面向实用型氢气传感器的贝叶斯优化：钯基等离激元超表面的逆设计
Bayesian Optimization for Practical H2 Sensors: Inverse Design of Pd-based Plasmonic Metasurfaces
AI总结：
本文开发基于贝叶斯优化的逆设计框架，结合第一性原理与电磁模拟，在五维设计空间中优化PdAu纳米盘超表面，实现1至100 mbar范围的高效氢气传感，为多目标传感器设计提供路径。
AI中文摘要：
随着氢气在能源与工业系统中的应用不断增长，氢气检测变得愈发重要。基于钯（Pd）纳米颗粒的光学传感平台因氢气吸收可直接改变其等离激元响应而备受关注；将此类纳米颗粒排列成周期性二维阵列（即超表面），还能通过集体共振进一步增强光学响应。然而，化学成分、纳米颗粒几何结构与阵列结构构成的庞大设计空间，亟需系统方法优化复杂纳米合金超表面的几何结构。本文开发了一种基于贝叶斯优化的逆设计框架，该框架将第一性原理介电函数与电磁模拟相结合，用于识别适用于1至100 mbar范围（氢气在此范围存在易燃隐患）内氢气传感的高性能PdAu纳米盘阵列。我们利用该方法在包含纳米盘高度、半径、阵列周期、聚合物涂层厚度及金（Au）占比的五维设计空间中搜索，以最大化特定波长下氢气诱导的消光变化。结果表明，将第一性原理光学模型与数据高效的优化相结合，可得到适配目标氢气压力、具备实验可行性的纳米颗粒超表面，同时为未来多目标传感器设计提供了路径；研究还揭示了建模方法中仍存在的、限制该方法定量可靠性的缺陷。
英文摘要：
Hydrogen detection is becoming increasingly important as its use grows across energy and industrial systems. Optical sensing platforms based on palladium (Pd) nanoparticles are attractive for this task because hydrogen uptake directly alters their plasmonic response. Organizing such nanoparticles into periodic two-dimensional arrays, known as metasurfaces, further enhances their optical response through collective resonances. However, the large design space presented by chemical composition, nanoparticle geometry, and array structure calls for systematic approaches for optimizing complex nanoalloy metasurface geometries. Here, we develop an inverse-design framework based on Bayesian optimization that couples first-principles dielectric functions with electromagnetic simulations to identify high-performance PdAu nanodisk arrays for hydrogen sensing in the 1 to 100 mbar range where the flammability of H2 becomes a concern. We use our approach to search a five-dimensional design space, comprising nanodisk height and radius, array pitch, polymer coating thickness, and Au fraction in order to maximize the H-induced change in extinction at a single wavelength of choice. The results show that integrating first-principles optical models with data-efficient optimization yields experimentally feasible nanoparticle metasurfaces tailored for targeted hydrogen pressures, while providing a pathway to future multiobjective sensor design. They also reveal remaining gaps in the modeling methodologies that still limit the quantitative reliability of the approach.
