Scholarlink

Abstract <p>In this paper, a thermal explosion of a hydrogen-oxygen mixture in air is simulated using a kinetic mechanism for 11 reactive particles. Two data sets are created: an ODE solver’s generated data set (DS1) and a physically enriched data set (DS2), which complements the uniform sample with additional states near sudden temperature changes, identified using second temperature derivatives. Both datasets are used to train an identical u-net like neural network with five fully connected layers and residual connections to predict the evolution of the thermochemical state at several time stages. On the test set, the physically-enriched dataset reduces the mean squared error from <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(1.8\times 10^{-3}\)</EquationSource> <!--LobJMat2561490Kolesnikova-m1--> </InlineEquation> to <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(1.0\times 10^{-3}\)</EquationSource> <!--LobJMat2561490Kolesnikova-m2--> </InlineEquation> and decreases the standard deviation of prediction errors from <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(1.5\times 10^{-2}\)</EquationSource> <!--LobJMat2561490Kolesnikova-m3--> </InlineEquation> to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(7.3\times 10^{-3}\)</EquationSource> <!--LobJMat2561490Kolesnikova-m4--> </InlineEquation>. This improvement is achieved without increasing dataset size or altering the model architecture. The results confirm that physics-guided data selection enhances accuracy and stability, providing more robust neural network models and bridging the gap between deep learning and traditional physical modeling in chemical kinetics.</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

Influence of Training Data Consideration on Neural Network Approximation of Thermal Explosion

  • 期刊论文

  • 发表日期：

    2026年1月1日

查看全文
  • O. P. Kolesnikova,
  • M. Yu. Malsagov,
  • E. V. Mikhalchenko,
  • I. M. Karandashev

摘要

Abstract

In this paper, a thermal explosion of a hydrogen-oxygen mixture in air is simulated using a kinetic mechanism for 11 reactive particles. Two data sets are created: an ODE solver’s generated data set (DS1) and a physically enriched data set (DS2), which complements the uniform sample with additional states near sudden temperature changes, identified using second temperature derivatives. Both datasets are used to train an identical u-net like neural network with five fully connected layers and residual connections to predict the evolution of the thermochemical state at several time stages. On the test set, the physically-enriched dataset reduces the mean squared error from \(1.8\times 10^{-3}\) to \(1.0\times 10^{-3}\) and decreases the standard deviation of prediction errors from \(1.5\times 10^{-2}\) to \(7.3\times 10^{-3}\) . This improvement is achieved without increasing dataset size or altering the model architecture. The results confirm that physics-guided data selection enhances accuracy and stability, providing more robust neural network models and bridging the gap between deep learning and traditional physical modeling in chemical kinetics.