(Invited) Chemical Sensing in the Big Data Era: How and Where Does the Chemical World Store Its Information?

Introduction In recent years, data analytics have emerged as an important tool for understanding many business and societal trends, and this has been made possible by new algorithms for collecting and processing large data sets. An attendant improvement in sensing technology has also allowed for data analytics to now be applied to physical phenomena. This has been most apparent, for example, in video analytics where CMOS imagers now cost as little as US$3, leading to the commoditization of digital imagery and the corresponding widespread use of image and video analytics across many industries. More recently, big data methods that collect and exploit chemical data have been demonstrated, where web-connected wearable sensors that track a user’s local chemical exposure1,2 alsoupload and aggregate the data to provide real-time air-quality maps.1 Another recent example of chemical big data has shown that chemical monitoring of wastewater systems can provide data useful for understanding drug-abuse patterns3,4 and can inform public policy to create more resilient cities.4 Given the complexity and dynamic nature of the chemical world, comprised of chemicals associated with public health and safety, law enforcement, food quality, agricultural, environmental, and climate issues, it is clear that the opportunity exists for new and creative applications of wide-scale chemical sensing. If these can be combined with big-data analytics, it will provide new and better ways of monitoring our world for the common good. The hope is that the emergence of new low-cost chemical sensors will eventually do for chemical sensing what the US$3 CMOS imager did for video analytics. However, these advances in sensor technology must be accompanied by the identification of appropriately useful new data sources. To facilitate this objective, this presentation attempts to answer the question “how and where does the chemical world store its information?” The hope is that the answer can serve to guide new strategies to improve our means to collect, aggregate, and exploit chemical big data. Chemicals as Sources of Information Chemicals encode information multiple ways, with the two most common being composition and concentration, where composition associates a chemical with a specific phenomenon of interest and concentration indicates amplitude. It is for this reason that most chemical sensors focus primarily on selectivity (composition) and sensitivity (concentration). Recent trends have shown that simultaneous improvements in both can allow for the transition from targeted to non-targeted analyses, thereby providing new opportunities for information discovery.5 This capability has proven useful for biomarker discovery, environmental toxicology, and, more recently, food quality assurance6 and environmental monitoring.7 Applying these trends to ubiquitous chemical sensing8,9 will open up many new roles for chemical big data, but implementing non-targeted analyses requires chemical sensors with very high informing power,10 and thus development efforts should specifically focus on means to increase both the resolving power11 and the dynamic range of new sensors in order to realize these capabilities. One immediate benefit of doing so would be to better exploit information contained in lower-concentration volatile organic compounds (VOCs). For example, recent measurements of urban air identified only seven VOCs whose peak concentrations were >100 parts per billion, but 112 VOCs whose peak concentrations were >100 parts per trillion.12 Clearly, the ability to detect and discriminate chemicals at lower concentrations provides opportunities to collect more information. However, recent work using ultra-trace (sub part-per-quadrillion) vapor detection has shown that unidentified compounds interfere even when using instruments with very high informing power such as MS/MS.13 This reinforces the importance of instrument channel capacity or, alternatively, the need for highly selective low-channel capacity systems when seeking information contained in ultra-low-concentration chemicals. In addition to composition and concentration, chemicals also encode information via spatial and temporal patterns. This is most apparent in climate studies, where spatiotemporal trends in both O3 and CO2 have played central roles in understanding global atmospheric processes and have served to influence regulatory policy. On a more local scale, these data features also play a central role in efforts to create urban sensing networks that extract specific information from amidst the normal patterns of life. This includes networks that provide early warning against chemical threats,14 illicit activities,15,16,17 and high-granularity maps of urban air quality.18 In addition to these millennial, annual, seasonal, diurnal, and hourly variations in chemical signals, higher frequency (<seconds) concentration fluctuations have been associated with turbulent plume phenomena and their measurement has been shown useful in the performance of source-seeking autonomous robots19 that mimic the way that the insects tracks plumes.20,21 One final important data feature is the correlation of chemical compounds across the concentration, spatial, and/or temporal domains. Compounds whose concentrations vary in concert with one another can be associated as originating from the same source, which provides useful information in untargeted analyses.22 This presentation will review these concepts as a means to reframe chemical sensor requirements in their context as “information collectors”. This information-based approach should facilitate the matching of new sensing technology with new big data opportunities applied to our chemical world. Such high fidelity data sets, as they become available, will also provide new opportunities to provide machine learning approaches that provide even greater insight into the chemical phenomena important to human quality of life. Approved for public release. Distribution is unlimited. This material is based on work supported by the U.S. Air Force under Contract No. FA8702-15-D-0001. Any opinions, conclusions, or recommendations expressed in this material are those of the author and do not reflect the views of the U.S. Air Force. Figure 1