AI Deep Learning for Perovskite Solar Cell Review: A Comprehensive Analysis

Introduction

Traditional single-crystalline silicon solar cells, despite significant efficiency improvements over five decades, face limitations in fabrication and inflexibility. The escalating global demand for green energy has spurred research into novel photovoltaic materials. Metal-halide perovskites have emerged as promising candidates, potentially offering equivalent or superior power conversion efficiency (PCE) through cost-effective fabrication processes. This review delves into the application of artificial intelligence (AI), specifically deep learning, in advancing perovskite solar cell (PSC) technology.

Perovskite Solar Cell Structure and Composition

A typical PSC device structure consists of several layers, including an active layer composed of metal-halide perovskite with an ABX3 structure. In this structure, 'A' represents an organic/inorganic cation (e.g., methylammonium (MA), formamidinium (FA), cesium (CS)), 'B' is a metal cation (e.g., lead (Pb), tin (Sn)), and 'X' is a halide anion (e.g., iodide (I), bromide (Br), chloride (Cl)). Novel perovskite compositions are created by changing or mixing ions at these sites. Additional layers can be incorporated to reduce Auger recombination or enhance electron-hole transport efficiency.

The Shockley-Queisser Limit and Non-Radiative Recombination

The power conversion efficiency (PCE) of PSCs remains below the Shockley-Queisser (SQ) limit. The SQ theory assumes purely radiative recombination pathways within a single bandgap PSC. Non-radiative recombination pathways in the perovskite layer (Shockley-Read-Hall recombination) and at the interfaces between layers contribute to this disparity. Improving existing solar cell designs requires understanding how ionic composition and additives interact with the electronic properties of each layer.

The Role of Machine Learning in PSC Development

Traditionally, improvements to solar cell designs relied on researcher intuition and trial-and-error experimentation, which is time-consuming. Machine learning (ML) and Big Data now offer the possibility of expediting this process through the analysis of large datasets. ML is better suited for modeling the non-linearities among the many features found in perovskite solar cells when compared to other modeling techniques like density functional theory (DFT) or molecular dynamics. ML analyses can range from pattern recognition to predicting key performance metrics like short-circuit current density (Jsc), open-circuit voltage (Voc), PCE, and fill factor (FF). The quality and quantity of data are critical for ML model performance.

Data Challenges and the Perovskite Database Project (PDP)

Previous research using ML to predict PSC performance required creating custom training sets from literature data. This made it difficult to replicate or compare performances across models. Creating an ML pipeline based on human-collected data necessitates a database devoid of erroneous data points, formatted for ML readability. PSC data is varied, with materials, chemical ratios, and manufacturing processes taking numeric and categorical forms.

To address these challenges, the open-source Perovskite Database Project (PDP) was created. The PDP is a large, open repository of PSC device information with detailed documentation, making it an ideal starting point for an open-source ML pipeline. The PDP includes perovskite compositional, additive, solvent, and depositional information, enabling ML models to learn mappings between relevant PSC qualities and overall device performance. Studies have already emerged using the PDP for ML-based PSC performance predictions. For example, researchers have experimented with synthetic data generation techniques in conjunction with ML to predict the PCE of PSCs, while others have used ML to identify superior configurations for Sn-based PSCs. These studies highlight the importance of using ML to expedite the PSC design process.

A Modular ML Pipeline for PSC Analysis

A modular and flexible workflow for ML, adaptable for any feature set or target within the PDP, is essential. This pipeline should account for the layered nature of solar cell data and handle common data errors. Representations of features should be chosen with model interpretability in mind, establishing clear relationships between granular quantities controlled in a laboratory setting (e.g., perovskite composition, deposition solvents and their ratios) and performance. This provides actionable and physically interpretable information for research and development.

Implementation of the ML Pipeline

An ML pipeline can be implemented in conjunction with automated hyperparameter tuning to build XGBoost, ANN, and Gaussian process regressor (GPR) models. Data is sourced from the PDP, and relevant raw columns are selected. Data is then extracted, cleaned, and split into training, test, and validation sets. The training set is used by the Optuna search algorithm, which performs automated hyperparameter tuning for the Neural Network and XGBoost models. A portion of the training set is used as a "verification set" in conjunction with a custom error metric for Optuna to conduct searches for candidate best models. Manual hyperparameter tuning can be used to generate GPR and ANN model candidates. The candidate models are tested on the validation set to identify the best-performing models for XGBoost, ANN, and GPR architectures.

Data Selection and Preprocessing from the PDP

Data can be selected from the PDP, consisting of tabular data recording device parameters, electrical characteristics, fabrication methods, and compositional information for perovskite solar cells. A subset of the data, containing entries published between specific dates, can be selected for cleaning. Limiting the data range can improve data quality, as earlier data may reflect a less mature technology, while more recent crowd-sourced data may contain a higher frequency of errors.

For the target variable, the short-circuit current density Jsc can be chosen because of its fundamental nature. Unlike PCE or FF, Jsc is agnostic to the load of any additional circuitry, allowing it to serve as a stand-in metric for other utilizations of perovskite, such as LEDs, which have similar device structures.

The selection of relevant data involves grouping raw columns into categories such as "Perovskite Deposition," "Perovskite Composition," "Transport Layer Characteristics," "Backcontact," "Electrical Parameters," and "Substrate." This selection can be motivated by literature and the correlation of the data contained in the groups with the target Jsc, where feature groups with higher correlation are considered favorable for training. Missing or unknown data can be set to NaN values. To ensure sufficient data for model training, columns that contain a sufficient number of non-NaN entries can be selected.

Feature Encoding and Data Cleaning

The PDP contains numerical, boolean, and categorical features. To convert raw columns from the PDP into a format readable by an ML model, all boolean and categorical features must be converted into numerical values. Boolean values can be represented by 1 for True and 0 for False. For categorical features, one-hot encoding can be used to represent the presence of a category with a 1 or absence with a 0. A "two-column" encoding strategy can be employed for categorical columns paired with associated numerical columns to quantify the category, such as "Perovskitecompositionaions" and "Perovskitecompositionaions_coefficients." To encode the layered nature of data, entries can be split into layered groups, and the same encoding strategy can be applied with the addition of an organizational tag denoting the layer group.

Throughout the encoding process, data cleaning steps are crucial to ensure reasonable results for training. For example, in two-column encoding of perovskite composition, there must be the same number of entries in the compositional and coefficient columns. In addition, if the material has multiple layers, the same number of layers must be listed across the A, B, and C site coefficients to be considered valid. After encoding all columns, NaN entries can be removed, depending on the model being used. A statistical cleaning step can be undertaken, where each encoded column's mean and standard deviation are calculated. If any row has an entry greater than three standard deviations away from its respective column's mean, that row is removed. The final cleaning occurs post duplicate removal.

Machine Learning for Perovskite Photodetectors

Perovskite materials have emerged as promising candidates for photodetectors (PDs) due to their high photoelectric conversion efficiency, broad spectral response, tunable optical properties, low cost, flexibility, and lightweight. Machine learning (ML) is being used to predict the performance of PDs and design more efficient devices. Researchers are gradually turning their attention to the field of combining ML and perovskite PDs.

For example, a comprehensive dataset was compiled by reviewing publications focused on high-performance 2D MHPs PDs. Leveraging this dataset, ML techniques were used within a scientific search network framework to predict key performance metrics: responsivity and detectivity. This analysis revealed a consistent trend, where increased nitrogen content in spacer cations inversely correlates with responsivity but positively impacts detection rates. In another study, a vertical matrix perovskite X-ray detector was designed for multi-energy detection based on the attenuation behavior of X-rays inside the detector and a ML algorithm.

Read also: An Overview of Deep Learning Math

The Workflow of Machine Learning for Perovskite Devices

The main purpose of ML is to classify data according to different algorithms to achieve the evaluation of the performance of optoelectronic devices and the prediction of material screening. The general workflow of ML is divided into data collection, data preprocessing, feature engineering, model selection, model training, and model validation.

Data Collection

Data collection is a key step, which determines the quality of the data underlying the model and the effectiveness of the subsequent work. Data can be collected through authoritative databases, public datasets, or experiments. Data types can be structured (e.g., Excel files, CSV files) or unstructured (e.g., images, texts). In addition to this approach, data can also be obtained from experiments to build datasets. One can enrich oneâs database by retrieving data from articles through a literature search. Computationally generated data can also be used to enhance experimental results. It is important to unify the data in a suitable ML algorithm format.

Data Preprocessing

Data preprocessing is an essential step, which aims to transform raw data into a data format suitable for model training, improve model performance, and ensure the reliability of the results. This step includes data cleaning, data conversion, data encoding, and data segmentation. Data cleaning is to remove some duplicate values or outliers, and data conversion is to unify the data into the same standard and specification. Data segmentation involves dividing the data into training, validation, and testing sets. Some of the datasets consisting of images also require operations such as cropping, adjusting pixels, and color modification.

Feature Engineering

Feature engineering is defined as the selection of a suitable feature for a target after data processing. The success of any ML approach aimed at predicting material properties hinges on the meticulous selection of a descriptor set. Descriptors should be comparable and be able to characterize the system with a high degree of recognition and accuracy for different response variables. Conventionally, descriptors are formulated based on the intrinsic characteristics of a material, including ionic radii among others. To mitigate the risk of dimensionality explosion and ensure effective model training without succumbing to overfitting, it is imperative to maintain a descriptor count that is subordinate to the sample size in the dataset. Furthermore, the selected features must uniquely encapsulate the pertinent information, while simultaneously purging any redundant data, thereby enhancing the modelâs ability to extract meaningful insights. There are three algorithms for feature selection in feature engineering: filtering, packing, and embedding. To prevent a feature from being over-represented, the features are also processed by normalization or normalization.

Model Selection

ML methodologies are typically categorized into three distinct paradigms: supervised learning, unsupervised learning, and reinforcement learning. Within the realm of supervised learning, a further differentiation exists between regression and classification techniques, each serving unique purposes within the domain of data analysis and prediction.

Machine Learning in Perovskite Optoelectronic Devices

ML is being used to tackle the practical challenges within perovskite optoelectronic devices, including the optimization of perovskite active layers, the selection of transport layers, and the elucidation of underlying functional mechanisms. Interpretable ML methodologies are increasingly being applied to analyze and comprehend the factors affecting the performance of perovskite devices.

Machine Learning Workflows for Perovskite Devices

ML workflows usually include data collection, feature engineering, model selection, and performance evaluation. Data collection is the foundation of ML, where the key lies in selecting appropriate data sources and ensuring the accuracy of the data. In the perovskite optoelectronic devices, data typically originate from laboratory measurements, as well as results from computational simulations. Following data collection, feature engineering is undertaken with the objective of converting raw material data into a format that ML algorithms can interpret and utilize effectively. Subsequently, the selection of ML models is contingent upon factors, such as the complexity of the task at hand and the specific characteristics of the dataset. Performance evaluation is the process of assessing an ML modelâs predictive capability and accuracy.

Machine Learning Algorithms

ML algorithms cover a spectrum of approaches including supervised, unsupervised, and semi-supervised learning techniques. Supervised learning is an ML method that involves training a model using labeled data sample. Unlike supervised learning, unsupervised learning is an ML method that focuses on uncovering patterns, structures, or relationships in unlabeled data. Semi-supervised learning is an ML method that combines supervised learning and unsupervised learning. It leverages a combination of labeled and unlabeled data during the training process.

Machine Learning for Perovskite Solar Cells

PSCs represent a promising photovoltaic technology due to their superior optical absorption properties and charge transport capabilities. ML has already been applied to various aspects to enhance the stability and power conversion efficiency (PCE) of PSCs, such as the device structure, perovskite layer, transport layer, and interface engineering.

tags: #AI #deep #learning #perovskite #solar #cells