[论文解读] Emissivity Prediction of Functionalized Surfaces Using Artificial Intelligence
本文提出了一种数据驱动的人工智能(AI)框架,用于预测飞秒激光处理的6061铝合金表面的半球发射率。通过结合三维激光共聚焦显微镜(LSCM)图像、激光加工参数和表面特性,作者训练了卷积神经网络(CNN)及机器学习模型(如决策树、人工神经网络)来预测发射率,误差低于4%,展示了该方法在热辐射工程中作为物理建模高度精确且可扩展的替代方案。
The radiative response of any object is governed by a surface parameter known as emissivity. Tuning the emissivity of surfaces has been of great interest in many applications involving thermal radiation such as thermophotovoltaics, thermal management systems, and passive radiative cooling. Although several surface engineering techniques (e.g., surface functionalization) have been pursued to alter the emissivity, there exists a knowledge gap in precisely predicting the emissivity of a surface prior to the modification/fabrication process. Predicting emissivity by a physics-based modeling approach is challenging due to surface's contributing factors, complex interactions and interdependence, and measuring the emissivity requires a tedious procedure for every sample. Thus, a new approach is much-needed to systematically predict the emissivity and expand the applications of thermal radiation. In this work, we demonstrate the immense advantage of employing artificial intelligence (AI) techniques to predict the emissivity of complex surfaces. For this aim, we fabricated 116 bulk aluminum 6061 samples with various surface characteristics using femtosecond laser surface processing (FLSP). A comprehensive dataset was established by collecting surface characteristic data, laser operating parameters, and measured emissivities for all samples. We demonstrated the application of AI in two distinct scenarios. First, the range of emissivity of an unknown sample was shown to be estimated correctly solely based on its 3D surface morphology image. Second, the emissivity of a sample was precisely predicted based on its surface characteristics data and fabrication parameters. The implementation of the AI techniques resulted in the highly accurate prediction of emissivity by showing excellent agreement with the measurements.
研究动机与目标
- 为解决在制造前预测功能化表面发射率方面的关键知识空白。
- 克服基于物理的模型所面临的局限,这些模型受限于复杂的表面相互作用和高计算成本。
- 开发一种数据驱动的AI框架,实现在热辐射应用中对发射率进行精确、快速且系统化的预测。
- 证明利用AI从表面形貌图像和制造参数预测发射率的可行性。
提出的方法
- 使用飞秒激光表面处理(FLSP)以不同的激光能量密度和脉冲数制备了116个6061铝合金样品。
- 收集了250张三维LSCM图像以捕捉表面形貌,并提取了表面特性。
- 使用热成像相机测量方向发射率(7.5–14 µm),并通过数值积分获得总半球发射率。
- 在LSCM图像上训练卷积神经网络(CNN)作为特征提取器,将样品分类为七个发射率类别。
- 将LSCM提取的特征与激光参数及表面特性结合,形成综合数据集用于机器学习。
- 使用80%训练集、10%验证集和10%测试集评估多种模型——k近邻(kNN)、人工神经网络(ANN)、广义线性模型(GLM)、W-M5P和决策树(DT)。
实验结果
研究问题
- RQ1AI能否仅从3D表面形貌图像准确估算功能化表面的发射率范围?
- RQ2结合深度学习与机器学习的混合AI模型,能否利用表面特性与制造参数精确预测FLSP处理的6061铝合金表面的发射率?
- RQ3不同机器学习模型在从多模态输入数据(图像、参数、表面特征)预测发射率方面表现如何比较?
- RQ4基于图像的特征提取误差对最终发射率预测精度有何影响?
主要发现
- 基于CNN的图像分类模型成功以高精度将样品分入发射率类别,证明了仅从形貌信息实现AI驱动发射率范围估算的可行性。
- 决策树(DT)和人工神经网络(ANN)模型表现最佳,测试集上的平均绝对误差分别为3.31%和3.88%。
- DT和ANN模型的R²值分别为0.979和0.980,表明预测值与实测发射率高度一致。
- W-M5P和kNN模型也表现出色,R²值均高于0.96,均方根误差(RMSE)低于0.05。
- 有两个样品的预测误差显著,可能源于模型1中CNN的特征提取误差在模型2中的传播。
- 本研究建立了一种可扩展的数据驱动替代方案,用于预测复杂功能化表面的发射率,替代基于物理的建模。
更好的研究,从现在开始
从论文设计到论文写作,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。