Context <p>The discovery of semiconductor materials for photovoltaic and optoelectronic applications is limited by the computational cost of DFT calculations and the requirement for complete 3D crystal structures. Existing deep learning approaches, including graph neural networks (CGCNN, MEGNet, ALIGNN) and transformer architectures (Matformer, CrysCo), require structural information, limiting their applicability to hypothetical materials of which only the composition is known. We developed MGPP-DL (Materials Graph Property Prediction via Deep Learning), a structure-agnostic approach that predicts the band gap and formation energy directly from the chemical composition of the material structure. Evaluated on 389,000 materials from the Materials Project, MGPP-DL achieves unprecedented accuracy with mean absolute errors of 0.0178 eV/atom for formation energy and 0.0619 eV for band gap, representing a 67% improvement over state-of-the-art models. The model demonstrates a robust generalization (R<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$^2$</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mrow /> <mn>2</mn> </mmultiscripts> </math></EquationSource> </InlineEquation> = 0.9869 and 0.9182, respectively) and provides an acceleration of 10,000<InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\times $</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> par rapport à la DFT, allowing the high-throughput screening of millions of hypothetical compositions.</p> Methods <p>Chemical compositions are represented as graphs of completely connected elements with nodes encoding 33 elementary properties extracted via Pymatgen, including electronegativity, atomic radius, and stoichiometric coefficients. Three GNN layers with message passing refine the embeddings of elements (64-dimensional) by multi-scale aggregation. GNN output is processed by two EfficientNet blocks (dimension 128, expansion ratio 4) with squeeze-and-excitation attention mechanisms (reduction ratio 0.25). The training employed the optimizer Adam (learning rate 0.001), MSE loss, dropout (0.3), and batch standardization over 40 epochs (batch size 32). The Materials Project dataset (DFT-GGA, functional PBE) has been divided by scaffold methodology (70/15/15) to ensure structural diversity. Implementation: Python 3.10.12, TensorFlow 2.12.0/Keras (CUDA 11.8), trained on NVIDIA RTX 3080 GPU (10 GB VRAM), with Pymatgen 2023.8.10, NumPy 1.24.3, Pandas 2.0.3, Scikit-learn 1.3.0, Matplotlib 3.7.2, and Seaborn 0.12.2.</p>

MGPP-DL: deep learning approach for material graph properties prediction

期刊论文
发表日期：

2026年4月6日

查看全文

Outhman Abbassi,
Hicham Labrim,
Soumia Ziti,
Mohammed Benaissa,
Nassim Kharmoum

摘要

Context

The discovery of semiconductor materials for photovoltaic and optoelectronic applications is limited by the computational cost of DFT calculations and the requirement for complete 3D crystal structures. Existing deep learning approaches, including graph neural networks (CGCNN, MEGNet, ALIGNN) and transformer architectures (Matformer, CrysCo), require structural information, limiting their applicability to hypothetical materials of which only the composition is known. We developed MGPP-DL (Materials Graph Property Prediction via Deep Learning), a structure-agnostic approach that predicts the band gap and formation energy directly from the chemical composition of the material structure. Evaluated on 389,000 materials from the Materials Project, MGPP-DL achieves unprecedented accuracy with mean absolute errors of 0.0178 eV/atom for formation energy and 0.0619 eV for band gap, representing a 67% improvement over state-of-the-art models. The model demonstrates a robust generalization (R $^2$ $^{2}$ = 0.9869 and 0.9182, respectively) and provides an acceleration of 10,000 $\times $ $\times$ par rapport à la DFT, allowing the high-throughput screening of millions of hypothetical compositions.

Methods

Chemical compositions are represented as graphs of completely connected elements with nodes encoding 33 elementary properties extracted via Pymatgen, including electronegativity, atomic radius, and stoichiometric coefficients. Three GNN layers with message passing refine the embeddings of elements (64-dimensional) by multi-scale aggregation. GNN output is processed by two EfficientNet blocks (dimension 128, expansion ratio 4) with squeeze-and-excitation attention mechanisms (reduction ratio 0.25). The training employed the optimizer Adam (learning rate 0.001), MSE loss, dropout (0.3), and batch standardization over 40 epochs (batch size 32). The Materials Project dataset (DFT-GGA, functional PBE) has been divided by scaffold methodology (70/15/15) to ensure structural diversity. Implementation: Python 3.10.12, TensorFlow 2.12.0/Keras (CUDA 11.8), trained on NVIDIA RTX 3080 GPU (10 GB VRAM), with Pymatgen 2023.8.10, NumPy 1.24.3, Pandas 2.0.3, Scikit-learn 1.3.0, Matplotlib 3.7.2, and Seaborn 0.12.2.