基于神经网络与 MC X-ray 蒙特卡罗模拟的 20 kV 均质多元素块体样品定量分析方法A neural network-based Monte Carlo X-ray (MC X-ray) approach for quantitative analysis of homogeneous multi-element bulk samples at 20 kV
麦吉尔大学采矿与材料工程系与 Dragonfly（Comet Group，蒙特利尔）合作，用蒙特卡罗程序 MC X-ray 在 20 kV 条件下模拟约 3000 组 EDS 谱图及对应元素质量分数，训练多层感知机（MLP）直接从整条谱图预测成分，实现无需物理标样的 SEM-EDS 定量分析。在 11 种经认证的标准矿物上验证，RMSE 为 0.0056–0.0179，R² 为 0.9873–0.9981，质量分数不低于 0.05 的元素平均绝对百分比误差 4.41%，在消费级笔记本上约 15 分钟完成；质量分数低于 0.05 的微量元素仍不可靠。论文指出 MC X-ray 已集成进图像分析平台 Dragonfly，并可借 Dragonfly 的 Python 扩展用一个脚本批量生成训练数据，后续将利用 Dragonfly MC X-ray 模块的几何配置扩展到薄膜、多层等复杂样品。McGill University's Department of Mining and Materials Engineering and Dragonfly (Comet Group, Montréal) used the MC X-ray Monte Carlo program to simulate about 3,000 EDS spectra with matching elemental weight fractions at 20 kV and trained a multilayer perceptron to predict composition from the full spectrum, enabling standards-free SEM-EDS quantification. Validated on 11 certified mineral standards, it gave RMSE of 0.0056–0.0179 and R² of 0.9873–0.9981, with a 4.41% mean absolute percentage error for elements at weight fractions of at least 0.05, in about 15 minutes on a consumer laptop; trace elements below 0.05 remain unreliable. The paper notes that MC X-ray has been integrated into the Dragonfly image analysis platform, whose Python extension generates large training datasets from a single script, and plans to use the Dragonfly MC X-ray module's geometry options to extend the method to thin films and multilayer samples.