Page 84 - Read Online

Page 8 of 31 Shu et al. J. Mater. Inf. 2025, 5, 36 https://dx.doi.org/10.20517/jmi.2025.13

[94]
convolutional neural network (CNN) , which is commonly used for image processing, object detection,
and spatial data analysis. CNN excels at automatically learning hierarchical feature representations from raw
data, making them particularly effective for tasks in materials science that involve complex spatial
structures, such as analyzing material microstructures or predicting material properties from images. While
CNN requires large labeled datasets and substantial computational power, its ability to reduce the need for
[95]
manual feature extraction is invaluable in material science applications. Recurrent neural network (RNN)
is designed to handle sequential data and is ideal for time-dependent tasks such as time series forecasting. In
optoelectronic materials, RNN could be used to model the temporal evolution of material properties under
varying conditions, such as light exposure or electrical stress. However, training RNN can be challenging,
especially when dealing with long sequences, due to issues such as vanishing gradients. Generative
[96]
adversarial networks (GAN) have gained prominence for their ability to generate realistic synthetic data,
which is useful for augmenting datasets in material science. GAN can help create new material designs by
generating plausible material configurations based on learned data distributions. However, GAN is
challenging to train and can suffer from instability or mode collapse, which may hinder their effectiveness
in practical applications. Recently, Transformer-based architectures have revolutionized deep learning,
particularly in natural language processing (NLP) and sequence modeling [97-99] . These architectures excel in
handling long-range dependencies, which can be advantageous in materials science for predicting complex
relationships between different material properties or performance over time. Transformers are highly
scalable and flexible, making them an exciting avenue for developing advanced models in optoelectronic
material design. They hold the potential to improve predictive modeling, accelerate the discovery of new
materials, and optimize material performance. Despite their computational cost, Transformer-based models
are increasingly becoming central to many cutting-edge AI applications, and their potential in materials
science is beginning to be realized.

Before a model can be used for prediction, it should be able to adjust its parameters automatically to
improve performance. During model training, an independent dataset is often used to validate the training
results and detect potential overfitting. The validation set, which usually accounts for about 10% of the
available data, monitors the training process, helps select hyperparameters, and prevents overfitting. After
training the model, an external dataset (test set) is used to evaluate the generalization ability of the model.
This test set, comprising about 10% of the available data, is critical for assessing the performance of the
model on unseen data, thereby providing an unbiased evaluation of its predictive ability . The training,
[100]
validation, and testing process in practical applications often repeats multiple times to optimize the model
performance. This process is known as model iteration or model optimization. Through repeated iterations,
the best model parameters and structure can be identified, thereby enhancing the predictive and
generalization capabilities of the model.

Model evaluation and optimization
Upon completion of model training, evaluating the model is a crucial step to ensure its practical
effectiveness. The purpose of evaluation is to measure the predictive performance of the model, including its
accuracy and error [101,102] . The following are common methods for model evaluation and optimization.

Cross-validation (CV) : CV is a widely used model evaluation technique to prevent overfitting. Common
[103]
CV methods include k-fold and leave-one-out CV. In k-fold CV, the dataset is divided into k subsets; the
model is trained on k-1 subsets and validated on the remaining one. This process is repeated k times, with
each subset used once as the validation set. The final evaluation result is the average of the k-validation
results. Leave-one-out CV uses one sample as the validation set, and the remaining as the training set each
time. Although this method is computationally expensive, it is particularly effective for small datasets.
Bootstrap sampling : Bootstrap sampling is a statistical resampling method that involves repeatedly
[104]