1. Project summary.

The global pandemic of COVID-19 has clearly demonstrated the need to re-envision our approaches to build safe indoor infrastructure designed to protect human health. Most notable are multi-user indoor facilities, e.g., hospitals, educational institutions, businesses, retail, entertainment, and sports facilities. However, while the green building movement has focused on reducing resources, it has not addressed the control of chemical and biological contaminants in the indoor environment. This proposal aims to develop a real-time virtual environment for multi-user facilities capable of controlling indoor air quality to reduce the exposure from chemical and biological contaminants. This virtual environment, the so-called digital twin, will be developed based on knowledge of pathogen persistence under different environmental conditions (Task 1), real-time measurement of air quality in selected multi-user facilities (Task 2), computational fluid dynamics simulation of the air pollutant transport by physics-informed deep learning (Task 3), privacy-preserving analysis of air quality data and user profiling by federated learning (Task 4), and designing an artificial intelligent management algorithm for controlling air quality (Task 5). Three multi-user facilities with a potentially high risk of pathogen transmission, Vinmec hospital in Times City, Technopark in Ocean Park, and VinUni, are selected for this study. The project’s outcome will be an accurate assessment of the health risks from indoor contaminants and real-time recommendations for intervention strategies to reduce these risks.

2. Background and Significance.

The quality of the indoor environment has a critical impact on people’s health because on average, we spend more than 90% of our time indoors [1]. Indoor air pollution, which ranks 9th among 69 risk factors included in the global burden of disease report, is estimated to cause > 3 million premature deaths worldwide [2]. In addition to induce chronic diseases, several studies have also shown the association between indoor particulate concentrations and SARS-CoV-2 cases or mortality rates around the world [3,4,5]. The COVID-19 pandemic has accelerated the need for improved indoor environments to reduce the transmission of the virus [5]. Hence, providing a healthy and safe indoor environment can save lives, reduce diseases, and increase our quality of life. The economic growth and rapid urbanization in Vietnam in the last two decades have significantly improved housing conditions in terms of reduced crowdedness, clean water supply, and adequate sanitation facilities. However, there was much less attention to the other aspects of housing quality, i.e., air quality and thermal comfort. Previous studies have shown that the levels of indoor air pollutants such as black carbon and particulate matter in different buildings in urban areas of Vietnam were similar to or sometimes even higher than the level of those pollutants outdoors [6,7,8], exceeding the WHO guidelines and thus posing elevated risk for the occupants. Better management of indoor environmental quality and saving energy consumption at the same time is of critical national and international significance. Such impacts are expected to increase in the future, as they will be exacerbated by climate change, and the aging population will make more people susceptible. The outcomes of this project are expected to deliver the necessary tool to achieve the target of making healthier buildings worldwide to protect human health.

3. Approach. 

Indoor air quality is highly dynamic depending on both indoor and outdoor emission sources and building envelope. For example, a large number of occupants likely leads to high concentrations of CO2, particulate matter (PM), and pathogens (if infected individuals are present). Modern multi-user facilities usually have a centralized heating ventilation and air conditioning (HVAC) system. The thermostat of the HVAC system is often set at a temperature decided by facility managers to enhance energy efficiency, rather than on reducing the concentration of indoor air contaminants. We propose to develop a virtual environment linked to an actual physical environment through sensors of common air pollutants. This virtual environment will be coupled with computational models developed based on the knowledge of pathogen persistence in the indoor environment, physics-informed machine learning of air pollutant transport, privacy-preserving machine learning of monitoring air quality data, and AI-empowered air quality control algorithm. This goal will be achieved through five tasks as described below:

Task 1. Quantify environmental persistence and infectivity of respiratory viruses [Nguyen, Verma, Phuong and Taylor-Robinson]

The persistence and infectivity of coronavirus and influenza viruses will be measured under various indoor environment (e.g., temperature (T), relative humidity (RH), and lighting conditions). We will study two coronavirus and influenza virus strains – CoV-NL63, and Porcine respiratory coronavirus (PRCV). CoV-NL63, PRCV and SARS-CoV-2 are all coronaviruses. Thus, findings with CoV-NL63 and PRCV will likely be translatable to SARS-CoV-2. Human coronavirus CoV-NL63 causes a seasonal cold-like disease [9]. CoV-NL63 also shares the same receptor as SARS-CoV-2, the virus causing the current COVID-19 pandemic [10]. PRCV is a common collected using a biosampler connected to a vacuum pump, operated at a flow rate of 12.5 L/min, and make-up air purified via a HEPA filter will be fed to the box using a compressed air system. The air, after passing through the biosampler will be passed to a disinfection sink containing bleach. The suspensions collected in the biosamplers will be transferred to 50 mL tubes and subjected to infectivity assays and genome extractions before PCR. The level of viral inactivation will be calculated as the ratios of infectivity before and after exposure. Molecular assays will be performed to determine the inactivation mechanisms as have been done in Nguyen’s previous studies [11,12]. Knowledge of the aerosol size (measured by a particle counter) is important because larger size aerosols will deposit and not be captured by the biosampler. The size distribution of the viruses will be used to determine the deposition rate in assessing the risk of infection.

Task 2. Quantify the concentration of air pollutants in indoor environments with and without intervention [Verma, Nguyen, Le and Taylor-Robinson]

We will deploy air samplers at different locations in hospital facilities (e.g., Vinmec Times City in Hanoi) and other buildings (e.g., VinUni and TechnoPark building in Hanoi) to collect bioaerosols and reactive volatile and particulate matter (PM) components. Sampling will be conducted in both summer and winter seasons to account for the effect of environmental conditions on indoor contaminants. Some potential locations include the front office, emergency rooms, pediatrics center, and classrooms since these locations are expected to experience relatively high concentrations of pathogens. In addition, other locations storing the cleaning supplies will be sampled to assess the exposure concentrations from reactive chemicals. Two different sets of samplers will be deployed to collect bioaerosols and the reactive PM components. We will use sorbent tubes (Anasorb and/or charcoal) and an active sampling method for the volatile compounds. All the samplers for collecting bioaerosols, volatile compounds, and reactive PM species will be operated using a battery-operated vacuum pump. Sampling for bioaerosols and reactive PM species will be conducted at a flow rate of 15 LPM, while 0.5 LPM will be used to collect the volatile compounds using sorbent tubes. These pumps are very quiet and will not disrupt the daily activities of the inhabitants. Integrated sampling will be conducted continuously for one week to collect enough substrate for the analysis. To the best of our ability, other metadata will also be collected to characterize each space (e.g., HVAC air handler flow rates and outdoor air supply fractions). Environmental conditions (e.g., temperature and relative humidity) will be measured with Onset HOBO U12 data loggers throughout sampling. The bioaerosol samples collected onto gelatin filters will be analyzed to determine the concentration of bacteria and viral pathogens. DNA and RNA from the air samples will be extracted using MagMax Pathogen DNA/RNA kit. For SARS-CoV2 and its variants, we will use the method developed recently by Nguyen’s lab [13]. We will also use the recently released CDC assays to detect SARS-CoV2 and seasonal influenza viruses simultaneously. For bacterial pathogens, we will target the mip gene of Legionella pneumophila, the gyrB gene of Pseudomonas aeruginosa, and the groEL2 gene of Mycobacterium tuberculosis. The primer sets for these common bacterial respiratory pathogens have been adopted from previous studies by Nguyen’s lab [14].

We will measure air exchange rate (AER) in the buildings using the CO2 tracer method. AER is defined as the number of times air gets replaced in each room every hour. Commercially available CO2 sensors are small, quiet, easy to use, relatively inexpensive (<$200), and measure real-time CO2 concentrations. With an intensive measurement of CO2 concentrations indoor and outdoor, AER can be calculated using eq. 

(1):

Here,    (m3) is space volume,  (mg/m3),(mg/m3) and (mg/m3) are CO2 concentrations in the inflow, indoor air and outflow, respectively. = = (m3/hr) are air flows into/out of the building/space, (mg/hr) is the indoor emission source of CO2, and (m3/hr) is the first-order degradation constant. Since = 0 for CO2(conservative pollutant), AER can be determined by a simple decay method. When there is no source, i.e., = 0, AER, denoted by , can be resolved from eq.

(2):

Based on the measured CO2 data, we will compute the AER from this decay model using non-linear regression analysis to minimize squared residuals for selected periods when a smooth decay curve in CO2 levels is observed. We will also collect size-segregated aerosols using Canaree sensors and other environmental parameters (T, RH) in different-sized indoor spaces. The relationship between AER and CO2 measurements will be validated for a couple of traditional indoor settings, as chosen by our partners (i.e., Vinmec Times City, VinUni, and TechnoPark). We will target to establish a correlation between the concentrations of aerosols and CO2.

Several intervention strategies will be considered to reduce the exposure of inhabitants to indoor contaminants. Specifically, we will consider the effect of commercially available air purifiers, enhancing ventilation, higher efficiency filters in the HVAC system, and installing UV lights. We hypothesize that installing air purifiers and using higher efficiency filters in the HVAC system will have a direct impact on reducing the concentration of particulate contaminants (including pathogens), while enhancing ventilation will dilute the concentration of both volatile and particulate compounds. Installing UV lights at selected places of potential sources (e.g., patient rooms) should inactivate the pathogens and thus have a direct impact on reducing the microbial concentration in the indoor environment. However, to quantify the effect of these interventions, we will repeat environmental sampling for reactive volatile and particulate compounds and pathogens for at least one month at all locations with every intervention implemented.

Task 3. Generate full-scale full-geometry predictive tools to understand airflow and aerosol transport in indoor environments [Yan, Verma, Nguyen, and Le]

The airflow motion and heat transfer in indoor environments are heterogeneous and turbulent, significantly affecting the transport of virus-containing aerosols. We will create a computational fluid dynamics (CFD) model that integrates the Navier-Stokes equations, heat transfer, and a scalar convection-diffusion equation to describe the airflow, temperature, and their efforts on aerosol transport. The evolution of aerosol concentration is determined by two competing physics: convection and diffusion. The convective velocity is obtained by superposing the formulation (VMS) to solve the aerosol-laden flows in a fully coupled way by fully resolving the geometry of indoor environments and HVAC systems. As a large eddy simulation (LES) model, the FEM-based VMS approach has been applied to various large Reynolds number turbulent flows, including particle-laden flows (see Fig. 2 for a preliminary study of aerosol transport in a lab at UIUC). It should be emphasized that the methodology proposed here is based on first conservation principles, with a minimal number of model parameters. These parameters will be calibrated by solving an inverse problem via a recently developed physics-informed neural network for CFD. We will leverage the measured data of aerosol concentration, CO2 concentration, and other available fluid-related measurements for the machine learning (ML) tasks.

Fig. 2. An example of full-scale full-geometry CFD model for aerosol transport in indoor in environments showing the distribution of contaminated aerosol around the infected students and a purifier can prevent the spread of the viruses emitted from the infected students

Task 4. Privacy-preserving analysis of air quality data and user profiling by federated learning [Wong, Le, Yan, and Verma]

The indoor air quality data will be collected from various sensors and internet-of-things (IoT) devices. We will leverage Machine learning/Deep learning by Watson Studio to manage the HVAC system based on its operating parameters and the decision of building managers. We will also develop a meta-learning model to learn the users’ usage of their HVAC. Our approach’s novelty lies in the idea that we will maintain a set of multiple pre-trained meta-models and efficiently switch between them in a real-time manner to accommodate sudden profile changes (e.g., multiple occupants or mood changes).

The success of ML systems depends on the availability of high-quality data collected from various sources such as sensors and internet-of-things (IoT) devices. However, a single entity might only own some of the data it needs to train the ML model it wants; instead, valuable data examples or features might be scattered in different organizations or entities. For instance, IoT sensing data sit in data silos, and privacy concerns limit sharing such data for ML tasks. Consequently, large amounts and diverse data from different stations/facilities are only partially exploited by ML. The abundance and availability of personal information online and from IoT sensors have led to an increasing rate of privacy breaches and potential security harm to individuals. Nevertheless, access to this information is critical for the success of many beneficial applications, such as population health studies or environmental planning.

Our indoor air quality control system deals with sensitive information about people’s living patterns and usage habits in the building. This information can be used to create sophisticated profiles of individuals and facilities. Some of the data involve personally identifiable information (PII), such as geolocation, and unique device information that, when combined, can be used to identify an individual or private facility. Due to privacy concerns, the concept of federated learning (FL) was first introduced by McMahan et al. [15]. The main idea is to train ML models in a decentralized fashion where no private dataset is sent to a central repository [16]. Specifically, the data for the learning tasks are acquired and processed locally at the edge node, and only the updated ML parameters are transmitted to the server for aggregation [17]. FL aims to train a single ML model using all the data available cooperatively without moving the training data across organizational or personal boundaries [18].

The available data sets will be divided into training and testing subsets for cross-validation. The ML model will be developed based on the PI’s previous code in TensorFlow. Once the parameters are identified and validated, the forward predictive models can calculate airflow, temperature, and aerosol transport. Aiming to build an easy-to-use cyberinfrastructure for the community, we will develop predictive models in the finite element automation framework, FeniCS [19, 20]. The code can be deployed to any indoor environment with full geometry details (e.g., furniture). The model can accurately predict the velocity, pressure, temperature, and time history of the spatial distribution of pollutant concentration everywhere in indoor environments to provide deep insights into airflow, heat transfer, and contaminant transport in indoor environments and generate practical guidelines to mitigate the risk of viral infection.

Task 5. Design an AI management platform for controlling air quality [Le, Wong, Yan, Verma]

Our final task is to develop an AI-integrated privacy-preserving analytics platform that supports indoor air quality monitoring, detecting, and preventing anomalies. Specifically, the platform will offer interactive interfaces for infection control and facility managers to make informed and optimal intervention strategies as per different intended uses of the multi-used indoor environments. We envision a system that will allow one to (1) have a visual understanding of the airflow motion and heat transfer and the multi-human behavior of indoor spaces, (2) simulate different intervention strategies based on customizable preferences and purposes, and (3) generate reasonable recommendations for the intervention strategies that optimize for different objectives such as reducing exposure to contaminants, reducing transmission risks, reducing energy consumption, and maximizing indoor thermal comfort.

This goal will be achieved by a data-driven algorithm using advanced Artificial Intelligence (AI) to manage and control HVAC and other devices, which are tested in Task 1, aiming to satisfy conflicting operational goals, the reduction of indoor contaminant’s exposure and infection risks, the comfort level of users and the efficient energy management. This AI optimization problem is challenging due to: (1) the wide range of appliance models being used in reality, (2) the diversity in building architecture that would impact airflow and correlation among appliances distributed in multiple rooms, and (3) the various activities of the occupants. These challenges will be overcome by the findings from the previous tasks. Specifically, the algorithms will be trained using the big data collected from the sensing platform and airflow (Tasks 2 and 3) and deployed to automate HVAC management in a number of selected residential houses for fine-tuning and validating its performance. We will use reinforcement learning (RL) to solve this complex optimization as follows:

(i) Quantifying the objectives of the task: minimizing total energy consumption, minimizing discomfort, and reducing the contaminant/pathogen concentration.

(ii) Formulating the energy controlling problem by reinforcement learning-based planning methods on the following variables: the house architecture, the distribution of appliances, the appliance energy consumption data, occupant activities, and habits. We will employ a state-of-the-art algorithm called Poly-HOOT [21] to learn flexible consumption models for appliances.

(iii) Iteratively learning and training the RL model based on the observed air quality data and updating the control plan of HVAC accordingly. The key challenge here is that if this iteration is done in a real-time manner, there is a high chance that it would pick up a strongly suboptimal behavior (due to the stochastic nature of the approach). If this happens too often, it will cause much annoyance and distrust to the user. To avoid this situation, we aim to use the offline RL approach [22] that uses previously observed data and learns the underlying model offline. In addition, the Poly-Hoot algorithm is also run offline to identify the most suitable plan.

4. Recruitment of Postgraduate Students.

If awarded, we will recruit five excellent Ph.D. students to be part of this project. Each will focus on a specific task in the proposal. Specifically, the plan is outlined as follows:

• In UIUC, we will conduct empirical studies in the laboratory to understand the environmental persistence of viruses. [one student – Task 1]
• In UIUC and VinUni, we will develop methods for field monitoring in indoor environments to measure the concentration of physical, chemical, and biological contaminants. Also, we will work on developing a mathematical model to assess the distribution of contaminants in the indoor environment. [two students – Tasks 2 and 3]
• In UIUC and VinUni, we will assess the effect of intervention strategies on the concentration of indoor contaminants. [one student – Task 4]
• In VinUni, we will design the front-end platform to provide optimal intervention strategies for specific use-based indoor environments. [one student – Task 5]

All students will be supervised/co-supervised by members of the research team, who will also serve as the student’s thesis committee. Monthly virtual meetings with the whole team will be conducted to share results, research progress, and feedback. The students enrolled at either university will exchange between institutions to conduct laboratory and field monitoring and learn computational fluid dynamics modeling and computing skills required to accomplish the proposed tasks. We anticipate having at least ten peer-reviewed high-impact publications from the proposed work, first-authored by the hired Ph.D. students.

The $20K fund will be mainly used at the home institution to buy necessary chemicals (reagents) and supplies to conduct laboratory experiments, and build sensors for field monitoring. When the student visit to host institution, the $20K fund will be used for conference expenses (travel and registration), environmental data collection expenses, and write grants to obtain external funding.

5. Expected Outcomes.

The proposed study is built on the complementary expertise from VinUni and Illinois faculty to conduct laboratory and field measurements in the indoor environment (UIUC), mathematical tools to model indoor contaminants transport (UIUC and VinUni), and sensors technologies and advanced computing skills (AI/ML; VinUni and UIUC). We will train the next generation of scientists and engineers who can work in trans-disciplinary teams, communicate effectively to a wide range of audiences, and distill research into practical solutions. The methods and products developed through these efforts will be applicable to the design of greener and healthier buildings worldwide, and therefore the collaboration are bound to be sustainable with strong potential to secure future external funding.