研究目的
To address the formability and stability issues in perovskites by developing a machine learning model based on high-throughput DFT calculations to predict thermodynamic stability more accurately than empirical descriptors.
研究成果
The ML model based on DFT-calculated data provides a more accurate and efficient method for predicting perovskite stability compared to empirical descriptors. It successfully guides experimental engineering of stable perovskites, with validation showing high agreement with experimental results. The approach enables exploration of a wide stability landscape for mixed perovskites, offering insights for materials design.
研究不足
The decomposition energy calculations consider only one decomposition pathway, potentially overestimating stability for some perovskites (e.g., those with Cu+ or In+). The ML model may not generalize well to perovskites with non-alkali metal elements on the A site due to limited training data. Overfitting occurred when electronegativity features were included.
1：Experimental Design and Method Selection:
High-throughput DFT calculations were used to compute decomposition energies (ΔHD) for 354 halide perovskites. A machine learning model (Kernel Ridge Regression with radial basis function kernel) was trained on this dataset to map the relationship between ionic radii features (RA, RB(I), RB(III), RX) and ΔHD. The model was validated against experimental data and compared with empirical descriptors like tolerance factor t.
2：Sample Selection and Data Sources:
The dataset included 354 halide perovskites (single and double) with compositions from the periodic table. Additional validation used 246 A2B(I)B(III)X6 compounds not in the training set and experimental data from literature.
3：List of Experimental Equipment and Materials:
Computational tools: VASP software for DFT calculations, scikit-learn package for ML. Materials: Various halide perovskite compositions.
4：Experimental Procedures and Operational Workflow:
DFT calculations were performed using VASP with PBEsol functional, cutoff energy of 300 eV, k-mesh density set by KSPACING=0.12, geometry relaxation until forces <0.01 eV ??1, and energy convergence criterion of 0.5 meV. ML model training involved fivefold leave-one-out cross-validation to optimize hyperparameters.
5：12, geometry relaxation until forces <01 eV ??1, and energy convergence criterion of 5 meV. ML model training involved fivefold leave-one-out cross-validation to optimize hyperparameters.
Data Analysis Methods:
5. Data Analysis Methods: Root-mean-square error (RMSE), precision, recall, and F1 score were used to evaluate model performance. Statistical comparisons were made with empirical descriptors.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
