Scholarlink

<p>Molecular properties prediction using machine learning is an active research area in materials science and computational chemistry. To predict properties of molecules, such as total energy, one uses molecular features, such as coordinates (positions) and charges, to develop a neural network. One challenge with this approach is that the predictor property should be invariant to rotation, translation, and permutation. In this regard, some types of descriptors, such as the Coulomb matrix, have been introduced. Here we propose a new model using a matrix-based descriptor to predict the molecular total energy. The model uses a multilayer perceptron neural network applied to a subset of the Retrievium dataset, containing 319,380 organic molecules. The results show that the proposed method is capable of accurately predicting the total energy of molecules using measures such as the mean absolute error (MAE) and the root mean squared error (RMSE). In the best case, the proposed method obtains an MAE of 8.91&#xa0;kJ/mol and an RMSE of 14.47&#xa0;kJ/mol. Our experiments also show that the results are statistically significant using a <i>t</i>-test.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural network potentials using matrix-based descriptors

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Omid Tarkhaneh,
  • Sharene D. Bungay,
  • Raymond A. Poirier,
  • Robert C. Mawhinney

摘要

Molecular properties prediction using machine learning is an active research area in materials science and computational chemistry. To predict properties of molecules, such as total energy, one uses molecular features, such as coordinates (positions) and charges, to develop a neural network. One challenge with this approach is that the predictor property should be invariant to rotation, translation, and permutation. In this regard, some types of descriptors, such as the Coulomb matrix, have been introduced. Here we propose a new model using a matrix-based descriptor to predict the molecular total energy. The model uses a multilayer perceptron neural network applied to a subset of the Retrievium dataset, containing 319,380 organic molecules. The results show that the proposed method is capable of accurately predicting the total energy of molecules using measures such as the mean absolute error (MAE) and the root mean squared error (RMSE). In the best case, the proposed method obtains an MAE of 8.91 kJ/mol and an RMSE of 14.47 kJ/mol. Our experiments also show that the results are statistically significant using a t-test.