Multi-fidelity machine learning models for structure–property mapping of organic electronics
Multi-fidelity machine learning models for structure–property mapping of organic electronics
复制标题
DOI：
10.1016/j.commatsci.2022.111599
复制
发表时间：
2022-10
影响因子：
3.3
通讯作者：
Chih-Hsuan Yang;B. Pokuri;Xian Yeow Lee;S. Balakrishnan;C. Hegde;S. Sarkar;B. Ganapathysubramanian-B.-Ganapathysubramani
中科院分区：
文献类型：
--
作者：
Chih-Hsuan Yang;B. Pokuri;Xian Yeow Lee;S. Balakrishnan;C. Hegde;S. Sarkar;B. Ganapathysubramanian-B.-Ganapathysubramani
机器学习方法在构建、管理和探索微观结构与属性之间的关系方面取得了重大成功。然而，这些方法的一个主要限制是需要大量的由微观结构-性能对组成的训练数据。获取与特定微观结构相关的属性值通常需要部署详细的物理模拟器，这将成为资源密集型。虽然使用低保真度属性量化器可以抵消创建训练数据集的成本，但在估计属性的准确性/保真度方面存在权衡。在这里，我们利用低保真度和高保真度属性模拟器的可用性，使用深度卷积神经网络构建从微观结构到属性的多保真度映射。从代表有机光伏器件有源层的大型形态学数据集开始，我们从基于快速图形的低保真形态学表征中吸收数据，并从高保真的激动子漂移扩散详细物理模拟器中获取有限数据。我们表明，我们的方法在保持竞争性性能的同时提供了显著的计算节省。这项工作可以很容易地扩展到其他应用，我们设想它作为加速材料量化和发现的基础。
Machine learning approaches have been used with significant success in constructing, curating, and exploring relationships between microstructure and property. However, one major limitation of these approaches is the need for a significant amount of training data consisting of microstructure–property pairs. Getting property values associated with a specific microstructure typically requires deploying a detailed physics simulator which becomes resource-intensive. While using a low(er) fidelity property quantifier can offset the cost of creating the training dataset, there is a trade-off in terms of accuracy/fidelity of the estimated property. Here, we leverage the availability of low- and high- fidelity property simulators to construct a multi-fidelity mapping from microstructure to property using deep convolutional neural networks. Starting with a large dataset of morphologies representing the active layer of organic photovoltaic devices, we assimilate data from a rapid graph-based low-fidelity characterization of the morphology with limited data from a high fidelity excitonic drift-diffusion detailed physics simulator. We show that our method provides significant computational savings while maintaining competitive performance. This work can be easily extended to other applications, and we envision it as a basis for accelerated material quantification and discovery.