A Database of Experimentally Derived and Estimated Octanol–Air Partition Ratios (K_OA) Open Access

Understanding the affinity of a chemical for liquid octanol, liquid water, and the gas phase is often the first step to understanding its potential environmental and biological fate and behavior. The physical–chemical properties to quantify those affinities include equilibrium partition ratios, saturation solubilities, and vapor pressure, which are related to one another through a series of thermodynamic triangles (Fig. 1). Chemical equilibrium partition ratios, hereafter simply referred to as partition ratios, are a concept fundamental to physical chemistry, with many applications in the fields of environmental, medicinal, and pharmaceutical sciences. While in the literature the thermodynamic property is more commonly referred to as a partition coefficient, we follow IUPAC nomenclature guidelines and describe the distribution of a chemical between two phases at equilibrium as a partition ratio.

The unitless octanol–air partition ratio (K_OA) describes the distribution of a chemical between octan-1-ol (CAS No. 111-87-5) and the gas phase at equilibrium,

where C_O and C_A are the concentrations of a compound in n-octanol and the gas phase in mol m⁻³, respectively. K_OA has many possible applications, most notably in linear free–energy relationships for predicting the equilibrium distribution of compounds between the gas phase and atmospheric particles (Finizio et al., 1997), blood (Batterman et al., 2002), soil (Hippelein and McLachlan, 2000; 1998), foliage (Müller et al., 1994; Paterson et al., 1990), and some of the polymers used in passive air samplers (Ockenden et al., 1998; Shoeib and Harner, 2002a).

Many comprehensive reviews (Mackay et al., 2015) and databases of octanol–water partition ratios (K_OW) (Leo et al., 1971), Henry’s law constants (k_H) (Mackay and Shiu, 1981; Sander 2015), and other physical–chemical properties [e.g., Mackay et al. (2006); Rumble et al. (2019); US EPA (2012)] exist in the literature. While Jin et al. (2017) compiled K_OA data for the development of an estimation model, there has been no comprehensive collection or review of K_OA data to date. Our aim is to assemble comprehensively and critically all previously published experimental and estimated K_OA data. This work further includes an overview of the different techniques that have been used to obtain K_OA values. The assembled database should be an easy-to-look-up repository of existing K_OA data but also be suitable for evaluating existing K_OA prediction techniques and the development of new ones.

In this section, we briefly review the various ways in which K_OA has been reported in the literature. In most cases, the values included in the database were reported as K_OA or log₁₀ K_OA values; however, 1409 values were derived from reported Ostwald coefficients in octanol (L_oct), Henry’s law constants in octanol (⁠$kHoct$⁠, Pa m³ mol⁻¹), the Gibbs energies of dissolution into octanol from the gas phase (∆G°, J mol⁻¹), or activity coefficients of a chemical in octanol at infinite dilution (⁠$γo∞$⁠). While various papers report values using different units for pressure, temperature, and volume, we have reported all equations and variables using SI units (e.g., Pa for pressure, K for temperature, and m³ for volume) unless otherwise stated.

The earliest measurement of the solvation of a compound in octan-1-ol from the gas phase that we found was published in 1960 and was reported as an Ostwald coefficient in octanol (L_oct) by Boyer and Bircher (1960). The Ostwald coefficient has been used for over a century to describe the solubility of gases in liquids (Ostwald, 1891). Since Ostwald initially coined the term, the following definitions for the Ostwald coefficient at equilibrium have been used (Battino, 1984):

• L_V⁰ is the volume of gas (V_G) dissolved in a volume of pure liquid (V_L⁰),

  $LV0=VG/VL0,$

  (2)
• L_V is the volume of gas (V_G) dissolved in a volume of solution (V_L),

  $LV=VG/VL,$

  (3)
• L_C represents the concentration of a gas in the liquid phase (C_L) divided by its concentration in the vapor phase (C_V),

  $LC=CL/CV,$

  (4)

  and
• L_C^∞ is L_C at the infinite-dilution concentration of the gas in the liquid,

  $LC∞=limCL→0CL/CV.$

  (5)

Battino (1984) reviews more comprehensively the differences between these definitions and judges concentration-based definitions for equilibrium ratios to be the most thermodynamically reliable and useful method for reporting Ostwald coefficients (Battino, 1984; Wilhelm and Battino, 1985). The use of L_oct to describe octanol–air partitioning is not altogether common and is to our knowledge limited to Boyer and Bircher (1960), Wilcock et al. (1978), Pollack et al. (1984), and Bo et al. (1993). Unless the reference states otherwise, we assume all published L_oct values to be concentration ratios, equivalent to the K_OA value (Abraham et al., 2001).

The Gibbs energy describing the energy required to transfer a solute between two phases can also be expressed in two ways (Schwarzenbach et al., 2005). If the Gibbs energy for octanol–air transfer is reported on a concentration basis (ΔG°), we can directly solve for K_OA using

where R is the ideal gas constant (8.314 J K⁻¹ mol⁻¹) and T is the absolute temperature (in K). If the Gibbs energy is reported using partial pressure and mole fraction (ΔG^*), a conversion to ΔG° is first required (Berti et al., 1986; Cabani et al., 1991),

where v_oct is the molar volume of octanol (0.000 158 m³ mol⁻¹ at 25 °C) (Rumble et al., 2019; Yaws, 2012).

Air–water equilibrium is often expressed with the Henry’s law constant (k_H) (Fig. 1), typically with units of Pa m³ mol⁻¹. Likewise, partitioning between octanol and air can be described as the k_H in octanol (⁠$kHoct$⁠) with units of Pa m³ mol⁻¹. Leng et al. (2015) and Roberts (2005) described octanol–air partitioning using the Henry’s law solubility constant, the reciprocal of $kHoct$ (k^′_H^oct, mol m⁻³ Pa⁻¹). K_OA is obtained using

K_OA can be related to a chemical’s solubility S_O (in units of mol m⁻³ octanol) or activity coefficient at infinite dilution $γo∞$ (hereafter referred to as the activity coefficient) in octanol,

where P_L is the liquid-phase vapor pressure (in Pa). A K_OA value can therefore be calculated from a reported activity coefficient $γo∞$ using the thermodynamic triangle of Eq. (9) if the vapor pressure of the liquid solute P_L is available. For the purposes of this database, we include K_OA values calculated using Eq. (9) if $γo∞$ and P_L were measured for the same system (Hussam and Carr, 1985) or if the P_L was used to derive $γo∞$ (Bhatia and Sandler, 1995; Dallas and Carr, 1992; Fukuchi et al., 2001; 1999; Tse and Sandler 1994). Chemicals for which the reported solubilities or activity coefficients and the vapor pressures derive from different studies are not currently included in the database of measured K_OA values.

Using another thermodynamic triangle, K_OA can be related to the ratio of the octanol–water (K_OW in units of m³ water m⁻³ octanol) and air–water partition ratios (K_AW in units of m³ water m⁻³ air) or the Henry’s law constant in water (k_H in units of Pa m³ water mol⁻¹),

Because during a K_OW determination, water–saturated octanol is being equilibrated with octanol–saturated water, the thermodynamic triangle of Eq. (10) yields the partitioning ratio between water–saturated octanol (referred to occasionally as “wet” octanol) and the gas phase, which we call K^′_OA. The presence of water in octanol may increase the octanol solubility of more hydrophilic chemicals and reduce the octanol solubility of more hydrophobic chemicals (Beyer et al., 2002).

In most instances, the K_OA reported in the literature refers to the ratio of concentrations of a chemical in pure octanol and the gas phase at equilibrium. However, this is not always the case [e.g., Xu and Kropscott (2014; 2012)]. Therefore, we note within the database whether K_OA or K^′_OA is reported.

The thermodynamic constraints imposed on the partitioning properties by the four thermodynamic triangles displayed in Fig. 1 have been used to adjust properties that are subject to measurement errors to yield a set of properties, called final adjusted values (FAVs), that is internally consistent and, by inference, subject to reduced error (Beyer et al., 2002). Those efforts also take into account the potential discrepancy between K_OA and K^′_OA. Whereas FAVs for the K_OA of hexachlorocyclohexanes (Xiao et al., 2004), other organochlorine pesticides (Shen and Wania, 2005), polycyclic aromatic hydrocarbons (Ma et al., 2010), polybrominated diphenyl ethers (Wania and Dugani, 2003), polychlorinated biphenyls (Li et al., 2003), polychlorinated dibenzo-p-dioxins and -furans (Åberg et al., 2008), and volatile methylsiloxanes (Xu et al., 2014) have been reported in the literature, this database does not include them.

K_OA is often highly temperature dependent. At higher temperatures, the K_OA of a chemical will be lower, as it becomes more volatile; at low temperatures, K_OA is higher. As an example, Fig. 2 plots log₁₀ K_OA of DDT (CAS No. 50-29-3) against reciprocal absolute temperature T, where K_OA spans multiple orders of magnitude over a 50 °C temperature range. The slope m of the linear regression between log₁₀ K_OA and 1/T is related to the molar internal energy of octanol-to-air phase transfer (∆U°_OA, J mol⁻¹),

If ∆U°_OA is assumed to be constant over a small range of temperatures, the van’t Hoff equation can be used to calculate the K_OA at different temperatures,

Here, ∆U°_OA expresses the temperature dependence of a partition ratio with the gas phase if the abundance of the chemical in the gas phase is expressed using a volumetric concentration (Goss and Eisenreich, 1996; Atkinson and Curthoys, 1978). The molar enthalpy of solution in octanol from the gas phase (∆H°_OA, J mol⁻¹) is used when the chemical’s abundance in air is expressed as partial pressure. ∆U°_OA is related to ∆H°_OA as follows:

While K_OA is almost invariably reported on a volume basis, we found that in some instances ∆U°_OA has been mistakenly referred to as ∆H°_OA. We note the difference between the two variables because prediction techniques for ΔH°_OA (Mintz et al., 2008; 2007) and direct measurements of ∆H°_OA using calorimetric techniques (Fuchs and Stephenson, 1985; Stephenson and Fuchs, 1985a; 1985b; 1985c; 1985d; 1985e) exist in the literature.

The ∆U°_OA must be negative because the slope m in Eq. (11) has a positive value (as log₁₀ K_OA decreases with increasing temperature). Many papers report a positive ΔU°_OA value, which we believe to be ΔU°_AO values.

The different experimental techniques used to measure log₁₀ K_OA can be grouped into three broad categories: dynamic, static, and indirect. Many of the reported values are direct measurements made using the dynamic generator column technique or indirect measurements using gas chromatography retention time (GC-RT) methods. Dynamic techniques typically involve streaming air through or over a stationary octanol phase. In static measurement techniques, the octanol phase and air phase are in direct contact with each other in a closed vessel; however, neither phase is moving. Indirect techniques require a reference compound with a well-established measured K_OA value, and the elution time of the analyte relative to that of the reference compound is used to determine K_OA. In this section, we discuss each of these measurement techniques in greater detail. Table 1 summarizes the different techniques used to measure K_OA. Most techniques have a specific applicability range for K_OA. We also list the temperature range for these different measurements.

In static techniques, either the gas phase, the octanol phase, or both are directly sampled and analyzed for the solutes once they have reached equilibrium within a closed system. This includes a variety of headspace techniques [e.g., Dallas (1995); Hussam and Carr (1985); Park et al. (1987); Treves et al. (2001); Xu and Kropscott (2014; 2013; 2012); and Lei et al. (2019)], a vacuum distillation method (Hiatt 1998; 1997), and a method based on measuring the kinetics of approaching an equilibrium distribution (Ha and Kwon 2010; Lee and Kwon 2016).

In the basic headspace technique, the solute is equilibrated between octanol and headspace in a closed container, whose temperature is controlled, for example, with a water bath. The concentration in the headspace is then quantified using gas chromatography and an external calibration. The concentration in octanol is determined by dissolving a known quantity of solute into a known volume of octanol, and the K_OA is then determined using Eq. (1). Headspace techniques can measure multiple solutes at the same time, at different temperatures, and at low solute concentrations.

Rohrschneider (1973) was one of the first to use headspace analysis to measure solvent–air interactions in many different solvents, including octanol. A small volume of solute was added to 2 ml of solvent and allowed to equilibrate for two to fifteen hours in a temperature bath. The headspace of the vial was sampled and calibrated against the response for the solute in a solvent for which K_iA is known (where i is a solvent).

The group of Carr et al. (Hussam and Carr, 1985; Park et al., 1987; Dallas, 1995; and Castells et al., 1999) refined the headspace technique for measuring solute partitioning between solvents and the gas phase. This technique has also been used by Cheong (1989), Abraham et al. (2001), and Dallas and Carr (1992). Typically, $γo∞$ and P_L are reported, allowing for K_OA to be derived using Eq. (9), or K_OA was reported directly. The data by Castells et al. (1999) are excluded from the database as no P_L values were reported.

Instead of a headspace vial, Xu and Kropscott (2013) used a 100 ml Hamilton syringe to equilibrate a solute between octanol and air. For analysis, air and octanol samples are taken through the same sampling port, with the former being collected onto a cold trap. A more complex apparatus involving two syringes connected by a small valve was used by Xu and Kropscott (2012) to simultaneously measure the partitioning equilibria between two solvents and the headspace. Using octanol saturated with water and water saturated with octanol as the two solvents, Xu and Kropscott (2012) measured the K^′_OA with this system. While this technique can determine multiple phase equilibria of relatively volatile chemicals at the same time, it is extremely challenging to implement because all three phases need to be sampled quickly to avoid disturbing the equilibrium of the system.

The variable phase ratio headspace technique introduced by Ettre et al. (1993), and first applied to the measurement of K_OA by Lei et al. (2019), improves on the basic headspace technique by doing away with the need to quantify the solute concentration in the headspace. Variable volumes of the same octanol solution are placed into sealed vials and allowed to equilibrate. The reciprocal signal strength obtained from headspace analysis is regressed against the phase ratio, which is the volume of air to the volume of octanol solution present in each vial (Lei et al., 2019). The K_OA is then determined as the intercept divided by the slope of the linear regression (Lei et al., 2019), i.e., no calibration or quantification is required.

Whereas headspace techniques work well for volatile compounds, they are unsuitable for chemicals with log₁₀ K_OA greater than about 4 (Lei et al., 2019). One challenge of applying headspace techniques to less volatile solutes is that the concentrations in the headspace are often too small for reliable quantification. Treves et al. (2001) used solid-phase microextraction (SPME) fibers to collect the solute from the headspace and thus increase the amount delivered onto the GC column for analysis. A quantification of the headspace concentration, however, would require knowledge of a solute’s gas–fiber partition ratio (K_FG) and a fiber-specific constant (k_F). Treves et al. (2001) eliminated the need to empirically determine K_FG and k_F of a chemical by using a reference compound with a known k_H. The response of each sample is plotted against the solution concentration, where the slope is equal to K_FG·k_F over $kHoct$·RT.

Seeking to measure the partitioning of anesthetic gases between air and blood, Strum and Eger (1987) developed a headspace technique that can work with small amounts of solvent, which is particularly advantageous when working with human samples (e.g., blood). This technique was widely used in the field of anesthesiology for a range of solvents. The technique as described by Taheri et al. (1991), and variations thereof, have been employed by Eger and colleagues to measure K_OA at 37 °C (Eger et al., 1997; 2001; Fang et al., 1997a; 1997b; 1996; Ionescu et al., 1994; and Taheri et al., 1993). A volume of the gaseous analyte is dissolved into octanol and the concentration of the solute in the headspace is determined using gas chromatography. A small aliquot of the octanol solution is then added to a larger evacuated flask. The pressure in the flask is slowly released, and a syringe is used to pump additional air to the system and mix the gaseous phase. The air in the syringe is then analyzed to determine the concentration of the solute in the gaseous phase. K_OA in this method is derived as a function of the volume of the flask, the volume of the aliquot of octanol solution in the flask, and the initial and final concentrations of the solute in the gas phase sampled above the octanol solution. This is a highly complex methodology and is therefore more likely to be prone to error. It is also limited to gaseous solutes, which must be available in a relatively pure form. These gaseous solutes will have low log₁₀ K_OA values.

Hiatt reported K_OA values while working to improve upon earlier designs of a vacuum distillation with gas chromatography and mass spectrometry (VD/GC/MS) technique for quantifying volatile organic compounds (VOCs) in complex environmental matrices, such as fish tissue and vegetation (Hiatt, 1998; 1997). A sample and a spike containing the analytes of interest are placed in the sample chamber and allowed to equilibrate for three hours (Hiatt 1995). The sample chamber is then evacuated using a vacuum pump for five minutes, and the evacuated air passes first through a condenser column, to collect water vapor, and then a cryo-loop, submerged in liquid nitrogen, to collect the distillate (Hiatt, 1995). A carrier gas is then used to push the distillate through to a GC/MS for analysis (Hiatt, 1995). K_OA is then calculated based on the analyte recovery from the organic phase and a calculated K_OA of surrogate analytes (Hiatt, 1997). A major flaw of this measurement technique is the use of calculated K^′_OA values for the surrogate analytes. The method also assumes that fish tissue and leaves are representative of pure octanol—however, we note that the values reported in these works are not explicitly indicated to be K_OA measurements. While the reverse can be used as an estimation technique, this assumption is not ideal for deriving physical–chemical properties of chemicals. Some of the reported K_OA values have a large degree of error (Hiatt, 1997).

Two general techniques were found to measure the L_oct of gaseous compounds. The first is used specifically for measuring the solubility of xenon (Xe, CAS No. 7440-63-3). Here, Pollack et al. (1984) used a NaI(Tl) crystal paired with a photomultiplier, which is directed at a fixed amount of gaseous Xe held within a sealed chamber (Pollack and Himm, 1982). The chamber is connected to a flask containing a known amount of solvent, in this case octanol, with some headspace (Pollack and Himm, 1982). The Xe is allowed to reach equilibrium with the solvent and excess gas. K_OA can be determined based on the volume of the gaseous phases in the two chambers and the volume of the solvent and by quantifying the amount of Xe present before and after equilibrium is reached (Pollack and Himm, 1982).

The second gas solubility technique often involves the use of specific equipment, such as the Van Slyke–Neill blood gas apparatus (Boyer and Bircher, 1960), modified Morrison–Billett apparatus (Wilcock et al., 1978), or the Ben–Naim/Baer-type apparatus (Bo et al., 1993). These techniques are scarcely described in the original literature; however, Battino and Clever (1966) described the technique using the Morrison–Billet apparatus and the Ben–Naim/Baer-type apparatus in an early review. An excess amount of gas is dissolved into a solvent and then the solution is degassed into an apparatus. The solvent is then saturated with the gas analyte at a constant temperature (Battino and Clever, 1966). Knowledge of the volume of the solvent in which the gas was dissolved and the pressure and volume of gas dissolved yields the gas solubility, and this combined with the partial pressure of the system can provide the L_oct.

In the technique by Ha and Kwon (2010), a tiny droplet of octanol is suspended above an octanol solution within a sealed vial. The kinetics of uptake in the droplet of the solutes of interest and of a reference chemical with a well-established log₁₀ K_OA is recorded by measuring the concentrations in the droplet after variable periods of time. The K_OA can then be derived from the kinetics of uptake if the thickness of the air boundary layer, the molecular diffusivity of the chemical in air, and the surface area and volume of the octanol droplet are known. The reference compound serves to calibrate the thickness of the air diffusive boundary layer. The K_OA of the analytes of interest must be sufficiently high so that the mass transfer resistance of the chemical in the octanol is negligible relative to that in air (Ha and Kwon 2010). The length of the experiment depends on the anticipated K_OA value, as it will take longer for a change in the chemical concentration in the octanol droplet to be quantifiable for chemicals with high log₁₀ K_OA values (Ha and Kwon, 2010). Although their measurements were conducted at 25 °C, Ha and Kwon suggested that this method can be used to obtain K_OA at different temperatures, as long as the octanol drop does not evaporate (Ha and Kwon, 2010). This measurement technique is applicable to chemicals with a log₁₀ K_OA between 5 and 9 (Ha and Kwon, 2010), i.e., it extends to higher values than are typically accessible with static headspace techniques.

Measurements of the partial vapor pressure of solutes in octanol can be used to determine the ΔG° of solvation into octanol (Berti et al., 1986; Cabani et al., 1991). The vapor pressure of octanol is first determined using a static apparatus (Berti et al., 1986). The partial pressure of the solute over solution is measured at varying molar ratios and is used to solve for ΔG^′_OA (Berti et al., 1986). By regressing the molar ratio with ΔG^′_OA, the authors extrapolated to solve for ΔG^*_OA where the pressure (in atm) and molar ratio are equal to 1 (Berti et al., 1986). Equation (7) is then used to solve for ΔG°_OA (Berti et al., 1986).

The challenge of static techniques for K_OA determination is that the amount of less volatile compounds in the gas phase is too small for reliable determination. It is therefore often necessary to greatly increase the volume of air that is being equilibrated with the octanol phase. If the determination is based on the amount of solute being lost from the octanol phase, it can also be beneficial to minimize the volume of octanol in the experimental system. Dynamic techniques for measuring K_OA involve passing a stream of air through or past a stationary octanol solution. Therefore, the volume of air can be increased by extending the length of time that the air is flowing past the octanol.

The generator column techniques require the amount of analyte transferred from octanol to the gas stream to be quantified, whereas in gas stripping techniques only the rate of change in the concentration of the analyte in the gas stream or the solvent must be recorded. In the dynamic gas–liquid chromatography technique, K_OA is derived from the time it takes for a chemical to travel through a gas chromatographic column with octanol as a stationary phase. The generator column technique is by far the most commonly applied dynamic method because it is one of the few techniques readily applicable to less volatile solutes.

The generator column technique, sometimes also referred to as the fugacity meter technique, involves passing large volumes of air through a stationary octanol phase. First used by Harner and Mackay (1995), this technique has since been used extensively in different configurations (Kömp and McLachlan, 1997; Dreyer et al., 2009; etc.). Either glass wool or glass beads are coated with a small volume of an octanol solution and are placed in a column. Air passing through the column at a controlled rate for a measured length of time equilibrates with the octanol. The air is saturated with octanol prior to passage through the column to prevent the vaporization of octanol. The amount of chemical that partitions from the spiked octanol into the air phase is trapped and quantified to determine a concentration in air, C_A. Using the known concentration of the chemical in octanol, C_O, yields K_OA from Eq. (1).

This method requires the validity of several assumptions to yield reliable results. The concentration of the analytes of interest in the octanol needs to be sufficiently high to remain constant throughout the measurement. The flow rate must be sufficiently slow for the chemicals to reach equilibrium between octanol and air. The length of an experiment must balance the need to collect an amount of chemical from the air stream that is sufficient for reliable quantification but not so much that it would deplete the chemical from the spiked column.

This technique is commonly applied for measuring k_H and involves passing air past a stationary solvent phase. Two variations of this technique have been applied to measuring K_OA.

Adopting the gas stripping method by Leroi et al. (1977), Fukuchi et al. (2001; 1999) moved small air bubbles through a very small volume of octanol containing the solute of interest. Equilibration is assured by a slow flow rate and small bubble size. A temperature bath allows for measurements at different temperatures. By recording the concentration change of the solute in the gas phase over time, Fukuchi et al. derived $γo∞$ from the gas flow rate and the solute’s estimated P_L. The volume of octanol is assumed to be constant (Leroi et al., 1977). We used Eq. (9), the measured $γo∞$⁠, and the estimated P_L to derive the K_OA value. This technique has only been used for four ether compounds.

In the technique by Roberts (2005), the solute is not added directly to the octanol, but the gas is first bubbled through a small volume of the liquid solute prior to being bubbled through a volume of octanol. Once the solute has reached equilibrium between the gas and octanol, the gas concentration of the solute at the outlet will be constant. At this point, the solute is removed from the gas flow, and the gas begins to strip the octanol of any solute (Roberts, 2005). Measuring the change in the concentration of the solute at the outlet allows for the determination of a first-order rate loss constant for the chemical from octanol. When combined with the octanol volume and gas flow rate, K_OA can be obtained (Roberts, 2005). This technique has been applied to measure the K_OA of peroxyacetyl nitrate (CAS No. 2278-22-0) (Roberts, 2005) and triethylamine (CAS No. 121-44-8) (Leng et al., 2015).

Among the advantages of the gas stripping techniques are that analysis of only one phase is required and that no quantification is necessary because the change in signal strength over time can be plotted in place of concentration. This also eliminates the need for a calibration curve. Finally, this technique uses multiple measurements to obtain a single K_OA value, which increases the reliability of the experimental value. However, solute volatility limits the applicability of gas stripping techniques to a fairly narrow range of K_OA. The technique employed by Roberts (2005) is also limited to liquid solutes.

Some dynamic methods rely on the determination of the retention of a solute in a gas chromatographic column containing octanol as a stationary phase. No quantification of the amount of solute in either octanol or gas phase is necessary. The use of octanol as a stationary phase sets these methods apart from other retention time techniques using commercial columns, which rely on correlations and always require reference compounds with a known K_OA. They will be discussed in the next section.

Gruber et al. (1997) recorded the net retention volume (V_N) on columns with variable volumes of octanol (V_L) coated on the inside. When V_N/V_L is regressed against the reciprocal of V_L, the intercept yields K_OA (Gruber et al., 1997). This technique has similarities with the static variable phase ratio technique by Lei et al. (2019) described above.

Sandler et al. (Tse and Sandler, 1994; Bhatia and Sandler, 1995) used a slightly different gas chromatographic method, relying on the use of a reference compounds with a known $γo∞$ (hexane and heptane), to measure the $γo∞$ of halogenated alkanes. The ratio of the elution time of the reference compound and the solutes of interest relative to that of methane is used, together with an estimated P_L. We utilize the reported $γo∞$ and P_L to calculate K_OA using Eq. (9).

Indirect techniques seek to derive K_OA from the retention time of solutes on commercial gas chromatographic columns, i.e., the stationary phases of those columns serve as surrogates for the octanol phase. Because these surrogates are imperfect, indirect methods always require a calibration and often relate the retention times of the analytes of interest to those of reference compounds with previously measured K_OA values. There are a few variations of the gas chromatography retention time (GC-RT) technique; however, they all have in common that at least one chemical with a well-established K_OA value at different temperatures is required.

The first instance of measuring K_OA using GC-RT was by Zhang et al. (1999), who regressed capacity factors of chemicals on multiple columns with their K_OA to obtain a multiple linear regression (MLR) equation. The K_OA of multiple calibration chemicals need to be known as a function of temperature, as separate MLR equations are required for different temperatures. While the use of multiple columns with different solid phases is meant to better account for different types of interactions of a chemical with octanol (Zhang et al., 1999), Su et al. (2002) showed that a linear regression with a single column’s capacity factor worked equally well and yielded K_OA values with a smaller error.

The retention time index (RTI) method is essentially a technique for extrapolating known K_OA values within a group of structurally related compounds by linearly regressing directly determined log₁₀ K_OA values against the compounds’ RTI (Harner et al., 2000). The RTI relates the retention time of the solute to that of linear alkanes. The regression equation is then used to estimate K_OA for other related compounds using their RTI. A separate regression for different experimental temperatures is required. By further regressing the slope and intercept of these linear regressions against temperature, the log₁₀ K_OA at different temperatures can be determined solely from the RTI of a chemical. This method relies heavily on having direct measurements of K_OA at different temperatures for different congeners and RTI values for each congener. When applying this method to polychlorinated dibenzo-dioxins and -furans (PCDD/Fs), Harner et al. (2000) also accounted for the position and number of chlorine substitutions because measurements with the generator column technique had revealed that tetra-, penta-, and hexa- PCDD/Fs with 3-4 chlorines in the 2,3,7, and/or 8 positions had a higher affinity to the octanol phase (Harner et al., 2000). This illustrates the need for good calibration and reference data when using indirect K_OA measurement techniques.

Adapting a technique for the determination of P_L, Wania et al. (2002) used the retention time of a chemical relative to a single reference chemical in order to obtain K_OA. In principle, the relative retention times of the analyte (t_Ri) and the reference compound (t_Rref) are proportional to the partition ratio between the stationary phase of the column and air, which is also proportional to K_OA (Wania et al., 2002). Thus, this method only requires a single reference compound to have well established K_OA values at different temperatures. These log₁₀ K_OA values are plotted against ln (t_Ri/t_Rref) to produce a linear regression with a slope equal to ∆U_OAi/∆U_OAref − 1. The internal energy of octanol gas phase transfer ∆U_OAi then allows for the determination of K_OA at different temperatures (Wania et al., 2002). The obtained K_OA values are then regressed against literature values of K_OA obtained using direct measurement techniques. Therefore, even though only one chemical is needed as a reference compound, calibration compounds with established K_OA values are needed to improve the reliability of the results. Wania et al. (2002) also showed the importance of selecting an appropriate reference compound because interactions of different compounds with the stationary phase and octanol may be dissimilar. This technique is the most commonly applied GC-RT for K_OA determination.

Numerous techniques for estimating K_OA exist. We describe here a few of the major techniques if they had been specifically designed for estimating K_OA and if K_OA estimated with those techniques have been reported in the literature. If K_OA values had been calculated in the context of studies on passive air sampling, atmospheric particle–gas partitioning, or environmental fate modeling, they are not considered. Only articles focusing on physical–chemical property estimation techniques or work comparing experimental and/or estimated K_OA values are included within the database and in this review. Table 2 summarizes the different techniques for estimating K_OA. These techniques tend to have a wider applicability range than the experimental ones. We also list the temperature range for these methods. Most of the estimation models for K_OA are Quantitative Structure–Property Relationships (QSPRs). Density functional theory-based solvation models have also been used to determine K_OA by first obtaining ΔG°_OA of a chemical in octanol [see Eq. (6)].

QSPR techniques typically involve the regression of descriptors against the property of interest to obtain an equation of best fit that will most accurately predict K_OA. These models can be very simple, using basic thermodynamic relationships and linear regressions or using machine learning algorithms to estimate K_OA based on a series of chemical descriptors.

K_OA can be derived from other properties using thermodynamic triangles [see Fig. 1 and Eqs. (9) and (10)]. The two property values used in such an estimation should ideally be experimentally derived. If they are themselves estimated values, the uncertainty of their prediction propagates to K_OA.

Most estimations of K_OA reported in the literature are derived using Eq. (10), using either experimental or estimated values of K_OW and K_AW. This can be a useful estimation method for chemicals with well-established K_OW and K_AW values. However, Finizio et al. (1997) already noted that six K_OA values estimated this way were between 0.48 and 1.04 log₁₀ units smaller than experimental values, which may be related to the estimation yielding wet octanol–air partition ratio (K^′_OA) (see Sec. 1.1.5). Meylan and Howard (2005) conducted the first comprehensive assessment of this technique for estimating K_OA using K_OW and K_AW. They also explored the temperature dependence of K_OA by combining a temperature-adjusted K_AW value with the K_OW of a chemical at 25 °C and using K_AW and K_OW values estimated with EPISuite^TM’s HENRYWIN and KOWWIN (Meylan and Howard, 2005). This estimation technique is what is used in the KOAWIN model included in EPISuite^TM (EPI Suite Data, 2012)_.

The use of Eq. (9) is less common but advantageous as it does not yield a K^′_OA value. Abraham et al. (2001) presented the K_OA of some chemicals derived from measured P_L and S_O. Sepassi and Yalkowsky (2007) used P_L and S_O estimated from other physical–chemical properties of a compound, including boiling-point temperature and enthalpy of boiling. Best et al. (1997) applied a combination of Eqs. (6) and (10) to estimate ΔG°_OA using ΔG°_AW and log₁₀ K_OW.

Many works report K_OA values calculated using thermodynamic triangles or estimated using EPISuite^TM [e.g., Alarie et al. (1995), Sühring et al. (2016), Tamaru et al. (2019), and Xu et al. (2014)]. We have elected to not include all these K_OA values. When we did include K_OA values obtained through thermodynamic triangles in the database, we also report the original source of the two property values in the property table (see Sec. 4.2.5).

Numerous regression models for predicting K_OA exist, most frequently restricted in applicability to a specific set of closely related compounds, such as the polychlorinated biphenyls (PCBs) and naphthalenes (PCNs) or the polybrominated diphenyl ethers (PBDEs). The models differ based on the compound group, the type of regression, and the source and type of chemical descriptors. The statistical techniques applied include ordinary or MLRs, partial least-squares models, and principal component regression models. Some models also incorporate temperature into the regression analysis (Chen et al., 2003c; 2003b; 2002b; Jin et al., 2017; and Li et al., 2006). Tables 2 and 3 include a list of the K_OA models whose predictions are included in the database and the parameterization as described in this paper.

The Unified Physical Property Estimation Relationship (UPPER) model by Yalkowsky et al. (1994a) uses the thermodynamic triangle between K_OA, P_L, and S_O [Eq. (9)]. Molecular descriptors are obtained from the structure of a chemical using additive-group contribution estimations or the geometry of the structure (Lian and Yalkowsky, 2014). The descriptors are then used to derive basic physical–chemical properties (referred to as component properties, including melting and boiling points), which allow for the calculation of K_OA, K_AW, and K_OW (Yalkowsky et al., 1994a).

The UNIFAC model estimates $γo∞$ with an additive fragment-based approach with group-interaction parameters (Fredenslund et al., 1975). It also considers the volume, surface area, and the number of different groups present in the solute (Fredenslund et al., 1975). Dallas (1995) used UNIFAC to estimate K_OA and compare it to direct measurements and the MOSCED model (see Sec. 3.2). This author also compared the performance of the UNIFAC model with an infinite-dilution activity based UNIFAC model, which uses calculated interaction parameters using activity coefficients at infinite dilution, and a modified UNIFAC model that combines the original and the infinite-dilution activity based UNIFAC models. A summary of publicly available group-interaction parameters can be obtained from the UNIFAC Consortium webpage (http://unifac.ddbst.de/unifac_.html). Note that within the database, we include estimates that directly report K_OA or include a P_L for the calculation of K_OA. Papers that only report a UNIFAC-estimated $γo∞$ are not included [e.g., Castells et al. (1999), Eikens (1993), and Li et al. (1995)].

While machine learning algorithms resemble regression models in that they use descriptors to predict K_OA, they differ in the approach to correlating the different variables. Jiao et al. (2014) created an artificial neural network model that uses molecular distance-edge vector index descriptors to predict K_OA. The model is designed to have the smallest RMSE for the validation set (Jiao et al. 2014).

The OPERA model (Mansouri et al., 2018; Mansouri and Williams, 2017) is also a QSAR model developed using machine learning. OPERA uses the k-nearest neighbor approach and PaDEL descriptors for the number of hydrogen bond donor and the hexadecane-air partition ratio to estimate K_OA. While estimates from OPERA are not included in the database, these values are easily obtained from the CompTox Dashboard (Williams et al., 2017) or the model can be downloaded from GitHub (https://github.com/kmansouri/OPERA).

Solvation models estimate the ΔG°_i of a chemical in a solvent i. The difference of ΔG°_i in octanol and the gas phase can be used to estimate ΔG°_OA, which, in turn, can be used to estimate K_OA (Nedyalkova et al., 2019). Such models have been applied to estimate K_OA of a wide range of chemicals, and numerous variations of models for estimating ΔG°_OA exist in the literature. The information included in the database is limited to models that have been specifically designed to estimate ΔG°_OA and to predictions made during the comparison and assessment of these solvation estimation techniques. A subset of universal solvation models that estimate ΔG° for various air–solvent interactions are also considered. Specifically, this includes estimates from MOSCED (Modified Separation of Cohesive Energy Density) (Thomas and Eckert, 1984) and various universal solvation models [e.g., Best et al. (1997)].

The MOSCED model estimates ΔG°_i in a solvent as the difference between the cohesive energy density of the pure phase and the solution (Thomas and Eckert, 1984). The SM8AD and SMD models are universal solvation models that solve for the electrostatic contribution using either the generalized Born approximation with asymmetric de-screening (SM8AD) (Marenich et al., 2009a) or the nonhomogeneous Poisson equation (SMD) (Marenich et al., 2009b). These models can be parameterized using different density functionals, which can produce slightly different results (Nedyalkova et al., 2019). Multiple variations of these solvation models for multiple solvents exist, and we have included a selection of estimates, such as Best et al. (1997), Duffy and Jorgensen (2000), Giesen et al. (1997), Li et al. (1999), and Zhu et al. (1998). While there are very likely far more universal solvation models for ΔG°_OA in the literature, we have included only selected estimates in the database because these models often merely improve upon previous iterations of the SM-AD and SMD models and predict the ΔG°_OA for sets of chemicals that also have experimental ΔG°_OA values.

The COnductor-like Screening Model for Realistic Solvents (COSMO-RS) software suite can also be used to estimate K_OA for chemicals [e.g., Parnis et al. (2015)]. COSMO-RS applies quantum chemical density functional theory and statistical thermodynamics to derive ΔG° values (Klamt et al., 2009). Endo and Hammer (2020) introduced a fragment contribution model for extrapolating COSMO-RS predicted K_OA for short-chain chlorinated paraffins, which reduces calculation times.

Another tool for estimating K_OA is SPARC Performs Automated Reasoning in Chemistry’s online physicochemical calculator (SPARC) (available at http://archemcalc.com/). The details on how exactly SPARC works are not widely available. However, it is noted that linear free energy relationships are used to estimate thermodynamic properties such as K_OA (Hilal et al., 2003). While SPARC has been applied repeatedly to estimate K_OA (Zhang et al., 2016), the calculated K_OA values are not often reported [e.g., Stenzel et al. (2014) and Wang et al. (2012)]. We only include K_OA estimates from SPARC in the database if they are compared to experimental values or other estimates; thus, K_OA values reported in papers such as Weschler and Nazaroff (2010) have not been included.

Some other publications on K_OA estimation models, including various MLR models such as poly-parameter linear free-energy relationships (ppLFERs), COSMOtherm (Klamt, 2018; 2011), and OPERA (Mansouri et al., 2018), do not always report the estimated K_OA values. Thus, K_OA estimates made with these approaches are not included in the database. In addition, there are published MLR models for predicting K_OA that do not report estimated K_OA values and thus are not included in the database; a summary of these models is included in Table 4.

Some models have been designed to estimate the temperature dependence of log₁₀ K_OA for a series of compounds. For example, the model by Yang et al. (2018) estimates the temperature dependence of K_OA for PBDEs. Mintz et al. (2008; 2007) published two ppLFERs using Abraham descriptors to estimate ∆H°_OA of a wide range of chemicals.

This database includes all measured or estimated K_OA values that we could locate in the literature using the Web of Science using variations of the keywords: Octanol–air partition coefficient (K_OA), octanol–gas partition coefficient, Ostwald coefficient octanol, and Gibbs free energy octanol. References were also found by looking up citations included in the identified papers. A total of 112 literature sources were found to contain K_OA data. Forty-seven sources included estimated K_OA values, while 70 contained measured values. The database incorporates 209 K_OA values from three dissertation theses (Cheong, 1989; Dallas, 1995; and Özcan, 2013). While a large portion of the work by Dallas (1995) was published in Abraham et al. (2001) and Dallas and Carr (1992), a portion of K_OA estimates from this thesis are not available in the peer-reviewed literature. To the best of our knowledge, K_OA data from Cheong (1989) and Özcan (2013) have not been published in the peer-reviewed literature. The search was limited to publications written in English, although articles containing log₁₀ K_OA estimates have also been published in other languages [e.g., Zhang et al. (2005) and Zou et al. (2005)].

We have included the error of a measurement or estimation in the database. We have also noted where the K_OA reported is for a mixture of isomers or the chemical structure is ambiguous [e.g., Harner and Bidleman (1998), Kömp and McLachlan (1997), and Vuong et al. (2020)]. In some instances where a single paper has reported more than one K_OA using different techniques, we note which technique or value was recommended by the authors [e.g., Su et al. (2002)].

The database is provided in a Microsoft Excel workbook and as an R package. The data are stored in seven distinct tables (Fig. 3) to allow users to sort and filter the data based on various criteria, including author, publication year, and measurement or estimation technique.

The Chemical Table provides information regarding the name, CAS number, SMILES notation, and other chemical identifiers for each solute. Each unique chemical is associated with a chemical identification number within the database (chemID). Similarly, each unique literature source and author are assigned a unique identifier in the Reference Table (refID) and Author Table (auID). The Author-Reference Table is used to connect the information presented in both. Each method for predicting or estimating K_OA within a paper is also assigned a unique identifier (methID). The Property Table contains the P_L, γ_oct, K_AW, K_OW, and ΔG° values originally reported in the paper in SI units (with the exception of ΔG°, which is in kJ mol⁻¹) and used to calculate the K_OA; each value is assigned a propID. Finally, the K_OA values are reported in the K_OA Table, with each datapoint uniquely associated with a dataID.

The QSAR-ready SMILES notation for the 1643 compounds with literature data was taken from EPA’s CompTox Dashboard (Williams et al., 2017) or PubChem (Kim et al., 2017) or, if none were available, created using ChemDraw. CAS numbers and names of chemicals were verified using both SciFinder and the CompTox Dashboard. Canonical SMILES for compounds were produced using Open Babel (version 3.00) (O’Boyle et al., 2011). We also include the IUPAC name, common name, common acronym, and alternative names and acronyms for each chemical. The list of names for each compound is not exhaustive, and we recommend searching for chemicals using their CAS number. We also group chemicals into over 50 broad categories, including PCBs, PBDEs, PCNs, amines, ketones, and so on. This categorization of chemicals is also non-exhaustive as many chemicals may fall into more than one group. Within the database, each chemical is assigned a unique identifier, a chemID. In some instances, the K_OA of a mixture of two or more chemicals is also reported, these mixtures are also given a unique chemID, and the CAS number and the identifying information for all chemicals in the mixture are included.

K_OA data are stored in the K_OA Table. Each K_OA value is assigned a unique identifier (dataID), which is associated with a specific chemical (chemID) and method (methID). For each K_OA value, we also include the temperature of the measurement, any errors reported for the measurement, and comments or notes for the datapoint. The comments indicate if the K_OA reported is for an isotopically labeled species or if any typo corrections and assumptions were made during the data curation process of the original work.

Each reference is stored in the Reference Table and assigned a unique identifier (refID). Some references may report or compare the results of different K_OA measurement or experimental techniques; therefore, a single reference can be associated with multiple experimental and estimation techniques. Each technique from each reference is also assigned a unique identifier (methID). Different measurement and estimation techniques have been employed and published at different times over the past few decades, which is shown in Fig. SI 7 in the supplementary material.

In Sec. 1.1, we discussed the different ways log₁₀ K_OA has been reported in the literature. The Property Table (dark blue table, Fig. 3) includes the data originally reported in the literature in standardized units, except ΔG°, which is reported in units of kJ mol⁻¹. This includes converting all reported k_H values to K_AW for convenience. Each property value within the database is associated with exactly one K_OA datapoint in the main K_OA Table; however, a single K_OA datapoint may be associated with more than one property value. Each property datapoint is associated with a chemical (chemID), method (methID), reference (refID), and K_OA value (dataID). There are 2228 dataIDs that are associated with 3723 propIDs. Figure SI 4 shows the distributions of the different property data included in the database.

Each technique and method (i.e., methID) for determining and reporting K_OA values was assigned a score based on whether

• the method is an estimation or empirical technique,

• the description of the used methodology is sufficiently detailed,

• some analysis of the methods is provided (including, e.g., their possible limitations and scope), and

• an error of the K_OA value is reported and/or whether an external validation of the K_OA value was performed.

Each factor is weighted equally, and the papers are thus scored from 0 to 4. The point assigned for each factor is binary, and thus, there are no half points allocated. These points are only ascribed to method, description, analysis, and error of the reported K_OA value. For example, a method from a paper about empirical measurements of K_OA, providing a detailed description and analysis of the methodology and the error of the prediction, will have a score of 4, whereas a thermodynamic triangle method without any additional details or analysis will have a score of 1. Based on this categorization, a total of 36 methods scored 4, 81 methods scored 3, 32 methods scored 2, and 14 methods scored 1. Note that we define each method as having a unique identifier (methID). Figure SI 2 shows the distribution of the scores across the log₁₀ K_OA range for experimental and estimated data.

The database contains 13 264 K_OA values for 1643 different chemicals between −50 and 110 °C. Of these, 2517 values (19%) are experimentally derived and 10 747 values (81%) are estimated (Fig. 4). Notably, the data are bimodally distributed with respect to log₁₀ K_OA, with one maximum between log₁₀ K_OA 2 and 4 and a second between log₁₀ K_OA 6 and 11. If we consider only measurements or estimates made at 25 °C, the peaks become even more pronounced (Fig. SI 1). These peaks are a result of the difference in the applicability range of the different measurement techniques. Many estimation techniques require a set of K_OA data to develop and train new models, so the distribution of estimated K_OA values follows a similar pattern.

There are 2517 measured K_OA values in the database for 704 chemicals. Of these, 1524 are directly measured values, while 993 are obtained by indirect measurements using chromatographic retention time techniques. The highest and lowest determined log₁₀ K_OA values are −1.8 and +14 for helium (CAS No. 7440-59-7) and trichloro-benzo[a]pyrene (CAS No. 97303-27-0), respectively. Most of the reported values (52%) are for log₁₀ K_OA values between 6 and 11, followed by measurements between 2 and 5 (30%). Few measurements have been reported for chemicals with a log₁₀ K_OA less than 2 (4%) or greater than 11 (8%). There is also a drop in the number of measurements between log₁₀ K_OA 5–6 (4.92%). Most log₁₀ K_OA values less than 4 are measured directly using static techniques, while dynamic or indirect techniques were used to measure log₁₀ K_OA values greater than 6 [Fig. 6, Panel (a)]. Approximately, a quarter (26.2%) of measurements in the database were obtained using static techniques, a third (34.2%) using dynamic techniques, and 39.4% using indirect techniques. Figure 5 displays in more detail the K_OA range in which different static and dynamic techniques have been applied.

There are 924 K_OA values (37%) for 573 chemicals at 25 °C, and most measurements (51%) are made within the 20–30 °C range. There are relatively more measurements made at high temperatures (>30 °C) when log₁₀ K_OA is less than 3, presumably to extend the applicability of static techniques to somewhat less volatile chemicals. More surprisingly, there are also relatively more measurements at cooler temperatures (<20 °C) when log₁₀ K_OA is greater than 7 [Fig. 6, Panel (b)]. This is likely because many of those less volatile chemicals are environmental contaminants and the partitioning behavior at environmentally relevant temperatures is of primary interest. Measurements in the log₁₀ K_OA range between 7 and 10 have been made at the most diverse range of temperatures.

The types of chemicals for which directly measured K_OA values have been reported are shown in Table 5. Chemicals with measured log₁₀ K_OA values greater than 6 are generally persistent organic pollutants, including CBz, PCBs, PAHs, PCNs, and PBDEs [Fig. 6, Panel (c)]. There is greater diversity among the chemicals with low measured log₁₀ K_OA values, including simple hydrocarbons such as alkanes, alkenes, cyclic hydrocarbons, haloalkanes, alcohols, and organosilicons. Measurements of K_OA for small, volatile molecules are often motivated by explorations of basic partitioning behavior (Abraham et al., 2001) or methodological issues (e.g., Lei et al., 2019). At the very low log₁₀ K_OA range (<1) are typically short chain alkanes, noble gases, and inorganic gases [e.g., xenon (CAS No. 7440-63-3) and carbon monoxide (CAS No. 630-08-0)].

A subset of experimental log₁₀ K_OA data was assessed to be unreliable. These measurements were made for polar compounds using a gas chromatography retention time (GC-RT) technique. Figure 7 compares K_OA values obtained using GC-RT methods against directly measured values, if they are available. While there is generally very good agreement between the reported values, some notable exceptions become apparent. K_OA values for a series of fluorotelomer alcohols (FTOHs) measured with the GC-RT technique are much lower than those measured using the generator column technique. We suspect that the GC-RT measurements are erroneous due to the high polarity of these compounds and their ability to undergo hydrogen bonding. These compounds would be expected to interact much more strongly with octanol than with the nonpolar GC column used, particularly relative to hexachlorobenzene (CAS No. 118-74-1), the reference compound used in the study (Lei et al. 2004). Thus, when measured with GC-RT, log₁₀ K_OA for such chemicals tend to be too low.

Large discrepancies are also apparent for benz[a]anthracene (CAS No. 56-55-3) and benzo[a]pyrene (CAS No. 50-32-8), where K_OA values from the GC-RT techniques (Odabasi et al. 2006a) are much higher compared to those obtained with the droplet kinetics technique (Ha and Kwon 2010). In Fig. SI 3, we compare the measurements made using the droplet kinetics technique against other experimental measurements for the same compounds. The K_OA for benz[a]anthracene by Ha and Kwon (2010) is almost an order of magnitude smaller than most other measured and estimated values for this compound. On the other hand, the value obtained by GC-RT is within 0.5 log₁₀ units of estimates using solvation models (Fu et al. 2016), the UPPER model (Lian and Yalkowsky 2014), and thermodynamic triangles (Sepassi and Yalkowsky 2007). The log₁₀ K_OA for benzo[a]pyrene by Ha and Kwon (2010) is also lower than almost all other reported values. The GC-RT derived log₁₀ K_OA for both PAHs is in excellent agreement with the final adjusted value derived by Ma et al. (2010). In addition, as these PAHs are relatively non-polar, the GC-RT technique should be applicable and, in any case, not lead to K_OA values that are too high. We therefore suspect that in this case, the values reported by Ha and Kwon (2010) are more likely to be erroneous than the GC-RT values.

There are many more GC-RT-derived K_OA values without complementary directly measured values. To identify other potentially flawed values, we compared the GC-RT measured value with predictions made by three different prediction models. Figure 8 displays the residuals between predicted and GC-RT measured values, whereby chemicals are color-coded by the strengths of their H-bonding with octanol. The latter is quantified as aA + bB, where A and B are the Abraham solute descriptors for hydrogen bonding acidity and basicity of the solute and a and b are the respective system constants from the poly-parameter linear free energy equation for log₁₀ K_OA by Endo and Goss (2014). Experimental solute descriptors were obtained using the UFZ-LSER Database (Ulrich et al. 2017); if experimental solute descriptors were unavailable, estimated solute descriptors were obtained using the IFSQSAR model developed by Brown and available on GitHub (https://github.com/tnbrowncontam/ifsqsar) (Brown 2014; Brown et al. 2012).

GC-RT-derived log₁₀ K_OA of hydrogen-bonding chemicals (aA + bB > 0.5) have unusually large residuals with all three prediction techniques, suggesting a large bias. Most residuals are negative, implying that the K_OA for such chemicals is biased low, which is consistent with expectations. We conclude that the GC-RT method is unsuitable for measuring the log₁₀ K_OA of polar, and especially hydrogen-bonding, chemicals because (i) the interactions between the octanol and the reference chemical are not necessarily comparable to the interactions between octanol and the analyte of interest and (ii) the way the analyte interacts with the stationary phase will not be similar to its interaction with octanol due to the latter’s capacity to undergo hydrogen bonding.

Within the database, we have noted which K_OA values obtained by GC-RT techniques may be erroneous due to the high polarity of the chemical.

The range of 10 747 estimated K_OA values in the literature is much larger than that of the experimentally derived values. The lowest estimated log₁₀ K_OA value is −3.02 for propylnitrile (CAS No. 107-12-0) by Best et al. (1997) using a thermodynamic triangle approach based on ΔG°_AW and log₁₀ K_OW. The highest estimated log₁₀ K_OA value, 30.20 for 1,2-bis[(2,3,4,5,6-pentabromophenyl) methyl] 3,4,5,6-tetrabromo-1,2-benzenedicarboxylate (CAS No. 82 001-21-6) by Zhang et al. (2016), was obtained using the thermodynamic triangle approach implemented in EPISuite’s KOAWIN.

The general distribution of estimated K_OA values is similar to that of the experimentally derived values, with the majority of estimated values (70%) within the log₁₀ K_OA 6–12 range (Fig. 9). A large portion of estimated log₁₀ K_OA values are also in the 2–5 range (17.5%). Fewer estimates are made above log₁₀ K_OA 13 (3.5%) or below 1 (1.3%). Between log₁₀ K_OA 5 and 6, there are also few estimates (3.4%).

Half (50.7%) of the estimated values are derived from some form of linear regression, as described in Sec. 3.1.2. Solvation models for estimating K_OA are also very commonly used (34.6%), followed by thermodynamic triangle estimation techniques (12.0%). Both models are used across a broad K_OA range. Likewise, the UPPER model is not restricted to a specific range of chemicals because it is rooted in principles applied to thermodynamic triangles. However, estimates are not commonly available in the literature, and almost two thirds of published values obtained with UPPER (63.7%) fall within the log₁₀ K_OA range between 1 and 5. The UNIFAC model is typically applied to estimate K_OA of volatile chemicals, and thus, reported values range only from 0.4 (tetrahydropyran; CAS No. 142-68-7) to 2.38 (dimethyl sulfoxide; CAS No. 67-68-5). Estimates made using machine learning techniques are limited to the work by Jiao et al. (2014) on PBDEs. As methods for estimating physical–chemical properties using neural networks and machine learning are developed further and because these approaches are not limited to a specific subset of chemicals, we expect their estimation range to widen significantly.

Most estimates are for log₁₀ K_OA at 25 °C (71.9%). There are 3023 estimated K_OA values for 486 different chemicals at non-standard temperatures, which have been reported by nine publications using either linear regressions, thermodynamic triangles, or solvation models. Linear regression models use temperature-dependent experimental K_OA values for training and validation (Chen et al. 2003c, 2003b, 2002b; Li et al. 2006; Mathieu 2020). The descriptors for these models are temperature-dependent (e.g., X_i/T) because temperature and K_OA are inversely correlated. Meylan and Howard (2005) estimated the temperature dependence of K_OA from that of k_H, i.e., ignore the temperature dependence of K_OW during the application of the thermodynamic triangle of Eq. (10). Li et al. (2020) estimated a temperature-dependent K_OA by estimating ΔG°_OA at 25 °C using a solvation model and then solving Eq. (6) with different values of T. The assumption that ΔG°_OA is not strongly temperature-dependent is similar to assuming that ΔU_OA or ΔH_OA are weakly temperature-dependent. Some solvation models, such as COSMOtherm, can directly estimate ΔG°_OA at different temperatures (Parnis et al. 2015).

The chemical classes for which K_OA is commonly estimated reflect the availability of experimental data for K_OA. The most commonly estimated K_OA values are for PCBs (22.7%), PCDDs (13.6%), PAHs (12.4%), PBDEs (11.1%), and PCNs (0.7%) within the log₁₀ K_OA 9–12 range. At the lower range of K_OA values, there are more estimates of alcohols and haloalkanes. There are also K_OA estimates for different CBz, arenes, alkanes, and OPEs. A full list of all methods and papers reporting estimated log₁₀ K_OA values is included in Table 6.