key: cord-1009934-pdw1m268 authors: Mesgarpour, Mehrdad; Abad, Javad Mohebbi Najm; Alizadeh, Rasool; Wongwises, Somchai; Doranehgard, Mohammad Hossein; Jowkar, Saeed; Karimi, Nader title: Predicting the effects of environmental parameters on the spatio-temporal distribution of the droplets carrying coronavirus in public transport- A machine learning approach date: 2021-10-07 journal: Chem Eng J DOI: 10.1016/j.cej.2021.132761 sha: 22212a8b4495c45d4eeb60c819a0863ac18b2413 doc_id: 1009934 cord_uid: pdw1m268 Human-generated droplets constitute the main route for the transmission of coronavirus. However, the details of such transmission in enclosed environments are yet to be understood. This is because geometrical and environmental parameters can immensely complicate the problem and turn the conventional analyses inefficient. As a remedy, this work develops a predictive tool based on computational fluid dynamics and machine learning to examine the distribution of sneezing droplets in realistic configurations. The time-dependent effects of environmental parameters, including temperature, humidity and ventilation rate, upon the droplets with diameters between 1 and 250 [Formula: see text] are investigated inside a bus. It is shown that humidity can profoundly affect the droplets distribution, such that 10% increase in relative humidity results in 30% increase in the droplets density at the farthest point from a sneezing passenger. Further, ventilation process is found to feature dual effects on the droplets distribution. Simple increases in the ventilation rate may accelerate the droplets transmission. However, carefully tailored injection of fresh air enhances deposition of droplets on the surfaces and thus reduces their concentration in the bus. Finally, the analysis identifies an optimal range of temperature, humidity and ventilation rate to maintain human comfort while minimising the transmission of droplets. Evaporation of droplet [ . - Density for kth fluid [ . -3 ] Heat flux for the kth fluid (droplet enthalpy) [ . -2 ] The density of the liquid phase [ . - Since late 2019, the transmission of SARS-CoV-2 [1] has significantly affected humankind in the entire world. It has been shown that coughing, sneezing [2] , and speaking [3] constitute the main routes of virus transmission. This could be through the deposited droplets over surfaces or the suspended droplets entering the mouth, nose and eyes. Virus-carrying droplets cover a range of sizes, and the larger they are, the more likely they become to deposit on surfaces, while the smaller ones may remain suspended in the air for a long time. It has been argued that these airborne droplets are responsible for virus transmission [2] . It follows that predicting the distribution of such droplets is an essential requirement for understanding the transmission routes of the virus. The suggested 1-2 m social distance by WHO is based on the study of Wells [4] in 1934 assessing the lifetime of droplets. However, the droplets can be carried to further distances if the airflow is considered. Sneezing and coughing are accompanied by complex turbulent respiratory flows, which can be responsible for virus spreading. The interactions between the moving droplets and the airflow break up the droplet and alter the subsequent evaporation and deposition processes. It has been reported that a single sneeze can produce droplets with velocities in access of 10-20 m/s [5] . The larger droplets settle on the surface by the gravitational forces. The smaller droplets, however, are transformed into nuclei droplets through evaporation. These airborne droplets can be carried everywhere in an enclosed environment by the natural/forced flow of air. The study of droplet-flow interactions can help understand the complex physics of virus transmission, particularly in enclosed environments such as public transport where the risk of infection is high. A sneeze is an involuntary reflex of the respiratory system that expels explosive air into the ambient. The transient high-pressure air sprays the mucus into the environment [6] .Experiments showed that the droplets from a single sneeze can be carried by a turbulent gaseous cloud even 7-8 meters far from the mouth [7] . This clearly shows the importance of analysing sneezing process in enclosed environments. Numerical analyses have been already successfully employed to simulate the travelling paths of virus-containing droplets; see for example [9] . Such investigations help to examine the efficacy of ventilation systems, design social distance layouts in indoor environments, and provide more effective health care protocols. For instance, by determining a wide range of initial injection velocity, size distribution , it was suggested that the maximum contamination distance can vary between 1 and 4 meters [9, 10] . Moreover, using the Monte-Carlo technique, it was shown that smaller droplets 8 could stay suspended in air for a long time, and that the initial velocity has a minor effect on the maximum suspension time. It was further demonstrated that a droplet with a 2.5 size remains suspended for approximately 41 minutes, whereas a droplet with a 100 size falls down after 1.5 seconds [8] . Also, increasing the relative humidity prolongs the evaporation time of the droplets [9] . Most recently, a comprehensive simulation of a single sneezing passenger inside a bus was conducted by Mesgarpour and co-workers [10] . The droplets size distribution was calculated by applying the volume of fluid method to the mouth and lips of the infected person. It was shown that the droplets with sizes less than 250 could travel the whole length of the bus. Further, evaporation of the droplets are of significance as it directly affects survival of the droplets and thus their virus carrying potentials [11] . It has been shown that it takes between 2 and 137 seconds for a droplet with a volume of 1 to 10 mL to completely evaporate over a surface [12] . In recent months, significant attempts have been made to numerically simulate the spread of droplets in public indoor environments, see, for example, Lu et al. [2] , Ren et al. [13] and Khosronejad et al. [14] . These studies indicated that ventilation systems could accelerate the dispersion of droplets in the room, yet they fail to remove the larger droplets before deposition. Abuhegazy et al. [15] showed that the size and location of the airborne droplets are strongly influenced by the ventilation system. The effects of relative humidity and wind velocity on the evaporation and dispersion of the droplets in an open space were evaluated by Li et al. [16] . By numerical simulation of sneezing in a room, Singh and Tripathi developed a model for droplets transmission [17] . Experiments of Chaudhuri et al. [18] showed that the droplets evaporation time depends on the ambient temperature. Although detailed simulation of the generation and dispersion of virus-carrying droplets is possible, the large number of parameters involved in this problem complicates such simulations. Further, in practice, many large environments with highly complex and varying configurations should be simulated. Hence, the computational time and cost of a purely computational approach would be forbiddingly high. To reduce the computational burden, techniques from machine learning (ML) can be combined with computational fluid dynamics (CFD). Such a hybrid approach has been 9 already applied to the problems in aero-propulsion [19] , heat and mass transfer [21 ,20] and spread of coronavirus [10] . In all these, substantial reduction in the computational expenses (up to 99%) was achieved, while predictions remained quite accurate. Given the size, variations and complexities of the indoor environments in which virus is spread, machine learning could significantly facilitate the analysis. Artificial Neural Networks (ANN) offers an accurate means of prediction on the basis of computational data. The capabilities of this method are already well-demonstrated. For instance, comparing CFD simulations of natural convection in a triangular cavity and ANN indicated that both soft codes predict the stream function and temperature parameters correctly [22] . Similarly, CFD simulation of natural convection of cold column cylinders was compared with a Multi-Layer Perceptron (MLP) network [23] . The training algorithm was resilient back-propagation, and the results indicated a good agreement between the two techniques. Further, ANN can be applied to more complex physics, such as multiphase flows [24] and porous media [26, 27] . For example, it was demonstrated that the Levenberg-Marquardt Algorithm (LMA) could replicate the unsteady simulation of a chaotic movement of gas-liquid flow in a square cavity with less than 10 % error [2] . Further examples of combining CFD and ANN can be found in Refs. [20, 21, [25] [26] [27] [28] . This study aims to provide a prediction of the distribution of the fine droplets and aerosols (1~250 ) generated during sneezing. A recent work of the authors showed that these droplets are most responsible for the transmission of the virus as they could travel a long distance before deposition/evaporation [10] . The indoor environment of a bus as an example of a realistic high-risk environment with complex geometry is considered. The influences of ambient temperature and humidity, as well as the effects of ventilation, are investigated using a novel hybridised CFD/ML tool. The inner space of a bus (Mercedes Benz, Tourismo (SKS E128I) with standard seats (Fig. 1a) is considered. Ambient air is injected from the fresh air duct and is collected by the ducts placed on the ceiling ( . Three different air temperatures and relative humidity (RH) are considered for × 36) simulations. The ventilation system circulates the air in the bus without affecting the air temperature, and RH. Air leakages are neglected, and the bus is assumed to move with a constant velocity. The range of flow rate for this type of bus is between 800 and 6,600 m 3 .hr -1 , based on the DENSO LD9 HVAC unit [29] [30] [31] [32] . Further, the entire bus is assumed to be under thermal equilibrium, and the heat generated by the sneezing person is ignored. Also, only one sneezing passenger inside the bus is considered. Fig. 1b shows the location of the sneezing person in the bus. This passenger is located in the last row of sits and on the left side of the bus, sneezing horizontally and directly forward. This position allows direct transmission of the droplets to the entire internal volume of the bus. The sneezing characteristics are summarised in Table 1 . [40] An in-house code (in Python) was developed to solve the governing equations with second-order accuracy. One hundred and twelve cores were used with a graphic card for simulations. The PISO method of velocity and pressure coupling was employed for turbulent modelling, and the SIMPLE method was used for laminar flow. The obtained solution was time-dependent with a 0.001s time step over 600 seconds. ANSYS 2021R1 was used for data post-processing, and the mesh was generated in ANSYS ICEM and implemented directly to the code. In order to enhance the accuracy, the According to Fig.2 , LES has been used for up to 500 mm from the sneezing person. This is because the high momentum sneezing flow can set a strong turbulent flow around the sneezing person. Between 500 ~ 1500 mm, a turbulence model with low intensity of turbulence was used to model the droplets. It should be noted that evaporation of droplets mostly occurs in this zone. There exists a relaxing zone stretched from 1500 ~ 12000 mm of the sneezing person [12 ,10] . According to Fig. 2 , a numerical simulation is performed in the area marked in red while machine learning is trained at each time step. Validation of machine learning with a numerical solution is done in the green area, while predictions are made by machine learning throughout the entire range marked with blue. The following assumptions are made throughout the proceeding analyses. 14 1-The seats are considered as solid and smooth surfaces. 2-The walls, ceiling, and windows are assumed to be adiabatic. 3-The head and body movements of the sneezing person during sneezing have been ignored. 4-Sneezing occurs once at the beginning of the analysis, and then the interactions of the ventilation system and droplet distribution are evaluated. The droplets are assumed to be initially spherical particles. 6-In the calculation of droplet shapes, the effects of collisions with surfaces are negligible. It is noted that most of the findings of the current study remain valid for other forms of public transport (e.g. train) and stationary environments such as classrooms. The investigated multiphase flow contains droplets in the range of 1~250 (liquid phase), air (gas phase), and vapour phase. The mathematical model includes the conservation equations in the Eulerian framework for any fluid, expressed by [41, 42] : Conservation of mass: Balance of momentum in x, y and z directions [43, 44] ∂( * * * ) The balance of energy for the mixture model of multiphase flow is written as [43, 45, 46] : LES was employed to simulate the vortical flow set by sneezing. The continuity and momentum equations in this method are expressed as [47] ( ∂ In this equation, the term S defines the body forces applied to the fluid. The stress tensor is expressed by where is the rate of the strain tensor. To calculate the turbulence kinetic energy (k), the following one equation eddy-viscosity model (OEEVM) was employed. The kinetic energy in the LES sub-grid model is defined by [48] : where . Considering Boussinesq approximation, a simple correlation for the incompressible Reynolds stress tensor is given * = ( │ ) -〈 〉 * by: The transfer of data between LES and RANS is based on the procedures in Refs [49] [50] [51] and are not further discussed here. The saliva droplets leave the mouth with a time-dependent sneezing profile. The considered exit velocity profile and pressure variation of sneezing are based on that in Ref. [6] , and therefore the maximum exit velocity was set to 100 (m.s -1 ). Further, there is a no-slip boundary condition for the flow inside the mouth. In keeping with the literature [52, 53] , the exit sneezing flow is set to have a relative humidity of 75%, while the droplets' temperature is between 34 and 36 . Also, the droplets are assumed to have a uniform distribution with a perfectly spherical shape at the exit ℃ ℃ moment [5, 6] . The physical properties of the saliva are based on those in Refs. [6, 54, 55] . In each time step, the droplet diameter was calculated following the methodology introduced in Refs. [12, 43] . Further, the merging of the droplets was modelled on the basis of the analysis detailed in Refs. [56, 57] . The droplets generated by sneezing have a size in the range of 1 and 1000 while, only those with diameters between 1 and 250 could travel a long distance [10, 34] . The droplets' shape changes significantly during the 18 advection process affecting their distribution, stability and suspension time. The droplet shape is calculated by considering the external and internal forces in each time step. These include drag force, droplets breakup, and collisions. Fig. 3 shows the variations of droplet mass and drag force with time and Reynold number, respectively. Expectedly, the droplet mass decreases over time [58] while drag force is calculated [38, 59, 60] . In Fig. 3a , the outcomes of adaptive method (boundary layer adaptation with geometry) have been compared with those provided by two other methods (bulk zone method [61, 62] , regular method [63] ) extracted from literature. Clearly, the adaptive method offers predictions closely matching the benchmark data. Similarly, Fig.3b shows that the adaptive methods can successfully predict the drag force in comparison to spherical shape [64] and deformable spherical shape [65, 66] . Temperature distribution inside the droplet is central to the modelling of droplet evaporation. Temperature distribution models are categorised based on the surface temperature distribution and ambient temperature. The current study utilises four different models, including constant temperature model [68] , uniform temperature model [69] , non-uniform temperature distribution [70] and non-uniform temperature distribution with boundary layer [71, 72] , to calculate the droplet evaporation rate and its temperature distribution. The results have been compared with those reported in Ref. [73] (see Fig. 4a ). Evidently, the non-uniform temperature distribution with boundary layer model [72] and adaptive computational domain on the 20 non-spherical shape of the droplet could provide the most accurate estimation of the droplet temperature. The droplet mass variation with evaporation alters the droplet mean diameter, which then changes the droplet deposition rate. Here, the calculation of droplet internal temperature is on the basis of the boundary layer sensitivity to accurately calculate the droplet temperature distribution. Indeed, the computational domain next to the outer droplet shell adapts itself to the droplet shape with an assumed isothermal boundary. Fig. 4b shows a variation of the mean droplet evaporation rate with the mean surface area. The results shown in Fig.4 confirm that the new model offers an accurate prediction of the droplets evaporation rate. The mass transfer for any fluid is governed by the following equation [43, 45] . For vaporisation processes, if the droplets kinetic energy and the viscous dissipation are ignored, the mass transfer on the droplets interface can be approximated by [74] : Combining the enthalpy and mixture velocity for integration of the phase and mass-weighted variables yields [45] : In equation 48, is the drift velocity for the m * th phase, which can be defined as . The evaporation of droplets is modelled According to Eq. (20) the decreasing flux of evaporating mass can be estimated by balancing energy and mass on both sides of the interface [43] . Stefan's correlation was used to find the concentration of vapour on the interface of the droplet; the mass concentration for each phase is given by [15, 62, 75, 77 , 78] The concentration equation is now substituted into the body force equations. This yields The droplet evaporation rate is then calculated by the following equation (see Ref. [73] for further details). In this study, four standard and robust artificial neural networks have been used for prediction. In several experiments, these networks' accuracy has been compared in our problem, and the best network has been selected. In the following, the examined networks are briefly introduced. Of course, since the Multi-Layer Perceptron (MLP.) network has provided the best results, this network will be presented in more detail. An MLP network consists of simple operational elements, most of which can work in parallel. These elements, called neurons, are inspired by biological nervous systems. An artificial neuron comprises inputs, outputs, weights, biases, and an activation function. The neural network configuration is determined by the characteristics of each neuron and how they are connected. MLP. has been used to estimate complex functions in a variety of contexts and articles [79] . The network generally consists of the input layer, the hidden layer, and the output layer. If needed, several hidden layers can be used. Data is injected into the network through the input layer. If is the number of input features of the model, this layer contains ( ) ( ) + 1 neurons. One of the neurons is used as a bias, and the other neurons are used to receive inputs. The hidden layer consists of neurons, ( ) + 1 one of which acts as a bias, and the other neurons, using equation 23, apply an activation function to the weighted sum of the outputs of ( ( )) input layer. In this equation, and are the values of the -th neuron of the input layer and the -th neuron of the hidden layer, respectively. is the output of -th neuron of the output layer. and are the weights of the bias and -th hidden neuron connection with the -th ( ) 2 0, 2 , output neuron, respectively. The activation function of output layer neurons can be different from that of hidden layer neurons. One of the essential characteristics of a neural network is its ability to learn. We used the supervised learning method to train the network. First, connection weights are randomly assigned. After that, the output of the hidden layer neurons and then the output layer neurons are calculated using equations 26 and 27. Afterwards, the weights change according to a specific learning algorithm. Weight changes should be made so that the neuron outputs are close to the actual outputs. The process of changing the neuron weights to achieve the desired output is called learning. A loss function is used to evaluate learning performance. The goal of learning is to reduce the amount of loss function to near zero. Therefore, an optimisation algorithm is used to find the weights that minimise the loss function. In this paper, the error backpropagation method is considered as the optimisation algorithm and the Mean Square Error (MSE) criteria as the loss function. The MSE is calculated by equation 28. The computational grid production in this analysis features significant complexities. The presence of droplets with variable geometry and micron size, the vapour layer and the large dimensions of the bus necessitates taking a novel approach to producing an adaptive grid. Hence, a fully automated algorithm based on the Three-Dimensional Multi-Block Advanced Grid Generation System (3DMAGGS) [80, 81] was utilised. In this method, the block separation pattern was applied to separate the areas required for re-meshing. An independent algorithm, based on the vertical velocity profile, created boundary layer grids around the droplets. Fig. 5 shows the grid independency diagrams for different modes. It should be noted that this grid adaptation varied at each time step according to the droplet geometry. Therefore, the grid matching code was based on the droplet geometry to redefine the new grid based on the modified geometry. All developed computational models of sneezing were validated against the existing data extracted from the literature. Fig. 6 provides a comparison between these datasets. Evidently, the current simulation data very well match those extracted from the references. The propagation spectra of different sneezing droplets are depicted in Fig. 7a , illustrating the significant effects of the environmental parameters. It has been already shown that the majority of the particles produced by sneezing have a short lifetime and cannot travel a long distance [10] . Here, an air-floating droplet is defined as a measure for 95% of particles with similar diameters, which have the following features. 1-Their standard momentum deviation is below 3%. 3-The standard deviation of their mass falls below 10%. Two types of droplets can be found in the range of 1-250 μm: those initially produced at this range and those experiencing evaporation and breakdown along the path. The former rapidly joins other droplets or evaporate due to their high velocity when leaving the mouth, while the latter mainly constituents the transmitted droplets to the other parts of the bus. The important point in this diameter range is the direct relation between the deposited droplets and their dimensions. With the decline in diameter, deposition of the droplets smaller than 25 μm becomes progressively more unlikely. Many references have considered the aerosols with sizes below 5 μm to be largely non-depositing particles [15, 84] . Environmental parameters can influence the propagation of droplets generated by sneezing through controlling the evaporation rate and droplet size [18, 85] . These effects are investigated in the current section. In Fig.7b , the interior space of the bus is divided into a few distinctive zones based on the distribution of sneezing droplets. It is observed that decreasing the diameter of droplets increases the average suspension time and distance. According to Fig.8a , when the passenger sneezes at the rear end of the bus, the entire interior space of the bus is affected. The provided supplementary data show the average velocity, concentration, and suspension time for all droplet diameters for the duration of 600s after sneezing. The results indicate that as the droplet diameter decreases, the suspension time increases. Also, droplets become smaller as longer distances are taken from the sneezing passenger [10] . The heating, ventilation and air conditioning system in the bus can significantly influence virus transmission by adjusting the inflow velocity, temperature, and relative humidity. In this investigation, the air is circulated with different velocities, temperatures, and relative humidity inside the bus in accordance with the data provided by the bus manufacturer [30, 32] . to the interactions between the seats and ventilation flow is evident. In particular, Fig. 8 shows that intensification of the ventilation rate results in the development of a series of vortical structures between the seats. Importantly, these vortical flows constantly evolve along the bus, and there is, essentially, no repeated pattern. It will be shown that this complex flow significantly influences the virus transmission in the bus. Temperature dominates the evaporation rate, and therefore the droplet volume is proportional to the ambient temperature. The air temperature may also have an effect on the droplet viscosity and thus on the shell deformation. Here, assuming constant relative humidity and airflow inside the bus, the effects of temperature on the droplet diameter and floating time is investigated. Here, 5 million droplets, produced at the onset, are analysed using the developed machine learning model (see Section 3.3). This method lowers the computational costs and, more importantly, enables pattern recognition. The evaporation rate of the propagated droplets is dominated by the surface area exposed to the flow. During the first second after sneezing, the maximum evaporation rate was observed due to the larger dimensions of the droplets and their breaking down. Fig. 9a shows the mean values derived for the droplet shell at the first two seconds subsequent to sneezing. Clearly, there is a substantial geometrical variation in the droplets. Further, increases in temperature significantly enhance the evaporation rate. Fig.9b illustrates the concentration of droplets ( ) in various 1 < < 250 locations throughout the bus. The results show that different locations are exposed to a range of concentrations at different times. Additionally, ventilation noticeably increases the rate of dispersion. As depicted in Fig. 9 b, the highest concentration of droplets is found near the person. There is no A decline in the droplet diameter results in the enhancement of the floating time. Fig. 10a shows the percentage of the droplets removed from the environment (through evaporation) at the ventilation rate of 0.1 m.s -1 and relative humidity of 20% for diameters of 1~100 μm and 100~250 μm. As expected, the results indicate that a greater percentage of droplets are eliminated at higher ambient temperatures. Fig. 10b shows the average suspension time for 95% non-spherical droplets in different ranges. The period of suspension is determined by internal and external forces acting on the droplet volume. Since the geometry of the droplet changes over time by the drag force, the suspension time is proportional to the drag force. Additionally, considering the impact of temperature on the droplet viscosity and its effect on the rate of evaporation, it is clear that the geometry of the suspended droplet is determined by the change in the outer shell (due to drag) and the rate of mass reduction. Fig.10b depicts the considerable effects of temperature upon the suspension time of droplets. Along with directly affecting the distribution of heavy particles in the near field region, ventilation can also have an effect on the droplets' diameter. Further, increasing the airflow velocity speeds up evaporation and increases the frequency of the droplets of less than in diameter. The flow 100 velocity inside the bus may also affect the forces acting on the droplets altering the buoyancy and drag forces. This, in turn, influences the particle lifetime and transmission range. Additionally, as shown in Fig. 8 , increasing the airflow velocity intensifies the vortex formation between the seat rows in the bus. These vortices can alter the path of the droplets, as droplets, less than in diameter tend to follow 250 the flow [88] . Increased airflow rates also help to disperse the heavy droplets, giving them a longer time to evaporate and shrink in size. Table 3 summarises the results for droplet velocity, suspension, and deposition. This table shows that the droplets with diameters less than were distributed evenly 20 throughout the bust in 600 seconds, with over 95% of them remaining floating. As an important observation, it appears that boosting the ventilation velocity can prolong the suspension time. Table 3 The average deposition range and floating time for different droplets' diameters. The transmission of contaminated droplets in the bus is constantly interrupted by the deposition process and re-circulation of air set by the ventilation system. Changes in the ventilation velocity can alter the duration of droplet transmission and suspension. Nonetheless, the details of these modifications are complicated due to the geometrical complexity of the current problem. It is recalled that the ventilation velocities considered in this study have been inferred from the bus manufacturer's data [32, 34] . Fig. 11 shows the average evaporation rate and concentration of droplets, indicating that for the evaporation rate of droplets remains almost constant = 0.1 / , this increase is 15~17% and 10~12%, respectively. The change in (50~120 ) and (120~250 ) the droplets' volume as a function of air velocity is depicted in Fig. 11a . According to this figure, increasing the airflow velocity causes an ascending trend in the droplets' evaporation rate, which then stabilises at a constant value. The slope of the diagram changes due to the interactions between the rate of surface evaporation and the heat flux across the droplet surface. According to Fig. 11b , increasing the ventilation velocity significantly increases the rate of droplet dispersion (up to 41% in zone 2 and 36% in zone 3 compared to without ventilation). The variations in ventilation velocity affect the droplet concentration in zone 1 by up to 50% when . However, the concentration of droplets in zone 2 increases over time, and ventilation = 0.5 / velocity contributes to reaching the maximum concentration in a short time ( 59 × 10 -4 grams/mm 3 in 44 seconds). In fact, ventilation velocity has a significant effect on increasing the droplet concentration in all selected zones except for zone 1. Changes in the geometry of droplets modify the drag force and convective heat transfer on the surface. Increasing the surface area of a droplet relative to its volume can aid increasing the rate of evaporation, and therefore, the rate of evaporation for nonspherical droplets is higher than that for spherical droplets [58, 89] . When the droplet diameter falls below 200 , the geometrical variance is negligible [44] . Hence, according to Fig. 12a Ambient humidity can affect the transmittance of contaminated droplets for a variety of reasons. Droplet evaporation is heavily influenced by humidity. Increased humidity reduces evaporation by reducing the difference in vapour concentration between the fluid and gas phases. This has the potential to increase the lifetime of the droplet and subsequently extend its transmission range. The changes in evaporation as a function of the mean diameter of the droplet are depicted in Fig. 13 for various relative humidity values. Reduced humidity increases the evaporation rate. By decreasing the concentration gradient around the droplets, the average rate of evaporation drops and therefore increases in humidity reduce the evaporation rate. Further, there is a relationship between the ambient temperature and relative humidity that affects the vapour concentration. These two factors are responsible for the reduction in the mass of droplets. (Table 4) , the concentration of droplets can be significantly reduced. The optimal thermal comfort (as defined by ASHRAE standard 55) is shown in Table 4 when the droplets concentration is minimised. According to Table 4 , the optimal combination of humidity, temperature, and velocity can reduce the concentration of droplets (1md p250m) by 13.62% when compared to the ASHRAE standard condition. A hybridised CFD/ML predictive tool was used to analyse the distribution of virus-containing droplets generated by human sneezing in public transport (a bus). The tool was rigorously validated and then utilised to explore the influence of environmental parameters including temperature, relative humidity and ventilation on the Spatio-temporal distribution of the droplets. In particular, emphasis was put on the droplets suspension time and transmission speed through the interior space of the bus. An earlier study of the authors showed that the droplets with diameters below 250 are most responsible for virus transmission. Hence, this investigation focused on the droplets with diameters between 1 and 250 . The key findings of this work could be summarised as follows.  When the ventilation system of the bus is switched off, the droplets (1~250 take more than ) 200 seconds to reach the front end of the bus. However, with the ventilation system running, this time drops to 22 seconds for , 15 seconds for , and 1.8 = 0.1 .  It was found that the droplets with diameters between 1 and 20 remain suspended in the air without deposition or evaporation for an extended period of time.  Environmental parameters can have significant effects on the concentration of the droplets with diameters between 1 and 250 . For instance, increasing the relative humidity by 10% can result in 30% increase in the droplet concentration at the farthest distance from a sneezing passenger.  Configuration of the ventilation system design can aid or hinder dispersion of small infectious droplets. According to the results, the optimal ventilation system can be designed by injecting fresh air from the ceiling and sucking air from the floor. Finally, it is essential to note that the number of viruses contained in the droplets can affect the risk of disease trasnmission. This biological factor was not included in the current analysis and therefore generalisation of the results should be done mindfully. A global health perspective on SARS-CoV-2-hazards, disaster and hope Influence of wind and relative humidity on the social distancing effectiveness to prevent COVID-19 airborne transmission: A numerical study Physics of virus transmission by speaking droplets On air-borne infection. Study II. Droplets and droplet nuclei The flow physics of COVID-19 Sneezing and asymptomatic virus transmission Turbulent gas clouds and respiratory pathogen emissions: potential implications for reducing transmission of COVID-19 Transmission of airborne virus through sneezed and coughed droplets Aerosol persistence in relation to possible transmission of SARS-CoV-2 Prediction of the spread of Corona-virus carrying droplets in a bus-A computational based artificial intelligence approach Transmission of COVID-19 virus by droplets and aerosols: A critical review on the unresolved dichotomy Characterizations of particle size distribution of the droplets exhaled by sneeze Numerical Study of Three Ventilation Strategies in a prefabricated COVID-19 inpatient ward Fluid dynamics simulations show that facial masks can suppress the spread of COVID-19 in indoor environments Numerical investigation of aerosol transport in a classroom with relevance to COVID-19 Dispersion of evaporating cough droplets in tropical outdoor environment Application of National Aerosol Facility (NAF) in Designing of a Ventilation System for Isolation Rooms to Minimize Interpersonal Exposure of Sneezing/Coughing Modeling ambient temperature and relative humidity sensitivity of respiratory droplets and their role in determining growth rate of covid-19 outbreaks State prediction of an entropy wave advecting through a turbulent channel flow Analysis of transport processes in a reacting flow of hybrid nanofluid around a bluff-body embedded in porous media using artificial neural network and particle swarm optimization Application of Machine Learning to Investigation of Heat and Mass Transfer Over a Cylinder Surrounded by Porous Media-The Radial Basic Function Network Prediction of flow fields and temperature distributions due to natural convection in a triangular enclosure using Adaptive-Network-Based Fuzzy Inference System (ANFIS) and Artificial Neural Network (ANN). International Communications in Heat and Mass Transfer Data-driven time-dependent state estimation for interfacial fluid mechanics in evaporating droplets Machine learning analysis for quantitative discrimination of dried blood droplets Direct numerical simulation and artificial neural network modeling of heat transfer characteristics on natural convection with a sinusoidal cylinder in a long rectangular enclosure A machine learning approach to the prediction of transport and thermodynamic processes in multiphysics systems-heat transfer in a hybrid nanofluid flow in porous media A machine learning approach to predicting the heat convection and thermodynamics of an external flow of hybrid nanofluid Artificial intelligence prediction of natural convection of heat in an oscillating cavity filled by CuO nanofluid The influence of school bus ventilation scenarios over incabin PM number concentration and air exchange rates. Atmospheric Pollution Research DENSO. Bus Air Conditioning Bus HVAC energy consumption test method based on HVAC unit behavior. International journal of refrigeration Air conditioning systems Prediction of the spread of Corona-virus carrying droplets in a bus-A computational based artificial intelligence approach Numerical modeling of the distribution of virus carrying saliva droplets during sneeze and cough Nozzle design influence on the steam-driven ejector A CFD Study of [C2mim][CH3SO3]/Al 2 O 3 Ionanofluid Flow and Heat Transfer in Grooved Tubes The δ-ALE-SPH model: An arbitrary Lagrangian-Eulerian framework for the δ-SPH model with particle shifting technique The effect of variable temperature and location on relative thermal conductivity (RTC) on the heat pipe in the presence of AL2O3 nanoparticles: Numerical and optimization approaches Hydrodynamic pattern in drying saline droplet with suspended nanoparticles Three-dimensional numerical simulation of violent free surface deformation based on a coupled level set and volume of fluid method. Ocean Engineering Computational methods for multiphase flow Multiphase flow dynamics: Fundamentals Computational techniques for multiphase flows Multiphase flows with droplets and particles Turbulent multiphase flows with heat and mass transfer Multiphase flow handbook Turbulent flows The physics of fluid turbulence. Chemical physics Embedded LES-to-RANS boundary in zonal simulations Hybrid LES-RANS using synthesized turbulent fluctuations for forcing in the interface region. International journal of heat and fluid flow On interface issues in LES/RANS coupling strategies: a method for turbulence forcing Airflow dynamics of human jets: sneezing and breathing-potential sources of infectious aerosols Physico-chemical characteristics of evaporating respiratory fluid droplets Effect of sneezing on the flow around a face shield Development of a Simulator to Model the Spread of Coronavirus Infection in a Closed Space Pillar-induced droplet merging in microfluidic circuits Improved droplet breakup models for spray applications Kinetics of droplet evaporation Particles, bubbles & drops: their motion, heat and mass transfer Drag coefficients of viscous spheres at intermediate and high Reynolds numbers Liquid flow pattern transition, droplet diameter and size distribution in the cavity zone of a rotating packed bed: A visual study The comparative investigation of three approaches to modeling the natural convection heat transfer: A case studyon conical cavity filled with Al2O3 nanoparticles A double droplet trap system for studying mass transport across a dropletdroplet interface Drag and lift forces acting on a spherical water droplet in homogeneous linear shear air flow Quasi-steady shape and drag of deformable bubbles and drops A numerical method for the calculation of drag and lift of a deformable droplet in shear flow On the transient flow inside and around a deforming millimetre class oil droplet falling under the action of gravity in stagnant air A water droplet evaporation and temperature model An Investigation of Liquid Droplet Evaporation Model Used in Multiphase Flow Simulation A simplified non-isothermal model for droplet heating and evaporation. International communications in heat and mass transfer Stable cluster of identical water droplets formed under the infrared irradiation: Experimental study and theoretical modeling Evaporation and movement of fine droplets in non-uniform temperature and humidity field Effect of heat conduction on droplet life time and evaporation rate under forced convection at low temperatures. Experimental Thermal and Fluid Science Mathematical modeling of the spread of the coronavirus disease 2019 (COVID-19) taking into account the undetected infections. The case of China. Communications in nonlinear science and numerical simulation What ENT doctors should know about COVID-19 contagion risks With COVID-19, modeling takes on life and death importance Modeling the role of respiratory droplets in Covid-19 type pandemics Understanding Transmission Dynamics of COVID-19-Type Infections by Direct Numerical Simulations of Cough/Sneeze Flows Neural network design for engineering applications The Three-Dimensional Multi-block Advanced Grid Generation System (3DMAGGS) A large-scale parallel hybrid grid generation technique for realistic complex geometry Violent expiratory events: on coughing and sneezing Effects of Ambient Temperature and Humidity on Droplet Lifetime-A Perspective of Exhalation Sneeze Droplets with COVID-19 Virus Transmission SARS-CoV-2 concentrations and virus-laden aerosol size distributions in outdoor air in north and south of Italy Evaporation-induced particle microseparations inside droplets floating on a chip Exhaled droplets due to talking and coughing The size and the duration of air-carriage of respiratory droplets and dropletnuclei On the dynamics of vortex-droplet interactions, dispersion and breakup in a coaxial swirling flow The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.