distinct relationships between Fermi-level splitting and excitation intensity. Including ss-PL data helps constrain the parameter space, reducing the ambiguity inherent in transient data alone.

Significance and Claims
The paper claims to provide a robust framework for rapid Bayesian parameter estimation in perovskite research. By replacing computationally expensive MCMC sampling with a neural-network-accelerated genetic optimization and local grid search, the authors reduce the time required for uncertainty quantification from hours to tens of seconds. This speed enables the analysis of high-throughput datasets and facilitates the exploration of complex, high-dimensional parameter spaces without sacrificing physical rigor (i.e., retaining detrapping terms).

The authors emphasize that their approach does not merely provide a single "best" parameter value but quantifies the confidence and limitations of the inferred parameters. They highlight that the non-identifiability of certain parameters (like defect density) is an intrinsic property of the differential equations governing the system within the experimentally observable regime, rather than a failure of the fitting algorithm. Consequently, the method shifts the focus from finding a unique solution to understanding the manifold of data-compatible solutions and their correlations. The inclusion of steady-state data is shown to be a critical step in breaking degeneracies that persist when analyzing transient data in isolation.