@Article{cmc.2026.089520,
AUTHOR = {Muhammad Imran, Javed Ferzund, Ateeq Ur Rehman Butt, Abdul Noman, Hessa Alfraihi, Mohamad Khairi Ishak},
TITLE = {Large-Scale Formation Energy Prediction in One Million Crystalline Materials Using Deep Learning and Principal Component Analysis},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/28373},
ISSN = {1546-2226},
ABSTRACT = {Predicting the formation energy of crystalline materials with high accuracy and low computational cost is important for accelerating large-scale materials screening. This study presents a systematic comparison of four deep learning architectures multi-layer perceptron (MLP), deep neural network (DNN), recurrent neural network (RNN), and long short-term memory (LSTM) and three classical machine-learning regressors, namely random forest, support vector regression, and gradient boosting. The models were evaluated using one million DFT-computed crystalline-material records containing nine structural and electronic descriptors, with an 80:10:10 train/validation/test split. Principal component analysis (PCA) reduced the numerical descriptor space from nine variables to six components retaining 95.1% of the variance. Under the original descriptor configuration, the DNN showed the lowest error among the evaluated models, and PCA was associated with a further reduction in error and faster convergence. However, energy_per_atom is included among the predictors and is mathematically related to formation_energy_per_atom; therefore, the very low absolute errors obtained in this configuration may be affected by target leakage and should not be interpreted as leakage-free estimates of generalization performance. Feedforward architectures consistently outperformed recurrent models on the tabular descriptor representation. The reliability of the comparative workflow was further assessed through independent held-out testing, training-set-only preprocessing, repeated-run statistical evaluation, and comparison across seven algorithms spanning deep and classical machine-learning paradigms. The study therefore provides a large-scale comparative benchmark and a preprocessing analysis, while highlighting the need for leakage-controlled descriptors, categorical treatment of crystallographic labels, and source-aware validation in future evaluations.},
DOI = {10.32604/cmc.2026.089520}
}