Physics-informed Machine Learning approach for a selective, sensitive, and rapid sensor for detecting unsafe levels of carcinogenic/toxic VOCs

项目摘要 每年有34万到90万人过早死亡与释放有毒气体造成的空气污染有关。 挥发性有机化合物（VOC），即，估计有18亿吨挥发性有机化合物排放到全球 每年的环境。此外，有些挥发性有机化合物即使在微量水平也会对健康造成严重的不良影响 浓度，例如，癌症，中枢神经和免疫系统的损害。例如，EPA 确定了188种已知或怀疑会导致癌症或其他严重健康影响的有毒空气污染物， 生殖影响、先天性残疾或不利的环境影响。现有的商用传感器 检测VOC，例如光致电离检测器，是非选择性的。因此，它们不适合用于检测 同时存在多种致癌/有毒VOC的不安全水平，例如，苯和甲苯。此外，目前 选择性检测技术，例如气相色谱-质谱（全局色谱 到2030年市场规模> 150亿美元）体积庞大（约5磅），昂贵（约2.5万至10万美元）、缓慢（约2分钟），并且需要 熟练的/受过训练的操作员。因此，Prometheus Technologies正在开发一种专利传感器平台， 具有选择性、低成本、快速、小型化监控解决方案等特点， 熟练/训练有素的操作员检测不安全的致癌/有毒VOC水平。一个重大的技术障碍， 开发选择性VOC传感器是来自背景混杂因素的一小部分的干扰， 特征有限的单波长解吸曲线用于定量。该应用程序的目标是1） 执行一系列经验证的基于物理的模型，以低成本生成相当大的光学传感器数据集 考虑到这一领域数据的稀缺性，这是必不可少的，2）开发一个基于机器学习的模型， 用于检测具有背景混杂因素的目标化合物的不安全水平。这 我们的工作是必要的，以推进我们的专利选择性和小型化VOC光学传感器。

Project Summary Each year, between 340,000 and 900,000 premature deaths can be linked to air pollution caused by releasing Volatile Organic Compounds (VOCs), i.e., an estimated 1.8 billion tons of VOCs are emitted to the global environment each year. Also, some VOCs cause serious adverse health effects even at the trace level concentration, e.g., cancer, damage to the central nervous and immune system. For example, the EPA has identified 188 toxic air pollutants known or suspected to cause cancer or other serious health effects, such as reproductive effects, congenital disabilities, or adverse environmental effects. Existing commercial sensors for detecting VOCs, such as photoionization detectors, are non-selective. So, they are unsuitable for detecting unsafe levels of multiple carcinogenic/toxic VOCs simultaneously, e.g., Benzene and Toluene. Also, the current selective detecting technologies such as gas chromatography-mass spectrometry (global chromatography market size >$15B by 2030) are bulky (~5 lbs.), expensive (~$25K - $100K), sluggish (~ 2 minutes), and requires a skilled/trained operator. Therefore, Prometheus Technologies is developing a patented sensor platform with features such as selectivity, low-cost, fast, small form factor monitoring solution that does not require skilled/trained operators to detect unsafe levels of carcinogenic/toxic VOCs. A significant technological hurdle to developing a selective VOC sensor is interference from a small subset of background confounders when a feature-limited single wavelength desorption curve is used for quantification. The goals of this application are 1) to perform a series of verified physics-based models to generate a sizeable optical sensor dataset at a low cost that is essential considering the scarcity of data in this field, and 2) to develop a machine learning model based on the dataset in step (1) for detecting unsafe levels of target compounds with background confounders. This work is necessary to advance our patented selective and miniaturized VOC optical sensor.