2.1 PSM measuring theory

A PSM measures the total particle number concentration of the activated particles. The sampled aerosol flow is mixed with a high-temperature flow containing saturated diethylene glycol coming from the saturator, and then the mixed flow passes through a low-temperature growth tube. The particles large than a specific diameter can overcome the Kelvin effect and grow to larger sizes due to the condensation of supersaturated diethylene glycol. The detection efficiency is mainly determined by the particle diameter and the saturator flow rate. The chemical compositions and charging state may affect the detection efficiencies (Kangasluoma et al., 2013, 2016a) and lead to errors in the reported particle size distributions (Kangasluoma and Kontkanen, 2017). However, we mainly focus on the inversion method in this study and assume the detection efficiency is only size dependent at a certain saturator flow rate. Since the temperatures in the saturator and the growth tube are fixed, a higher saturator flow rate leads to a higher supersaturation ratio of diethylene glycol in the growth tube and hence higher detection efficiencies (Fig. 1a). See Sect. 3.1 for the details on how to obtain the detection efficiency curves. The detected total particle number concentration varies with the varying saturator flow rate when the particle size distribution is unchanged. The relationship between the detected total particle concentration, R, the saturator flow rate, s, and the particle size distribution function, n, can be expressed in the Fredholm integral equation of the first kind:

where R_i is the number concentration recorded at the ith saturator flow rate, s_i; d_p is the electrical mobility diameter since the calibrating particles are classified according to their electrical mobility; η is the overall detection efficiency determined by s and d_p, including the detection efficiency and the sampling efficiency; n is the probability density of particle number concentration (particle size distribution function), dN∕dd_p and N is the accumulated number concentration of particles smaller than d_p; and ε_i is the error in the recorded particle concentration at s_i.

Figure 1(a) The fitted detection efficiency curves according to calibration data. (b) The estimated kernel function curves according to the fitted detection efficiencies. The kernel function is equal to the derivative of the detection efficiency with the respect to the saturator flow rate.

Download

There are many potential sources of the error, ε. For instance, the uncertainties in the calibrated detection efficiencies, the systematic errors caused by the non-ideal fitting formula of the detection efficiency curves, the CPC counting uncertainties, the uncertainties in the supersaturation ratio due to fluctuations in the flow rate and temperature, and the unstable aerosol source will all contribute to the difference between the detected number concentration and the expected particle concentration assuming there is no error.

As shown in Fig. 1b, the kernel function of the PSM, K, is defined as the derivative of the detection efficiency, η, with respect to the saturator flow rate, s. The area of the kernel function is equal to the difference between the detection efficiencies at the maximum and minimum saturator flow rates. Here we define r as the derivative of the detected number concentration, R, with respect to s. According to Eq. (1), the relationship between r and s is also a Fredholm integral equation of the first kind:

where r_m is the r at the mth saturator flow rate, s_m; and ${\mathit{\epsilon }}_m^{\prime }$ is the error in r_m. Although r is theoretically defined as the derivative of R, practically one can only approximate r using the difference between two adjacent R_i over the increment in s_i and approximate s_m with the mean value of the two corresponding s_i. These approximations also contribute to the uncertainties, ${\mathit{\epsilon }}_m^{\prime }$ in addition to the aforementioned sources for ε_i.

When using a PSM to determine particle size distributions, the PSM records the varying total particle concentration, R_i, and the corresponding saturator flow rate, s_i. The saturator flow rate may vary continuously in the scanning mode or fixed at different flow rates in the stepping mode. The particle size distributions are recovered using the recorded relationship between R_i and s_i or the relationship between the approximated r_m and s_m.

The sizing ability of the PSM can be described using the size resolution. Similarly to defining the sizing resolution of a differential mobility analyser (DMA, Flagan, 1999) to classify particles according to their electrical mobility, we define the resolution of a PSM as

where Res is the resolution at s^∗, s^∗ is the peak saturator flow rate of a kernel function, and Δs is the full width at half maximum of the kernel function peak. A relationship between the saturator flow rate and the electrical mobility diameter is defined to straightforwardly relate the resolution and the particle diameter. The peak saturator flow rate, s^∗ is defined as the corresponding saturator flow rate of the particle diameter. This definition is similar but different from the definition using the saturator flow rate at the half maximum detection efficiency in Lehtipalo et al. (2014) and in the commercialized PSM. The sizing resolution of a PSM can be estimated according to the relationship between s and d_p, as shown in Fig. 2. However, the resolution alone is not sufficient to indicate the possible reported size range when the PSM is measuring monodisperse particles because the kernel functions are asymmetric and the inversion method also affects the reconstructed peaks. One should especially keep in mind that the PSM does not measure particle diameter because the relationship between s and d_p is only a definition rather than an intrinsic correlation. A PSM only records the varying particle concentration against the varying saturator flow rate (as indicated in Eqs. 1 and 2). One can only obtain the particle diameters via proper data inversion.

2.2 The stepwise method

The resolution of the PSM is assumed infinite in the stepwise method. Thus, the integral equation relating n and r collapses into a one-to-one corresponding relationship (Lehtipalo et al., 2014).

where n_m is the particle size distribution function (dN∕dd_p) at d_m; d_m, d_i, and d_i+1 are the corresponding half-maximum cut-off diameters of s_m, s_i, and s_i+1; and s_m is the mean value of s_i and s_i+1. The relationship between particle diameter and the saturator flow rate is determined using the saturator flow rate at the half maximum detection efficiency (Lehtipalo et al., 2014). The stepwise method does not magnify the relative error in measurement since it is a one-to-one inversion method. However, the inverted results using the stepwise method are perhaps non-negligibly affected by the relatively low resolutions of the PSM.

Figure 2The saturator flow rate at kernel function peak and the resolution as functions of the particle diameter. Note that the resolution is defined using the saturator flow rate, but the horizontal axis is shown in the particle diameter corresponding to the peak saturator flow rate for a more straightforward understanding.

Download

2.3 The kernel function method

The kernel function method assumes that the particle size distribution can be approximated using several particle size bins and the detection efficiencies of particles in each size bin are the same. The mathematical description of this approximation is

where d_j is the representing particle diameter of each size bin; J is the number of d_j; n_j is the particle size distribution function (dN∕dd_p) at d_j; Δd_j is the length of each size bin; and I is the number of R_i. The symbol of ≈ is to emphasize that Eq. (5) is an approximation even if there are no measuring errors because it approximates the integral with a finite discrete sum and estimates r_m using the recorded R_i. Using a matrix, Eq. (5) can be rewritten as

where

The upper case letters in Eq. (6) indicate the dimensions of the matrix and the vectors, while the lower case letters in Eq. (7) represent the corresponding element. The particle size distribution is obtained via solving Eq. (6) using the non-negative least squares method.

2.4 The Hagen & Alofs method

The H&A method (Hagen and Alofs, 1983) was proposed to deal with the multicharging correction problem when using a DMA. It can also be used to solve the PSM inversion problem. Similarly to the kernel function method, a discrete sum is used to approximate the integral:

Equation (9) is the vector form for Eq. (8) and P is the matrix relating n_j and R. We use the symbol of = in Eqs. (8) and (9) rather than ≈ because the H&A method requires a J much larger than I. One should increase J if the error in approximating the integral with the discrete sum is still large. Usually, J is determined as 30 times that of I considering the computational expenses. However, Eq. (8) itself is not solvable because there are more unknown variables than the equations. Thus, additional constraints are required. The H&A method assumes that any n_j can be approximated using n_i, i.e.

where f is the function relating n_j and n_i (n_i is a vector); n_i is the particle size distribution function at d_i; n_j is estimated using more than one single n_i; and Eq. (11) is the vector form for Eq. (10). The determination of d_i is theoretically arbitrary as long as the number of d_i is the same as the number of R_i. For the details on determining f, please refer to Hagen and Alofs (1983).

Similarly to the kernel function method, the relationship between the particle size distribution and the number concentration recorded by the PSM can be described in the vector form:

P and F are determined according to Eqs. (8)–(11) and thus Q is determined by η, f, and Δd_j. One can directly solve Eq. (10) (e.g. via Gaussian elimination) since Q is usually non-singular. However, in contrast to the matrix obtained from a DMA, the matrix Q in PSM inversion problem is usually not a positive-definite matrix because the detected particle concentration sometimes decreases with the increasing saturator flow rate due to random errors. Simply solving Eq. (12) often obtains negative values in particle size distributions. Thus, the non-negative least squares method is suggested to determine the particle size distribution in the PSM inversion problem. The H&A methods can also reconstruct the particle size distribution according to the relationship between r_m and s_m. However, using the kernel functions instead of the detection efficiencies does not necessarily improve the accuracy or precision of the results. On the contrary, we found that using the kernel functions usually leads to larger uncertainties than when using the detection efficiencies, because of the errors caused by approximating r_m.

The H&A method is theoretically more stable than the kernel function method because of the more accurate assumption of the true aerosol size distribution. However, the H&A method adapted for PSM inversion may still report size distributions with large uncertainties because of the least squares method. The computational expense of the H&A method is similar to that of the kernel function method because the rate-limiting step is to solve the least square question. Their low computational expense is an advantage over other non-linear inversion methods.

2.5 The expectation–maximization algorithm

The EM algorithm is a statistical method dealing with inversion problems with unobserved latent variables. An explanation of the EM algorithm can be found in Do and Batzoglou (2008). In the PSM inversion problem, the latent variable is R_i,j, defined as the contribution of particles with the diameter of d_j to the detected number concentration, R_i (Maher and Laird, 1985). The algorithm obtains the recovered particle size distribution using two steps: the expectation step and the maximization step. In the expectation step, the values of R_i,j are estimated according to Bayesian theorem:

In the maximization step, the particle size distribution function is estimated according to the maximum likelihood:

The EM algorithm obtains the recovered particle size distribution by repeating the expectation step and the maximization step until convergence. The convergence can be measured by the likelihood function (Maher and Laird, 1985). The values and the number of d_j are not limited when using the EM algorithm, and a larger J can reduce the errors in approximating the integral using the discrete sum. Thus, the EM algorithm is able to report particle size distributions with more size bins compared to the stepwise method, the kernel function method, and the H&A method.

The EM algorithm is more stable compared to the algorithms based on the least squares methods (Maher and Laird, 1985). The convergence of the EM algorithm has been proved (Dempster et al., 1977), but the convergence speed is not mathematically guaranteed. Compared to the kernel function method and the H&A method, the computational expense of the EM algorithm is much higher. In addition, the EM algorithm is a greedy algorithm in that the iteration is easily trapped in a local optimum. To start the first expectation step, an initial guess of the particle size distribution is required. We suggest the initial guess to be a vector of all of them. Note that the EM algorithm is sensitive to the initial guess and uses a recovered particle size distribution obtained from another method; i.e. the stepwise method does not necessarily improve the iteration results.

3.1 Experiments

Laboratory experiments using particles with known peak sizes or size distributions were conducted to test the inversion methods (Fig. 3). Sub-10 nm tungsten oxide particles were generated using a wire generator (Peineke et al., 2006; Kangasluoma et al., 2015). In the narrow peak measurement, the negatively charged particles were classified using a high-resolution Herrmann DMA. The sizing resolutions of the Herrmann DMA in the experimental conditions were no smaller than 25 (Kangasluoma et al., 2016b). Thus, the classified aerosols out of the Herrmann DMA can be approximately regarded as monodisperse. The relationship between the Herrmann DMA voltage and the classified particle size was calibrated using standard molecular ions (Ude and de la Mora, 2005). A TSI 3068B aerosol electrometer using the same aerosol flow rate with the PSM (2.5 L min⁻¹) was used as the reference.

In the wide peak measurement, the particle size distributions classified using a TSI nanoDMA have wider peaks than those generated in the narrow peak measurement. The aerosol and sheath flow rates of the nanoDMA were 2 and 10 L min⁻¹. It should be clarified that the particle size distributions classified using the nanoDMA in the wide peak measurement were still narrow due to the limitation of the nanoDMA. A lower sizing resolution either achieved by a higher aerosol-to-sheath flow ratio will cause the nanoDMA to be out of work due to significant turbulence. A half-mini DMA (Fernández de la Mora and Kozlowski, 2013) with calibrated penetration efficiency and a downstream Faraday cage electrometer (FCE) was used to measure the classified particle size distributions in parallel.

Figure 3The experimental set-up to calibration the PSM and test the inversion methods.

Download

The PSM (Airmodus A11) was calibrated using negatively charged tungsten oxide particles before the test. The experimental set-up for the calibration was the same as that used in the narrow peak measurement. The influence of the finite resolution of the Herrmann DMA on the calibrated efficiency curves was negligible. The saturator flow rate of the tested PSM varied from 0.05 to 1.3 L min⁻¹. This saturator flow rate range is wider than that of a typical PSM and obtains a complete kernel function curve of 3 nm particles. The maximum background noise of the PSM was approximately 1 cm⁻³, which was negligible compared to the usually detected particle concentrations. The detection efficiency is determined as the ratio of the particle number concentrations reported by PSM over the number concentration reported by the electrometer. The detection efficiency curves of the PSM were fitted using a function (Eq. 15) modified from the Chapman–Richards growth curve (Richards, 1959), which fitted better than other tested functions for the tested PSM.

where s_max is the maximum saturator flow rate (1.3 L min⁻¹); a, b, c, and d are the fitting parameters. If not specially mentioned, the PSM was fixed at 18 different saturator flow rates when measuring the particle size distributions in this study. This operation in the stepping mode was to avoid the potential uncertainties introduced in the scanning mode. The stability of the particle size distribution was monitored using the reference FCE during the relatively long measuring period.

3.2 Simulation

The performance of the four inversion methods was also studied using Monte Carlo simulations. The detection efficiencies used in the simulations were determined according to the calibrated efficiencies but slightly adjusted towards smoother curves. The uncertainties in practical calibration were neglected in the simulation.

Figure 4The recovered particle size distributions using different inversion methods when measuring monodisperse particles. FCE, SW, kernel, H&A, and EM are short for the Faraday cage electrometer, the stepwise method, the kernel function method, the H&A method, and the expectation–maximization algorithm, respectively. The number concentration detected by the reference FCE and the sum of recovered sub-3 nm particle concentration in each size bin are shown in the text. The size distributions in panel (d) were recovered using the sum of the recorded number concentrations in panels (a), (b), and (c), i.e. assuming the PSM was measuring 1.51, 2.41, and 3.93 nm particles simultaneously. The sub-3 nm particle concentrations reported by different inversion methods are summarized in Table 1.

Download

The particle number concentrations detected at different saturator flow rates were simulated using a certain initial particle size distribution. The random error, ε_i, was inserted into the simulated particle concentration, R_i. The random errors were determined experimentally. The relative random errors were larger than the statistical relative errors predicted using Poisson distribution (Iida, 2008; Kuang et al., 2012; Kangasluoma and Kontkanen, 2017) and independent of the particle concentrations at a certain instrumental configuration, indicating that random errors were governed by the fluctuations of the source and/or the instrumental parameters (e.g. flow rate). We used the mean relative random standard deviation observed in the experimental tests with 3.7 % as the representative value. In total 10 data points were assumed to be collected at each saturator flow rate. Thus, the random errors inserted into the simulated particle concentrations; i.e. the relative standard deviations of the mean particles concentrations, were assumed to be 1.2 % ($=\mathrm{3.7}\mathit{\%}/\sqrt{\mathrm{10}}$). A relatively large random error of 10 % obtained from the ambient measurements was also tested. The Monte Carol simulation was conducted 10 000 times using each inversion method to estimate the accuracy and precision of the recovered particle size distribution indicated by the mean values and the standard deviations of the inverted results.

4.1 Sizing accuracy

The inversion methods tested in this study, i.e. the stepwise method, the kernel function method, the H&A method, and the EM algorithm, are able to estimate the classified particle diameters when the PSM was measuring nearly monodisperse sub-3 nm particles. When the classified particle diameters were 1.51 and 2.41 nm, all of the four inversion methods could recover single peaks around the classified diameter (Fig. 4a, b). The size distribution reported by the stepwise method was the largest because the stepwise method does not account for the resolution of the PSM. Note that the peak diameters reported by the kernel function method and the H&A method were also affected by the selection of the particle size bins. The total particle concentrations obtained via inversion were similar to the number concentration detected by the reference FCE, except for the number concentration of 1.51 nm particles reported by the kernel function method.

Table 1The inverted particle concentrations (in cm⁻³) using different inversion methods and the total particle number concentration (in cm⁻³) recorded by the Faraday cage electrometer when measuring monodisperse particles.

Download Print Version | Download XLSX

None of the four inversion methods could size particles larger than 3 nm with relatively good sizing accuracies. When the classified particle diameter was 3.93 nm, the four inversion methods failed to report narrow peaks with peak diameters approximating 3.93 nm (Fig. 4c). This is because the PSM resolution for particles larger than 3 nm is low; i.e. the resolution was ∼1.0 when measuring the classified 3.93 nm particles (Fig. 2). The 3.93 nm particles contribute to the signal for 2.17 nm particles when using the stepwise method (inferred from Figs. 1 and 2). When focusing on the sub-3 nm particle size range, the kernel function method, the H&A method, and the EM algorithm reported nearly no sub-3 nm particles. However, the stepwise method reported a non-negligible number of sub-3 nm particles with a total number concentration of 1591 cm⁻³ due to the low size resolution.

We further tested the sizing ability of the four inversion methods using the sum of the recorded particle concentrations when the PSM was measuring 1.51, 2.41, and 3.93 nm particles (Fig. 4d). The kernel function method, the H&A method, and the EM algorithm distinguished the particles with different sizes, and the reconstructed peaks were similar to the corresponding peaks when the PSM was measuring monodisperse particles. The inverted results using the stepwise method was also unaffected by the summation; however, it was difficult to distinguish the isolated peaks from the recovered particle size distribution due to the broadened size distribution.

The size distributions of particles larger than 3 nm could not be successfully retrieved via data inversion because of the low resolution of PSM for these particles. However, it helped to recover sub-3 nm particle size distributions. Most of the reported particle sizes using the kernel function method, the H&A method, and the EM algorithm were larger than 3 nm when the PSM was measuring 3.93 nm particles (Fig. 4c). This estimation of particles larger than 3 nm ensured a relatively accurate sizing of sub-3 nm particle size distribution (Fig. 4d). Thus, we recovered the particle size distribution up to 5 nm using different inversion methods but focus only on the sub-3 nm size range.

4.2 Uncertainties using different inversion methods

The stepwise method, the kernel function method, and the H&A method may report false sub-3 nm particles when there are only particles are larger than 3 nm in the input aerosol. A particle size distribution with a peak diameter of 5 nm and nearly no sub-3 nm particles was simulated (Fig. 5a). The detected particle concentrations were assumed to fluctuate with a 1.2 % relative standard deviation due to measuring uncertainties (Fig. 5b). The EM algorithm reported nearly no sub-3 nm particles except for the smallest size bin at 1.16 nm (Fig. 5c). The expected values of particle concentrations in the bins smaller than 3 nm recovered using the H&A method were near zero; however, false sub-3 nm particle concentrations were occasionally reported (Fig. 5d). Compared to the H&A method, the size distribution recovered using the kernel function method was more unstable, especially in the sub-2 nm size range (Fig. 5e). The simulated uncertainty is the main cause of the false sub-3 nm particle concentrations reported by the H&A method and the kernel function method in Fig. 5. When assuming that there is no error in the particle concentration detected by the PSM, the H&A method and the kernel function method report nearly no particles in the sub-3 nm size range. In contrast to the H&A method and the kernel function method that reported false results due to their instability, the stepwise method reported false particle size distributions when assuming there are no uncertainties (Fig. 5f). This is because the stepwise method assumes a simple one-to-one relationship between the saturator flow rate and the recovered particle diameter instead of accounting for the wide kernel function peaks. For sub-1.5 nm particles, the non-zero mean particle concentration reported by the stepwise method is due to the simulated uncertainties.

Figure 5The recovered sub-3 nm particle size distributions simulated using the Monte Carlo method when the detected particles were larger than 3 nm. (a) The assumed true particle size distribution. (b) The simulated particle concentrations recorded by the PSM. The concentrations were assumed to fluctuate due to random errors. The particle size distributions were recovered using (c) the EM algorithm, (d) the H&A method, (e) the kernel function method, and (f) the stepwise method. The error bar represents the standard deviation of the recorded particle concentration or the recovered size distribution, and the shaded area indicates the range determined by 3 times the standard deviation. The dashed lines represent the inverted results assuming there were no random errors in the recorded particle number concentrations. Note that the scale of the vertical axis in panels (c)–(f) is different and the appearing possibility of recorded counts or the recovered size distribution is not uniform in the shaded area.

Download

The false sub-3 nm particle concentrations due to improper inversion methods were tested experimentally. Particles larger than 5 nm were classified using the nanoDMA (Fig. 6a). No sub-3 nm particles were reported using the EM algorithm and the H&A method. On the contrary, the kernel function method and the stepwise method reported approximately 3×10³ particles when the total particle concentration measured using the DMA–FCE system was approximately 2.4×10⁴. Based on both the simulating and experimental results, we conclude that the PSM may report false sub-3 nm particle size distributions when there are actually no sub-3 nm particles because of the uncertainties and the non-ideal data inversion methods, especially the stepwise method. Note that large particles with detection efficiencies that do not vary with the saturator flow rate do not lead to a bias in the recovered sub-3 nm particle concentrations. We examined this theoretical deduction experimentally using a PSM to measure ambient particles in the room air and the recorded particle concentration did not significantly vary with the saturator flow rate.

The performance of the four inversion methods in the sub-3 nm size range under the influence of larger particles was tested using a bimodal distribution (Fig. 7a). Similar particle size distributions are usually observed in the atmospheric new particle formation events (Jiang et al., 2011) and in flame (Tang et al., 2017). As shown in Fig. 7, the particle size distribution recovered using the EM algorithm had the highest accuracy and the smallest uncertainties among the four methods. The recovered particle size distribution using the EM algorithm had a slightly different shape compared to the initial distribution because the results were trapped in the local optimum. However, the differences between the recovered and the initial size distributions were the smallest. The standard deviations of the size distribution recovered using the H&A method and the kernel function method were relatively large due to the unstable least squares method. Because of a better assumption of the initial particle size distribution, the H&A method resulted in smaller uncertainties compared to the kernel function method, especially in the sub-2 nm size range. The size distribution recovered using the EM algorithm has higher accuracy and stability compared to both the H&A method and the kernel method because the one-to-one inversion method does not magnify relative errors.

Figure 6The experimental testing results of the four inversion methods when the PSM was measuring particles larger than 3 nm. (a) The particle size distribution detected by the reference half-mini DMA–FCE system. (b) The particle concentrations recorded by the PSM. The error bars indicate the standard deviations of the recorded particle concentrations. (c) The recovered particle distributions using different inversion methods.

Download

Figure 7The input and recovered sub-3 nm particle size distributions simulated using the Monte Carlo method. Note the vertical axes in panels (c)–(f) are not the same.

Download

The experimental tests using bimodal distributions agreed with the simulation results. The particles with a peak diameter at approximately 2.3 nm were classified using the nanoDMA. We added the observed number concentration to those detected in Fig. 6a (particles larger than 5 nm) to account for the influence of large particles. Unfiltered room air served as the make-up flow to provide background particles. As shown in Fig. 8, all the four inversion methods recovered the peak around 2.3 nm, while the results reported by the H&A method and the kernel function method were less smooth compared to the EM algorithm and the stepwise method.

Smoothing the size distribution recovered using the H&A method and the kernel function method into fewer size bins can reduce the uncertainties. We determined the number of the size bins of the recovered distributions according to the number of the fixed saturator flow rates. Too many size bins will lead to relatively large uncertainties, but the uncertainties can be reduced by sacrificing the resolution, i.e. reporting the size distribution in fewer bins. The size distributions recovered using the kernel function method were typically reported in 4–6 bins (Lehtipalo et al., 2014). This was achieved by assuming fewer discrete particle diameters in Eq. (5). Another option is to merge bins into fewer numbers after inversion rather than assume fewer bins at the beginning. Note that the H&A method cannot assume fewer discrete size bins at the beginning. Instead, the H&A method assumes an adequate number of size bins to guarantee a relatively smooth distribution (Eq. 8). As shown in Fig. 9, the standard deviations of the reported size distribution with fewer size bins were comparatively smaller than the corresponding standard deviations with more size bins shown in Fig. 7. The H&A method reported size distributions with smaller standard deviations than the kernel function method, and the kernel function reported in merged size bins had smaller standard deviations than the kernel function method using fewer size bins at the beginning. This is because approximating the true particle size distribution, which is usually a smooth curve, with fewer discrete size bins will lead to larger uncertainties. Thus, we suggest merging the recovered particle size distribution into a few size bins to reduce the uncertainties when using the H&A method and the kernel function method.

Figure 8The experimental testing results of the four inversion methods when the PSM was measuring sub-3 nm particles with the influence of larger particles. The particle number concentrations for the inversion and the particle size distribution detected using the DMA–FCE system were the sums of two separate experiments rather than real data obtained in a single experiment.

Download

Relatively large uncertainties were found when recovering sub-1.3 nm particle size distributions. A particle size distribution with an increasing dN∕dd_p as a function of the decreasing particle diameter, which is a typical particle size distribution observed in the atmospheric new particle formation events (Jiang et al., 2011), was used to test the four inversion methods (Fig. 10). None of the inversion methods reported a particle size distribution with relatively small uncertainties comparable to the inverted results shown in Fig. 7c, especially in the sub-1.3 nm size range. Similarly to the results for particles larger than 3 nm, the low resolution of particles smaller than 1.3 nm (Fig. 2) is possibly the cause of the large uncertainties. In addition, incomplete kernel function peaks and the relatively low detection efficiencies of sub-1.3 nm particles may also contribute to the uncertainties (Fig. 1).

Figure 9Comparisons of the inverted results using (a) the H&A method smoothing the particle size distribution via merging size bins; (b) the kernel function method smoothing the particle size distribution via merging size bins; and (c) the kernel function method assuming fewer discrete particle sizes in Eq. (5).

Download

The performance of the inversion methods under relatively large random errors was also tested. The relative standard deviation used in the above simulations, 3.7 %, was estimated according to laboratory experiments. The relative standard deviations of the recorded particle number concentration obtained from the atmospheric measurement were usually similar to the value obtained in the laboratory, indicating that the random errors were governed by instrumental factors. However, relatively large uncertainties in the recorded particle number concentrations were sometimes observed due to the unstable atmospheric aerosol source. Thus, we simulated the performance of the four inversion methods using a relative standard deviation of 10 %. It should be clarified that the value 10 % only characterizes the random errors in the CPC since it was estimated using the data when the recorded particle number concentration did not vary with the saturator flow rate. Compared to the results in Fig. 7 simulated using the same aerosol size distribution, the uncertainties in the recovered particle size distributions using the larger relative standard deviation of 10 % was larger (Fig. 11). The EM algorithm still reported smaller uncertainties compared to the H&A method and the kernel function method. Note the expected value of sub-2 nm particle size distribution recovered using the kernel method was close to the input size distribution when the uncertainty was 3.7 % (Fig. 7); however, the recovered size distribution in the sub-2 nm size range was non-negligibly overestimated when the uncertainty was 10 % (Fig. 11).

Figure 10The input and recovered sub-3 nm particle size distributions simulated using the Monte Carlo method when the particle size distribution increases with decreasing particle diameter.

Download

Figure 11The recovered particle size distributions simulated using the Monte Carlo method when assuming the relative standard deviation of the recorded particle number concentration are 10 %. The reported size bins smaller than 1.3 nm recovered using the kernel function method and the stepwise method are not shown because of the large uncertainties.

Download

4.3 Uncertainties in the scanning mode

The PSM instrumental factors limiting the accuracy of the inversion were also tested. Although using the EM algorithm and the H&A method can reduce the errors in the recovered size distributions compared to the kernel function method and the stepwise method, relatively small measuring uncertainties are still vital to retrieve a particle size distribution with relatively high accuracy. The uncertainty in the scanning mode, for example, is one of the potential sources of the measuring uncertainties. The saturator flow rate of a scanning PSM increases linearly with time in previous studies. However, the relationship between the particle diameters and the saturator flow rates at the kernel function peaks is non-linear (Fig. 2). The detection efficiencies of particles larger than 1.6 nm vary mainly in the flow rate range from 0.05 to 0.3 L min⁻¹, while the corresponding scanning time is only 20 % of the whole scanning cycle. This non-linear relationship may result in non-negligible uncertainties in the recovered particle size distributions (Fig. 12). The EM algorithm recovered the single peak when using the particle concentrations recorded in the stepping mode. However, the recovered particle size distribution using the EM algorithm was not a single smooth peak when using data recorded in the scanning mode (Fig. 12). This difference can be illustrated using the raw data. The curves of the particle number concentration recorded in the stepping mode and the scanning mode are similar to each other and they both appear to be smooth (Fig. 13a). When presenting in the derivate of the particle number concentration with the respect to saturator flow rate, however, the curve corresponding to the stepping mode appeared to be a single peak, while the other curve corresponding to the scanning mode seemed to be composed of multiple single peaks (Fig. 13b). Since none of the four inversion methods tested in this study add smoothing constraints when solving the Fredholm integral equation of the first kind, this roughness in the raw data will lead to split peaks in the recovered particle size distribution unless one reports the size distribution using only a few size bins.

Figure 12The recovered particle size distributions using the particle number concentration recorded in (a) the stepping mode and (b) the scanning mode.

Download

Figure 13(a) The relationship between the recorded particle number concentration and the saturator flow rate in the scanning mode and the stepping mode. (b) The derivative of number concentration with the respect to the saturator flow rate.

Download

4.4 Implications on using the PSM

According to the discussion above, we provide the following suggestions on using a PSM to determine particle size distributions:

• a.

  Particle size range and saturator flow rate range

  Complete efficiency curves are preferable to determine the particle size distribution in a certain size range. For example, to reduce the uncertainties in the recovered size distribution of ∼3 nm particles, the saturator flow rate in this study was extended from the commonly used 0.1 to 0.05 L min⁻¹ where the detection efficiency of 3.11 nm particles was almost zero. The detection efficiency curves of particles larger than the maximum concerned diameter should also be calibrated to reduce the influence of large particles on the recovered particle size distribution and total concentration. The PSM can theoretically estimate particle size distributions larger than 3 nm or smaller than 1.3 nm; however, the uncertainties are usually large due to the low resolution and the incomplete detection efficiency curves. The particles with detection efficiency of constant values in the measuring saturator flow rate range cannot be determined using a PSM and they do not influence the recovered particle size distributions if their concentrations are stable during each scanning cycle.

• b.

  Scanning scheme

  The scanning scheme of the saturator flow rate is suggested to be improved to reduce the measuring uncertainties. The scanning scheme is preferably determined to ensure that the particle diameter corresponding to the saturator flow rate increases linearly with time so that the numbers of the recorded particle number concentration at each saturator flow rate are the same when the recovered particle size increases linearly. A convex function between the saturator flow rate and the scanning time, e.g. an exponentially increasing saturator flow rate, is also better than the linear scanning scheme. This improvement may require both the hardware and the software to be updated.

• c.

  Inversion method

  We suggest using the EM algorithm to address the PSM inversion problem because the particle size distributions recovered using the EM algorithm have the best accuracy and stability among the four tested methods. However, considering the relatively high computational expense of the EM algorithm, the H&A method reporting in merged size bins is recommended to be used for preliminary data analysis and to meet the need of fast inversion, e.g. real-time display on the instrumental screen. The accuracy of the recovered size distribution is also determined by the uncertainties in the recorded number concentration rather than the inversion method alone. The inversion methods suggested in this study do not necessarily ensure an accurate inverted result without properly determined detection efficiencies and an improved scanning scheme.

• d.

  Uncertainties in atmospheric measurement

  One should be always aware of the potential uncertainties in the recovered particle size distribution, especially when conducting atmospheric measurement. The reported sub-3 nm particle concentrations may give false results due to systematic and random errors, especially when using the stepwise method. The number of the reported size bins should also be carefully limited. For example, the EM algorithm can theoretically provide infinite size bins; however, we suggest reducing the reported size bins to avoid false fluctuations.