研究目的
To develop and evaluate deep learning methods for predicting molecular excitation spectra using neural networks, aiming to overcome the time and cost limitations of conventional spectroscopic and theoretical methods.
研究成果
Deep neural networks can accurately predict molecular excitation spectra, with the DTNN achieving the best performance (RMSE of 0.19 eV and spectral error of 3%). The method enables fast screening of large molecular databases, facilitating applications in materials science and spectroscopy, with potential for future extensions to inverse predictions and higher-fidelity data.
研究不足
The study uses Kohn-Sham spectra, which are approximate and may not fully represent true excitation energies. The spectral metric is sensitive to peak positions, potentially overlooking broader spectral shapes. The method is currently applied only to organic molecules and may require extension to higher-fidelity datasets like GW calculations or experimental data.
1：Experimental Design and Method Selection:
The study compares three neural network architectures (MLP, CNN, DTNN) for learning molecular excitation spectra from atomic coordinates and charges. Bayesian optimization is used for hyperparameter tuning, and backpropagation with the Adam update scheme is employed for training.
2：Sample Selection and Data Sources:
Datasets used are QM7b and QM9, containing 6k and 132k organic molecules, respectively. Molecules are optimized with PBE+vdW density functional, and excitation energies are computed from the highest 16 occupied eigenvalues, broadened into density of states spectra. A separate 10k diastereomers dataset is used for application testing.
3：List of Experimental Equipment and Materials:
Computational resources from Aalto Science-IT project are used. Software includes the FHI-aims code for quantum mechanical calculations.
4：Experimental Procedures and Operational Workflow:
Neural networks are trained on 90% of the data, with the rest split for validation and testing. Inputs are molecular structures represented via Coulomb matrices or atomistic coordinates. Spectra predictions are made and evaluated using RMSE and R2 metrics.
5：Data Analysis Methods:
Performance is assessed using root mean square error (RMSE) and squared correlation (R2) for excitation energies, and relative spectral error (RSE) for spectra comparisons.
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
