Scholarlink

<p>Ignition time (<i>t</i><sub>ig</sub>) of thermally thick solids, a crucial indicator of thermal safety and fire risk, was predicted by developing three machine learning (ML) models, namely backpropagation neural network (BPNN) with only one hidden layer (BPNN1), BPNN with two hidden layers (BPNN2), and support vector regression (SVR). A 1D numerical model was utilized to generate dataset. Six parameters highly related to <i>t</i><sub>ig</sub> served as inputs of ML models, including surface absorptivity (<i>ε</i>), density (<i>ρ</i>), specific heat (<i>C</i><sub>p</sub>), thermal conductivity (<i>k</i>), critical temperature (<i>T</i><sub>cri</sub>), and incident heat flux (<i>HF</i>). By tuning hyperparameters, the neuron number of BPNN1 (<i>N</i><sub>h,1</sub>), neuron numbers in hidden layers of BPNN2 (<i>N</i><sub>h,1</sub> and <i>N</i><sub>h,2</sub>), penalty parameter (<i>C</i>) and shape parameter (<i>γ</i>) of kernel function of SVR were optimized. Accuracy, convergence efficiency, and robustness of three models were compared using multiple statistical metrics. Feature importance of inputs was analyzed using two methods. Reliability of the models was verified by comparing with experimentally measured <i>t</i><sub>ig</sub> of five polymers. The results show that the optimal <i>N</i><sub>h,1</sub> of BPNN1 is 34; the optimal <i>N</i><sub>h,1</sub> and <i>N</i><sub>h,2</sub> of BPNN2 are 11 and 10; the optimal <i>C</i> and <i>γ</i> of SVR are 7 and 5. Accuracy, convergence efficiency, and residual of the three models share the same ranking of BPNN2 &gt; BPNN1 &gt; SVR, whereas SVR exhibits highest robustness when treating incomplete datasets. <i>T</i><sub>cri</sub> poses greatest impact on <i>t</i><sub>ig</sub> followed by <i>ε</i> and then <i>HF</i>, while <i>k</i>, <i>ρ</i>, and <i>C</i><sub>p</sub> exert approximately identical least effect on <i>t</i><sub>ig</sub>. Measured <i>t</i><sub>ig</sub> are relatively well predicted by the ML models.</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

Predicting ignition time of thermally thick solid fuels using backpropagation neural network (BPNN) and support vector regression (SVR)

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Haoyu Liao,
  • Xuguang Tang,
  • Junhui Gong,
  • Anran Sun,
  • Cunfeng Zhang,
  • Xuan Wang

摘要

Ignition time (tig) of thermally thick solids, a crucial indicator of thermal safety and fire risk, was predicted by developing three machine learning (ML) models, namely backpropagation neural network (BPNN) with only one hidden layer (BPNN1), BPNN with two hidden layers (BPNN2), and support vector regression (SVR). A 1D numerical model was utilized to generate dataset. Six parameters highly related to tig served as inputs of ML models, including surface absorptivity (ε), density (ρ), specific heat (Cp), thermal conductivity (k), critical temperature (Tcri), and incident heat flux (HF). By tuning hyperparameters, the neuron number of BPNN1 (Nh,1), neuron numbers in hidden layers of BPNN2 (Nh,1 and Nh,2), penalty parameter (C) and shape parameter (γ) of kernel function of SVR were optimized. Accuracy, convergence efficiency, and robustness of three models were compared using multiple statistical metrics. Feature importance of inputs was analyzed using two methods. Reliability of the models was verified by comparing with experimentally measured tig of five polymers. The results show that the optimal Nh,1 of BPNN1 is 34; the optimal Nh,1 and Nh,2 of BPNN2 are 11 and 10; the optimal C and γ of SVR are 7 and 5. Accuracy, convergence efficiency, and residual of the three models share the same ranking of BPNN2 > BPNN1 > SVR, whereas SVR exhibits highest robustness when treating incomplete datasets. Tcri poses greatest impact on tig followed by ε and then HF, while k, ρ, and Cp exert approximately identical least effect on tig. Measured tig are relatively well predicted by the ML models.