In the HUB-forward analysis we know or assume an underlying distribution Pu(T) of ice nucleation temperatures for the IN in the sample and generate from it an artificial IN dilution series similar to those obtained in experiments, from which we compute the cumulative freezing spectrum Nm(T) using Vali's equation (Eq. 1a). Using this approach, we investigate the relationship between Nm(T) and Pu(T) (Fig. 1) and the sensitivity of Nm(T) with respect to the number of droplets and dilutions. For generality, we represent Pu(T) as a linear combination of normalized continuous distributions PiT that represent subpopulations of freezing temperatures: 2Pu(T)=c1P1T+c2P2T+…+cpPpT, where p is the total number of subpopulations, P1T,P2T,…,PpT are normalized distribution functions, and c1, c2, …, cp are their weights such that ∑i=1pci=1. These subpopulations could correspond to different chemical, topographical, or structural motifs in the IN samples, although chemically distinct species could also produce overlapped freezing signatures, and a single species could display a broad freezing range. Our formalism does not require a mapping of subpopulations of freezing temperatures to physical IN sites. The units of Pu(T) are the same as for nm(T), i.e., those of the cumulative spectrum divided by a unit of temperature, but are generally omitted in what follows. Throughout this work we assume that PiT can be represented by Gaussian (i.e., normal) distributions: 3Pi(T)=1si2πe-12T-Tmode,isi2, where each subpopulation PiT is further characterized by its most likely temperature of freezing Tmode,i and spread of distribution of freezing temperatures si. We also provide in the HUB code the option for the user to use the log-normal distribution, which has a tail towards higher temperatures, or the left-tailed Gumbel distribution, which has a tail towards lower temperatures. In our model, we assume that the underlying distribution of ice-nucleating temperatures Pu(T) does not change with the concentration of INs. This last condition is violated when INs are involved in chemical, aggregation, or solubility equilibria that alter the proportionality between their concentration and the dilution factor of the sample, resulting in a lack of overlap of the pieces of the cumulative spectra Nm(T) obtained from different dilutions (Bogler and Borduas-Dedekind, 2020).

The number of INs in each droplet is given by the Poisson distribution: 4p(n,λ)=λnn!e-λ, where n is the actual number of INs in each droplet and λ represents the average number of INs among all droplets of the corresponding dilution. Figure 2a shows the probability mass function (PMF) for λ=1, 5, and 10, computed according to Eq. (4) and sampling over N0=104 droplets using the “SciPy Stats” Python framework (Virtanen et al., 2020). As λ increases, the probability that any droplet nucleates homogeneously rapidly approaches zero (inset of Fig. 2a). When there is one IN on average per droplet (λ=1),∼37 % of the droplets do not have any INs; i.e., they are “empty” droplets that would nucleate at the homogeneous nucleation temperature. We note that by performing dilutions until a sizeable fraction of droplets nucleate homogeneously, it is possible to calibrate the absolute concentration of ice nuclei in the original, undiluted sample.

To illustrate how the heterogeneous ice nucleation temperatures recorded in drop-freezing experiments depend on the number of INs in the droplets, we start from two examples with Pu(T) represented by one or two Gaussian subpopulations, shown with dashed black lines in Fig. 2b and c, respectively. We assign a temperature to each IN contained in droplets from a 10-fold dilution series of five solutions with λ=1, 10, 102, 103, and 104 average number of INs per droplet. If the droplet volume is constant, λ is proportional to the concentration of INs in the droplets. We sample N0=104 droplets for each concentration. This N0 is much higher than the ∼100 droplets usually sampled in laboratory experiments; we address the effect of sampling in Sect. 3.1 below.

To sample independent random values for each IN, the number of random variates, which are drawn from Pu(T), is the total number of INs among N0 droplets. Thereby, each droplet has a set of temperatures Tjλ=(T1λ,T2λ,…,Tkλ), where j is the droplet index and k is the IN index. Since we assume that freezing occurs at the characteristic temperature of the IN with the highest freezing temperature, the nucleation temperature for each droplet is defined as the maximum, i.e., the extreme upper value, of several independent freezing temperatures Thet,jλ=max⁡(T1λ,T2λ,…,Tkλ). Figure 2b and c show the normalized distribution of Thet,jλ for different values of λ, namely Pmaxλ(T). Therefore, Pu(T) represents the underlying probability of heterogeneous ice nucleation temperatures independent of the concentration of INs, while Pmaxλ(T) represents the concentration-dependent distribution and has the same units as Pu(T) and the differential spectrum. According to the Fisher–Tippett–Gnedenko theorem, the distribution of extreme upper values of the Gaussian distribution is the right-skewed Gumbel distribution (Castillo et al., 2005; David and Nagaraja, 2004; Gumbel, 2012; De Haan and Ferreira, 2006), which has a fatter tail on the high-temperature side of its maximum. Indeed, the shift in Pmaxλ(T) curves in Fig. 1b and c evinces that as the number of INs in the droplet increases, the probability of sampling the higher temperature tail of Pu(T) increases significantly. This skew is the reason why several dilutions are needed to sample the full population of ice nucleants.

HUB-forward computes the fraction of frozen droplets and cumulative spectra from a proposed underlying distribution of freezing temperatures, using extreme value statistics. The fraction of frozen droplets ficeλT can be calculated as a function of the concentration-dependent distribution: 5ficeλT=∫TmTPmaxλT′dT′×NFλN0, where the integration is from the ice melting temperature Tm to the temperature T,NFλ is the total number of droplets that freeze heterogeneously, and N0 is the total number of droplets. We note that the approach taken in this work differs from that of previous studies that start from a microscopic model for the nucleation sites and nucleation theory to predict the fraction of frozen droplets using Monte Carlo simulations, as well as from previous modeling using the singular approximation, which do not account for the statistics of extreme sampling.

To use the HUB-forward code, the user must define the total number of droplets “ndroplets” that serves as the total number of each concentration and the number of subpopulations “nsubpop”. If “nsubpop” = 1, the user must provide the temperature of maximum likelihood Tmode,1 and the spread s1. If “nsubpop” = 2, the user must provide Tmode,1, s1, Tmode,2, s2 and c2. If “nsubpop” = 3, the user has to provide Tmode,1, s1, Tmode,2, s2, c2, and Tmode,3, s3, c3. To generate the cumulative freezing spectrum NmT, the user needs to define the total number of concentrations “nconc”; the concentration of the parent suspension is defined in “density” and the droplet volume in “volumedrop”. The output is composed of different data plots and files: the normalized Pu(T) and PmaxλT, the artificially generated ficeλ(T), and NmT built from the 10-fold dilution series.

Figure 3a and b show the fraction of ice computed using PmaxλT of Fig. 2b and c, which correspond to Pu(T) with two subpopulations and one subpopulation, respectively. The intermediate plateau in Fig. 3b indicates that no droplets freeze at those temperatures. As discussed above, only 63 % of the droplets freeze heterogeneously for λ=1. We assume droplets of uniform volume Vdrop=0.1 µL obtained through 10-fold dilution of a parent suspension with λ=104 INs per droplet corresponding to a mass m=1 mg of IN in a volume Vwash=1 mL. We use Eq. (5) and the Pu(T) of Fig. 2b and c to generate ficeλ(T) (Fig. 3a and b), sampling either 100 or 104 droplets per dilution. We combine the ficeλ(T) using Eq. (1a) to build the cumulative freezing spectra Nm(T) shown in Fig. 3c and d (sampling 104 droplets per dilution) and Fig. 3e and f (sampling 102 droplets per dilution).

The ability of HUB-forward to generate the cumulative freezing spectrum Nm(T) from the underlying distribution Pu(T) allows for an analysis of the sensitivity of Nm(T) and Pu(T) to the number of droplets and dilutions, as seen in the comparison of NmT generated from the same underlying distributions using 100 and 104 droplets in Fig. 3. In Sect. 3.1 we show that the sampling with 100 droplets for only four dilutions of a system with two subpopulations of INs results in distortions of the distribution of freezing temperatures and the proportions of these populations in the differential spectrum.

The knee point in Nm(T) corresponds to the point of maximum curvature (Satopaa et al., 2011) and has been used to characterize the nucleation temperature of a particular subpopulation (Hartmann et al., 2022). Similar to Hartmann et al. (2022), we have identified in Fig. 3c and d the knee points (dotted magenta line) of the artificially generated Nm(T) by using a Python function named “kneed”. The Python function “kneed” using S=1, curve = “concave”, and direction = “decreasing”. The knee points Tknee are very close to the temperatures of maximum likelihood Tmode (dashed black lines) of the corresponding underlying distribution PuT, because under these conditions the differential freezing spectrum nm(T) is a very good approximant for Pu(T). However, we find that the removal of the more dilute solutions eliminates the plateau in Nm(T) and results in poor estimation of the modes of Pu(T) from the knee of Nm(T).

The HUB-backward code implements a stochastic optimization procedure to extract the differential spectrum nm(T) from a given cumulative spectrum Nm(T) or from an experimental ficeT curve. The latter is useful when data are available for a single concentration. One possibility for obtaining nm(T) from Nm(T) would be to follow the following steps: (i) propose a trial function nmtrial(T), (ii) use HUB-forward to predict the concentration-dependent distributions Pmaxλ,trialT for various IN concentrations, (iii) use these in Eq. (5) to predict the freezing fractions ficeλ,trialT, (iv) compute Nmtrial(T) from the freezing fractions using Eq. (1a), (v) evaluate the difference between that trial and the target (experimental) value using 6δT=log⁡10Nmtrial(T)-log⁡10Nmtarget(T), and then (vi) evolve the parameters that determine nmtrial(T) until the difference δT is minimized. However, the use of HUB-forward in steps (ii) and (iii) to generate and evaluate hundreds of droplets containing up to tens of millions of INs would require significant computations that render this optimization process inefficient.

The HUB-backward optimization procedure, sketched in Fig. 4, uses a shortcut for steps (ii) and (iii) above to directly predict Nmtrial(T) from nmtrial(T) with fast convergence. The shortcut is based on the understanding that, in the asymptotic limit in which the sample is extremely dilute (i.e., λ⟶0), each droplet that nucleates heterogeneously contains a single IN. In such a case, sampling an infinitely large number of droplets with Pmaxλ⟶0T is equivalent to sampling each and every IN, i.e., Pmaxλ⟶0T=PuT. In agreement with this ansatz, Fig. 2b and c show that the underlying distribution Pu(T) (dashed black line) and the concentration-dependent Pmaxλ=1T (blue squares) sampled with 104 droplets per dilution are already very close, i.e., Pu(T)≈Pmaxλ=1T.

With this insight and considering the intrinsic cumulative spectrum, IuT=∫TmTPuT′dT′×(1-e-1), we define the cumulative integral of the differential spectrum as 7IutrialT=∫TmTnmtrialT′dT′×β, where the integration is from the ice melting temperature Tm to the temperature T and β is an adjustable scaling factor to be obtained from the optimization. Likewise, a similar estimate can be made for a single fraction of ice curve ficetrialT=IutrialT using Eq. (7) and the mean squared error can be directly evaluated (Fig. 4). When the target is a cumulative freezing spectrum, HUB-forward uses nmtrialT to predict a trial cumulative freezing spectrum (Fig. 4), 8NmtrialT=-ln⁡[1-IuT]×1X, where 1/X corresponds to the maximum of the cumulative spectrum in the target distribution, 1/X=max⁡NmtargetT. With Eq. (8) we obtain an NmtrialT that we compare with the target using Eq. (6) (Fig. 4). To do the comparison, HUB-backward uses a spline fit to interpolate the experimental Nmtarget(T), in order to have equally spaced temperature points to compare with the estimates in Nmtrial(T). We use the “interp1d” algorithm, which is available in the Python SciPy library (Virtanen et al., 2020), with a linear interpolation to construct new equally spaced data points within the range of the lowest and highest temperature values in the freezing spectrum. The cost function for the optimization is the mean squared error (MSE), computed from the difference δT in Eq. (6): 9MSE=1t∑δ2, where t represents the total number of equally spaced points in δT.

We use a stochastic global optimization technique based on a simulated annealing algorithm to find the set of parameters of nmtrial (Eqs. 2 and 3) and β (Eq. 7) that globally minimize the MSE. We use the simulated annealing (SA) algorithm “dual annealing” that is part of the SciPy minimize library (Virtanen et al., 2020) with its default arguments predefined, except for the parameters “maxfun” that sets the maximum number of evaluations of the objective function (we select “maxfun” = 1 000 000 in the examples below) and the seed for the generation of random numbers (a new random integer is automatically generated every time the HUB-backward code is run). We show below that the optimized differential spectra, nmoptimizedT, are quite insensitive to the value of the seed.

The output of HUB-backward is an optimized differential spectrum nmoptimized(T) or an optimized fraction of ice ficeoptimized(T). To quantify how much this optimized prediction deviates from the known underlying distribution in the examples of Fig. 5, where Pu(T) is known, we define the mean relative error (MRE) for the set of parameters: 10MRE=13p∑i=1pTmode,ioptimized-Tmode,itargetTmode,itarget+sioptimized-sitargetsitarget+cioptimized-citargetcitarget, where p is the number of subpopulations.

We now turn our focus to how to select the input parameters required by HUB-backward to start the search for the underlying distribution, using the experimental Nmtarget(T) or ficetarget(T) as a guide. The code requires the user to define the number of distinct Gaussian subpopulationsPi(T) that comprise the underlying distribution (Eq. 2) and to provide upper and lower bounds for the weights ci, their modes Tmode,i, and spreads si of each of these populations. In general, we find that defining the minimum and maximum values for the weights to cimax=1 and cimin=0 (see constraint in Eq. 2), for the modes Tmode,imax and Tmode,imin to between the homogeneous nucleation temperature (about -30 ∘C) and the melting temperature (0 ∘C), and for the spreads to simax=10 ∘C and simin=0.1 ∘C works well. However, these bounds can be tuned in order to better fit the data (as we find for pollen in Sect. 3.2 below). If the existing experimental Nmtarget(T) data are very noisy, they can be interpolated in HUB-backward using the method “interp1d” with “npoints” = 100 and then smoothed with a Savitzky–Golay filter by changing the parameters “window_length”, which is the length of the filter window, and “polyorder”, which is the order of the polynomial used to fit the samples (“filter” in Fig. 4). The default values are 3 and 1, respectively. HUB-backward generates a plot that compares the original and the interpolated target data.

To identify the minimum number of subpopulations needed to represent a given freezing spectrum, we consider that every time a population is accumulated in Nm(T) or fice(T), these functions display a sharp increase. We note that assuming a large number of subpopulations may challenge the interpretability of the optimized differential spectrum nmoptimized(T).

We apply the HUB-backward procedure to the NmT obtained in Fig. 3c and d by sampling four 10-fold dilutions with 100 droplets, i.e., only a total of 500 droplets. Figure 5 shows the comparison between the predicted (solid magenta lines) and the target (dashed black lines) NmT (panels a and b) and nmT (panels c and d). Table 1 shows the predicted parameters and the precision of the optimization procedure to recover the known underlying distribution Pu(T). The MRE between the underlying distribution Pu(T) and the optimized differential spectrum nmoptimized(T) is 2 % for the system with one subpopulation and 13 % for the one with two despite the low number of droplets used to sample the cumulative freezing spectra in the computer-generated freezing experiments.

We conclude that the HUB-backward code gives a good estimate of the mode, spread, and weights of the populations of INs in a sample and it can be applied in a situation where Pu(T) is unknown. In Sect. 3.1 we discuss how the accuracy is of the underlying distribution recovered with HUB-backward impacted by various schemes of the sampling of the number of droplets and dilutions to construct NmT. In Sect. 3.2, we apply the HUB-backward procedure to obtain nmoptimized(T) from actual NmT of experiments with various soluble biological INs. In Sect. 3.3, we apply the HUB-backward procedure to obtain nmoptimized(T) from ficeT of experiments of insoluble crystal INs.

Figure 3d–f show Nm(T) generated with HUB-forward using five dilutions from λ=104 to 1 of a solution with Pu(T) containing two populations in a ratio of 9 to 1. The Nm(T) are different when the number of droplets per dilution is 100 (Fig. 3f) or 104 (Fig. 3d). As shown in the previous section, the freezing spectrum obtained with 100 droplets and five dilutions has enough sampling to recover this Pu(T) with good accuracy (Fig. 5c and d). We test different number of droplets and concentrations, defined by the average number of INs per droplet λ, to test the sensitivity of nm(T) to the number of droplets and dilutions when the underlying distribution Pu(T) is known. We use HUB-forward to build Nm(T) based on a combination of different numbers of droplets and concentrations, similar to the case shown in Fig. 3f. Then, we use HUB-backward to obtain nmoptimized(T), compare it to PuT, and test the accuracy of each prediction through its mean relative error (MRE) defined in Eq. (10).

The left panels of Fig. 6 show Nm(T) generated with HUB-forward based on a combination of different concentrations using 100 droplets for each dilution. The magenta lines are based on the data provided by the HUB-backward code. The right panels of Fig. 6 compare nmoptimized(T) in magenta and the known underlying distribution PuT in black. In this example, nm(T) is very close to PuT if both subpopulations are sampled enough. However, if the most dilute solution with λ=1 is not included in Nm(T) (second panel), the estimate of the underlying distribution is very poor. Thus, to improve the sampling of the lower tail of PuT, we recommend ending the dilution series always in the immediacy of λ=1, which can be gleaned from the temperature range for which NmT becomes flat and a sizeable fraction of droplets of the more diluted sample nucleates homogeneously (inset of Fig. 2a). The one-to-one correspondence between the fraction of droplets nucleated homogeneously and the average number of particles in the droplet in the highly diluted limit (inset of Fig. 2a) demonstrates that reaching this limit allows for an absolute calibration of the number of INs in the initial sample. Moreover, sampling to concentrations down to about one nucleant per droplet is essential to recover a proper weight of the poorly nucleating IN populations.

The relative weights of class A and C populations in Pseudomonas syringae is approximately 1 to 1000 (Sect. 3.2), while the ratio is 9 to 1 in the two-population system example of Fig. 6. To understand the impact of highly imbalanced populations on the sampling of the cumulative spectrum and recovery of the underlying distribution, we show in Fig. 7 the analysis of an example where the subpopulation of highly efficient INs is 3 orders of magnitude less likely to occur than the subpopulation at lower temperatures, mimicking the one of P. syringae. Our analysis confirms that it is important to end the dilution series in the immediacy of λ=1 to fully represent the contribution of the poorer INs (Fig. 7b–f). Furthermore, we find that it is important to sample a concentration high enough to account for the rare INs that nucleate at the warmest temperatures (Fig. 7d–h).

If only 25 droplets per dilution, instead of 100, are used to construct the cumulative spectrum, the impact of insufficient sampling at the higher concentrations is more pronounced: compare Fig. 8c and Fig. 7d obtained with the same underlying distribution PuT with 1000 to 1 subpopulation ratios and number of dilutions.

We conclude that an increase in the accuracy in the account of the subpopulations requires a higher number of dilutions and the checking of the predictions with the addition of each successive concentration to ensure convergence of nmoptimized(T). Measuring fewer droplets or fewer dilutions leads to poor statistics and results in incompleteness or the misrepresentation of the underlying distribution in samples with multiple subpopulations. In principle, increasing the number of droplets of the most concentrated solutions or adding more 10-fold concentrated ones until there are no changes in the cumulative spectrum is recommended to ensure complete sampling. When that limiting scenario is not attainable, the use of HUB-forward to produce synthetic data from a proposed underlying distribution, followed by the recovery of the differential spectrum from these data sets, allows for an estimation of the errors that may be incurred for putative, proposed underlying distributions with the sampling scheme available in the laboratory.

In this section we use the HUB-backward code to obtain the differential freezing spectrum nm(T) from the cumulative freezing spectra Nm(T) of the fungi Fusarium acuminatum strain 3–68 (Kunert et al., 2019), the bacterium P. syringae (Schwidetzky et al., 2021), and birch pollen (Dreischmeier, 2019). We select these systems because they are important biological INs and show increasing complexity in terms of the apparent number of underlying distributions that define their freezing spectra.

The experimental Nm(T) obtained for F. acuminatum (black squares in Fig. 9a) was obtained by sampling six 10-fold dilutions, each with 96 droplets (Kunert et al., 2019). Figure 9a shows the cumulative spectra optimized assuming one (green curve) and two (cyan curve) subpopulations; Fig. 9b shows the corresponding optimized differential freezing spectra. The nmoptimized(T) with a single subpopulation that peaks at -5.9 ∘C is unable to represent the cumulative density of the most potent nuclei and misses the inflection at around -5.9 ∘C in the experimental data, resulting in a mean squared error MSE = 0.05. The nmoptimized(T) with two subpopulations has a lower MSE = 0.003 and a better fit that suggests a population that peaks at -7.3 ∘C and another at -5.5 ∘C, in comparable amounts (Table 2). Most notably, the two subpopulations do not overlap in the differential freezing spectrum, supporting that they may indeed correspond to different physical entities. The improvement in the fit becomes apparent in the inset of Fig. 9a, which shows Nm(T) on a linear scale. The significant slope of Nm(T) even at the lowest temperatures indicates that the sampling of more diluted solutions is needed to capture the contribution of the less active INs. An attempt to represent F. acuminatum nucleation data with three different subpopulations resulted in two of them being almost identical. We conclude that adding a third subpopulation is unnecessary to reproduce the experimental cumulative freezing spectrum of F. acuminatum. We refer the reader to Schwidetzky et al. (2023) for an interpretation of the size of the ice-nucleating surface of F. acuminatum based on its differential spectrum and nucleation theory.

Next, we apply the HUB-backward code to analyze the experimental freezing spectrum of Snomax^®, i.e., inactivated P. syringae. The cumulative spectrum suggests the presence of two distinct subpopulations, usually called class A (at warmer temperatures) and class C (at colder ones). We first assume the differential freezing spectrum nm(T) of P. syringae is a combination of two Gaussian populations. The parameters of the optimized differential spectrum with two subpopulations are listed in Table 2, and the curve is shown in Fig. 9d with a cyan line. We use a logarithmic scale to represent this nmoptimized(T) because the population corresponding to class A accounts for less than 0.1 % of the total (Table 2). While the fit with two subpopulations results in a good overall account of the target data, we note that there is some difference in the region between classes A and C (Fig. 9c). The fitting for P. syringae achieves an excellent agreement between optimized and target cumulative spectra (Fig. 9c), through the prediction of an additional peak located between classes A and C (the elusive class B), with a population comparable to class A (Table 2 and red curve in Fig. 9d). However, more measurements and analyses are needed to establish whether this “class B” peak at -5.2 ∘C is reproducible and truly distinct from the one of class A at -3.7 ∘C to warrant a physical interpretation. Overall, both the analyses with two and three subpopulations agree with previous ones (Govindarajan and Lindow, 1988; Warren, 1987) that concluded that over 99 % of the INs active in P. syringae bacteria in Snomax^® belongs to class C. The analysis presented here for fungal and bacterial INs illustrates how HUB-backward can be used to reveal and characterize the underlying number of IN subpopulations of complex biological samples.

To further test the methodology, we model the cumulative freezing spectrum of birch pollen. Given that the original Nm(T) data for pollen in Fig. 3.1 of Dreischmeier (2019) consist of multiple independent curves, we took one of the many presented in this graph as the target Nmtarget(T) (black curve in Fig. 9e) and present some of the additional data – not used in the optimization – with gray circles in Fig. 9e. Section S4 in the Supplement shows that the differential spectrum optimized from the whole data set and its sparse sampling are almost identical because HUB-forward interpolates and smooths the input data to produce an equispaced data set. The Nmtarget(T) seems to contain three quite separated subpopulations, which is confirmed by the accuracy of the optimized cumulative spectrum in Fig. 9e. The parameters of the optimized differential freezing spectrum nmoptimized(T) and the MSE are shown in Table 2. Our analysis indicates that the two subpopulations that nucleate ice above -16 ∘C constitute less than 0.01 % of the active nucleating sites in pollen (Fig. 9e), consistent with drop-freezing assays that only measured solutions with low concentrations of birch pollen and did not observe freezing at higher temperatures (Augustin et al., 2013; Pummer et al., 2012; Felgitsch et al., 2018), while the more extensive data of Dreischmeier (2019) reveal two more active subpopulations of INs.

To further illustrate the use of HUB-backward, Fig. 10 shows the effect of pH in the modes, spread, and weights of the subpopulations that contribute to the nucleation spectrum of P. syringae (Snomax^®), using data from Lukas et al. (2020). Freezing in the temperature range of class A drops about 3 orders of magnitude when the pH is lowered from 6.2 to 4.4 (Fig. 10b). However, we note that the cumulative number of INs is preserved in the experimental cumulative freezing spectrum (Lukas et al., 2020), indicating that the change in pH did not impact the number of nucleants. Figure 10c and d demonstrate that the distributions associated with both subpopulations shift to lower temperatures when the pH decreases, and the range of freezing temperatures in class A becomes broader. An attempt to fit the cumulative spectra of Snomax at different pH values with the same subpopulations, allowing only for adjustment of their weights, resulted in a poor fit to the experimental Nm(T), supporting the conclusions of Lukas et al. (2020) of a central role of electrostatic interactions in the assembly of the bacterial ice-nucleating proteins and their ability to bind to ice. This analysis exemplifies how HUB-backward can be applied to quantify the dependence of IN on environmental variables.

Section 3.1 and 3.2 discuss how to obtain the differential spectrum from a target cumulative one. However, there are many cases where the results are presented as a fraction of frozen droplets as a function of temperature, fice(T). In these cases, the HUB-backward code can be used to obtain the optimized differential freezing spectrum nmoptimized(T) directly from ficetarget(T). Section S5 illustrates this approach for the analysis of droplet freezing data for a sample of lignin (Bogler and Borduas-Dedekind, 2020) in which the INs participate in aggregation equilibria. Here, we exemplify the optimization of the differential spectrum of cholesterol from experimental freezing data obtained at two cooling rates with droplets sampled from a single dilution.

In the analysis of drop-freezing experiments, it is assumed that each IN has a singular freezing temperature, independent of the cooling rate. However, ice nucleation is a stochastic process, and the underlying distribution of freezing temperatures PuT strictly depends on both temperature and cooling rate, as slower rates give more time for the system to cross the nucleation barrier at warmer temperatures.

The triangles and squares in Fig. 11a display the experimental fice(T) obtained by sampling the freezing of hundreds of 120 µL droplets pipetted from a suspension of cholesterol monohydrate crystals in contact with Teflon cooled at 0.18 K min-1 (triangles) and 0.06 K min-1 (squares) (Zhang and Maeda, 2022). Our analysis of the freezing data of cholesterol monohydrate shows that even a 3-fold change in the cooling rate can have a significant effect on the differential spectrum (Fig. 11b).

As expected, the modes of the three populations move towards warmer temperatures upon decreasing the cooling rate. We note, however, that the shift in the peaks is not uniform; the middle one seems to be more sensitive to the cooling rate. Different sensitivity of the freezing rate of subpopulations has also been reported in simulations of nucleation data of minerals using the stochastic and modified singular frameworks (Herbert et al., 2014; Murray et al., 2011). The modified singular model proposes an empirical correction to the relation between fice(T) and Nm(T) to account for the effect of the cooling rate on the shift in these quantities (Vali, 1994). That analysis could be extended to the analysis of the subpopulations of INs obtained with HUB-backward. Moreover, it would be interesting in future studies to use the rate dependence of the mode of the subpopulations to extract the steepness of the nucleation barrier with temperature using nucleation theory (Budke and Koop, 2015) and to investigate the relationship between the cooling rate dependence of the differential spectrum obtained in the singular approximation with the interpretation of the same data modeled with the stochastic framework, such as in Wright et al. (2013) and Herbert et al. (2014).