Estimating the silica content and loss-on-ignition in the North American Soil Geochemical Landscapes datasets: a recursive inversion approach

Rock, sediment and soil chemical and mineralogical characterizations are fundamental to the discipline of geochemistry, particularly when it comes to applications in the fields of mineral exploration, environmental management, agronomy, horticulture and forestry, and landuse decision making. Most rocks, sediments and soils on Earth contain minerals that include silicon (Si) and oxygen (O) in their structure (e.g. silicates), and often also hydrogen (H; e.g. phyllosilicates) and, perhaps less commonly, carbon (C; e.g. carbonates) (e.g. Finkl 1981; Deer et al. 2013; Schaetzl and Anderson 2007). Traditionally, total analyses for major elements have been reported as oxides (the term ‘major’ is used here to include components with generally greater than 0.1 percent abundance), typically obtained using X-ray fluorescence (XRF) as the analytical method, which determines total content (regardless of the host, speciation or oxidation state of each major element). The analysis of rock, sediment and soil samples by XRF is often complemented by the gravimetric determination of ‘loss-on-ignition’ (LOI) obtained by heating the sample to a set temperature (e.g. 900°C) and measuring the mass loss compared to the starting sample at standardized temperature, pressure and humidity. LOI has several components, including adsorbed water (H₂O; e.g. interlayer water in clay minerals), combined H₂O (e.g. hydrated minerals and labile hydroxyl-compounds), carbon dioxide (CO₂; e.g. from carbonates and organic matter) and volatile elements (e.g. Hg).

One advantage of reporting the major components of a rock, sediment or soil sample as oxides is that their sum, when complemented by LOI and trace elements (TEs), should add up to 100 weight percent (wt%) of the sample. Having a complete sample analysis, or at least as complete as practically feasible, is important to give confidence that the sample is well characterized, which implies that the composition is closed or full and not a subcomposition. This has implications in subsequent data processing, including in the development and application of Compositional Data Analysis methods (e.g. Chayes 1960; Aitchison 1986; Scealy et al. 2015).

Another benefit of a complete sample analysis is the direct relationship between the geochemical and mineralogical compositions, via the knowledge (or modelling) of the minerals’ stoichiometric compositions. Deriving the most plausible mineralogy from geochemistry is a non-unique inversion problem known as ‘normative analysis’ (e.g. Caritat et al. 1994; Aldis et al. 2023). This is a useful way to ensure that chemistry and mineralogy of a sample are mutually compatible, especially for finer-grained samples where optical or even electronic microscopic techniques reach their resolution limit to helpfully identify minerals.

In recent decades, another family of analytical methods has gained in popularity, mainly due to its high precision and multi-elemental capability, namely inductively coupled plasma atomic emission spectroscopy (ICP-AES) and -mass spectrometry (ICP-MS). The ICP-based methods typically require a digestion of the rock, sediment or soil sample to present it to the instrument as a liquid phase (laser ablation is an alternative input mode not discussed here). This digestion, which can range from near-total to weak, and from selective to nonselective, is crucial to document in detail as it controls how to interpret the geochemical results (e.g. Mann 2010). Analytical data generated by ICP often consist of a list of 30 or more elements, generally reported in parts per million mass/mass (ppm m/m; equivalent to mg kg^–1 and $μ$g g^–1). These element concentrations for the major elements (Al, Ca, Fe, etc.) can readily be converted to oxide equivalents (yielding a ‘pseudo-XRF’ result).

Generally a full or complete analysis of a rock, sediment or soil, expressed as oxides, consists of Al₂O₃, CaO, FeO, Fe₂O₃, K₂O, MgO, MnO, Na₂O, P₂O5, SiO₂, SO₃ and TiO₂ concentrations typically expressed in wt%, either directly obtained by XRF or converted from ICP elemental data. Often, the two Fe analytes, reduced and oxidized Fe, are reported together as Fe₂O_3tot. Added together and supplemented by LOI and TE concentrations, the full analysis is considered complete and should sum to (or close to) 100 wt%. Any discrepancy represents components not analysed for and/or uncertainty.

The North American Soil Geochemical Landscapes (NASGL) project is a recent continental-scale geochemical survey of the conterminous USA (Smith et al. 2013, 2014, 2019; see the recent project review in Smith 2022). It sampled soil from three levels (0–5 cm depth, A horizon and C horizon) at 4857 sites, the <2 mm fraction of which was analysed for 45 major and trace element concentrations by methods yielding ‘total or near-total’ elemental content. The chemical elements reported were Ag, Al, As, Ba, Be, Bi, C, Ca, Cd, Ce, Co, Cr, Cs, Cu, Fe, Ga, Hg, In, K, La, Li, Mg, Mn, Mo, Na, Nb, Ni, P, Pb, Rb, S, Sb, Sc, Se, Sn, Sr, Te, Th, Ti, Tl, U, V, W, Y and Zn. Elements were mostly analysed and quantified by ICP-AES or ICP-MS, after a four-acid (hydrochloric, nitric, hydrofluoric and perchloric acids) digestion of the milled samples at a temperature of between 125 and 150°C (see Smith et al. 2013 for more detail). Note that Si was not included in the contracted analytical package. As ICP-based analytical techniques cannot quantify O and H present (in fact, abundant) in most if not all rock, sediment, or soil samples, the sum of all its analytes (Al, …, Zn) falls well short of one million ppm (a complete composition); indeed in the C horizon dataset, for instance, the sum of all ICP analytes ranges from 1134 to 390 740 ppm (average 144 869 ppm). Table 1 presents a brief statistical summary of the geochemical composition of NASGL C horizon soil samples for the major elements. The C horizon dataset is used herein to illustrate our method.

The NASGL project also analysed and quantified mineralogy in those samples. The minerals quantified were quartz, K-feldspars, plagioclases, (total feldspars), 14 Å clays, 10 Å clays, kaolinite, (total clays), gibbsite, calcite, dolomite, aragonite, (total carbonates), analcime, heulandite, (total zeolites), gypsum, talc, hornblende, serpentine, hematite, goethite, pyroxene, pyrite, other and amorphous (phases in parenthesis are summations of other minerals). The amorphous phase typically consists of material that is poorly diffracting; this will generally include clay minerals, various forms of micro-quartz, Fe-, Mn- and Al-oxyhydroxides, organic matter, volcanic glass, etc. (e.g. Tan et al. 1970; Smith et al. 2018; Tsukimura et al. 2021). Minerals were analysed by X-ray diffraction (XRD) and quantified using a Rietveld refinement method (Smith et al. 2013). Unlike the geochemical data, the XRD data are ‘complete’ in the sense that they add up to 100 wt% (range 99.6 to 100.2 wt%, average 100.03 wt%, for the C horizon). Table 2 presents a brief statistical summary of the mineralogical composition of NASGL C horizon soil samples.

One of the shortcomings of the NASGL project is that neither Si/SiO₂ nor LOI were reported in the released geochemical datasets. The present paper aims to propose and test a method for estimating those missing, yet crucial, parameters.

As the NASGL project did not use XRF analysis, we first need to convert the 10 reported major elements (Al, Ca, Fe, K, Mg, Mn, Na, P, S and Ti) into oxides, which is readily achieved by dividing each elemental concentration by the atomic weight of the element, multiplying this by the molecular weight of the oxide, and adjusting for any unit change (e.g. dividing by 10 000 to convert from ppm to wt%). These oxides are hereafter referred to as other_oxides to indicate that they do not include SiO₂. For one of the most common soil components, Si, no reported elemental or oxide concentration exists and it must thus be estimated. The proposed method for estimating SiO₂, which draws upon both the geochemical and the mineralogical analyses of the NASGL samples, is described below and the workflow illustrated in Figure 1. A worked example is provided in a Microsoft Excel™ spreadsheet (see the section below on Datasets).

Initially, two estimates for SiO₂ are calculated by inverting mineralogical information; neither is ideal, as the first is likely to give a minimum, the second a maximum value for SiO₂. Next, a ‘consensus’ SiO₂ concentration is obtained recursively from the two aforementioned estimates. Finally, the LOI is calculated to obtain a closed full composition at 100 wt%. The detailed steps are described below.

• Step 1: Data preparation. The geochemical and mineralogical data for soil samples of the conterminous USA (A and C horizon datasets) were downloaded from https://mrdata.usgs.gov/ds-801/. Samples (rows) which had either incomplete or missing geochemical or mineralogical quantification (e.g. insufficient sample material) were removed. Analytes (columns) with excessive censored values (below detection/reportable limit) were removed (e.g. Ag, Cs, Te; Grunsky et al. 2018). Concentration units were unified (ppm) and censored data were imputed using the zCompositions package (lrEM function) in the R computing environment (Palarea-Albaladejo et al. 2014). Note that the imputation step is not critical to the present estimation workflow and other ways of handling censored data may be applied. After imputation, the major elements were converted to oxides and all analytes were expressed as wt%.

• Step 2: Inverting ‘normative’ SiO₂ due to silicate minerals. The ‘normative’ SiO₂ is the amount of SiO₂ each sample must contain to be consistent with its mineralogy (technically this is a reverse normative or inversion approach). This ‘normative’ SiO₂ calculates and sums the contributions in SiO₂ of each Si-bearing mineral (silicate), for example, 1 * quartz + 0.6476 * K-feldspar + … + AVERAGE (0.483, 0.5985, 0.5549, 0.5173) * pyroxene. The multipliers are the mass proportions of the relevant oxide (e.g. SiO₂) in each mineral (e.g. K-feldspar above), and were sourced from https://webmineral.com. Where more than one end-member mineral exists for a group (e.g. a solid solution), the average of the (most common) end-members is used (e.g. pyroxene above). Table 3 summarizes the proportional multipliers used in this paper. This first estimate of SiO₂ does not consider the phases ‘other’ and ‘amorphous’. Amorphous has a median abundance of 17.5 wt% and a maximum of 95.2 wt% in the NASGL C horizon dataset. It is likely to contain forms of microcrystalline silica, such as opal-A; e.g. Achilles et al. 2018), and therefore the ‘normative’ SiO₂ calculated here could, and most likely does, underestimate the real SiO₂ concentration.

• Step 3: Inverting LOI due to hydrate and carbonate minerals. The ‘normative’ H₂O and ‘normative’ CO₂ components of LOI in each sample were calculated to be consistent with the mineralogy (e.g. amounts of gypsum and calcite). This is done in a similar way as described above, but applied to all O- and H-bearing (hydrate) minerals and all C-bearing (carbonate) minerals, respectively. The relevant proportional multipliers used are also given in Table 3. As for SiO₂, the ‘normative’ H₂O and CO₂ contents of the amorphous phase are not known and likely important (e.g. Achilles et al. 2018). Thus this method could, and most likely does, underestimate the real LOI concentration.

• Step 4: Calculating a second estimate for SiO₂. A second (maximum) SiO₂ estimate is calculated by the difference 100 wt% – Sum(other_oxides, TEs, ‘normative’ H₂O, ‘normative’ CO₂). It could, and most likely does, overestimate the real SiO₂ concentration because LOI is almost certainly underestimated (see above). Note that in some instances, the first estimate of SiO₂ is larger than the second, which we interpret to result either from uncertainty in the mineralogical quantification (amounts of silicate, hydrate and carbonate minerals are not consistent with the geochemistry), or from overestimation of ‘normative’ H₂O (‘normative’ CO₂ being well constrained by carbonate minerals).

• Step 5: Recursively estimating a ‘consensus’ SiO₂. A ‘consensus’ SiO₂ is then calculated recursively by first taking the average of the above two SiO₂ estimates. For some samples, this SiO₂ estimate results in the Sum(all_oxides, TEs, ‘normative’ H₂O, ‘normative’ CO₂), where all_oxides include the ‘consensus’ SiO₂ determined at Step 4, to exceed 100 wt%; in these cases, the SiO₂ estimate is trimmed so that this sum is 100 wt%.

• Step 6: Calculating total LOI. Finally, the LOI_rest, that is volatiles other than the ‘normative’ H₂O and ‘normative’ CO₂ calculated at Step 3 above, are calculated as the difference 100 wt% – Sum(all_oxides, TEs, ‘normative’ H₂O, ‘normative’ CO₂). This LOI_rest is likely to comprise H₂O and CO₂ in the amorphous phase as well as any other volatiles not specifically accounted for above. From here, total LOI or LOI_tot is calculated as Sum(‘normative’ H₂O, ‘normative’ CO₂, LOI_rest). Note that in a few cases where LOI_tot is zero it is replaced by 0.0001 wt% to allow log-transformation.

The resultant final estimates for SiO₂ in the C horizon samples from the NASGL project have a distribution as represented in the Tukey boxplots (Tukey 1977) of Figure 2, which seem reasonable compared to the distribution of the other oxides. SiO₂ is clearly the most abundant major oxide in the NASGL soil samples, as is both expected and consistent with other regions (e.g. Australia; see Caritat and Cooper 2011a). The distribution of LOI is also illustrated in Figure 2. Table 4 summarizes the statistics of the estimated SiO₂ and LOI concentrations derived herein for both the A and C horizons.

Minerals not specifically quantified in the NASGL project are assumed to be included in the category ‘other’, so for instance if soils contain halite (NaCl) the mineralogy still adds up to 100 wt% but the geochemistry will not be explicitly accounting for halite. The contained Na will be part of the reported total Na, or converted Na₂O wt%, but the Cl, generally not be reported in an ICP-based analysis, is assumed to be part of the ‘missing’ composition (estimated LOI in this case).

Figure 4 shows the major oxide, including the SiO₂ estimated as described above, TEs, ‘normative’ H₂O, ‘normative’ CO₂ and LOI_rest of five selected samples from the NASGL C horizon dataset. Those samples were deliberately chosen to span the range of soil compositions in the dataset: sample from Site ID 7327 (California) is an Al-rich sample, 972 (Texas) is Ca-rich, 444 (Maryland) is Fe-rich, 12779 (Colorado) is K-rich and 3808 (Florida) is Si-rich. Without the estimates for SiO₂ and LOI (and its components), only between 0.2 (3808) and 54 wt% (972) of those samples would be geochemically characterized; the rest would be unknown. This unknown ‘gap’ is shown by the present estimation technique to comprise widely varying proportions of SiO₂ (from silicates), H₂O (mainly from silicates), CO₂ (from carbonates) and other volatile phases (from the amorphous phase and possibly other volatile components). It is thus important to provide estimates for each sample that honour the known mineralogical characteristics rather than apply a one-size-fits-all estimation of these parameters. Table 6 shows the mineralogy and geochemistry, including the ‘gap’-filling SiO₂ and LOI estimates, for these five samples.

For instance, sample 7327 contains significant clays (40.5 wt% kaolinite) and thus has not only elevated Al₂O₃, but also SiO₂ and LOI (H₂O) concentrations. Sample 972 holds significant carbonates (64.1 wt% calcite) as reflected not only by the elevated CaO, but also CO₂ concentrations. Sample 444 comprises significant amorphous material (41 wt%) as well as notable clay (24.2 wt% of combined 14 Å clay and kaolinite), pyroxene, talc and hematite contents, imparting a significant Fe₂O_3tot, MgO, moderate SiO₂ and relatively low LOI concentrations. Sample 12779 contains 60 wt% combined K-feldspar and plagioclase and some 10 Å clay, translating into an SiO₂-, Al₂O₃- and K₂O-rich geochemical makeup. Finally, sample 3808 contains 98.5 wt% quartz and 1.5 wt% K-feldspar, giving a geochemical composition overwhelmed by SiO₂ (estimated at 99.6 wt%); it probably also contains trace amounts of anatase or other Ti-bearing phase(s), undetected by the XRD method applied, to account for (some of) the 0.13 wt% TiO₂ reported geochemically.

Maps of the distributions of estimated SiO₂ concentrations in the NASGL A and C horizons are shown in Figure 5a and b, respectively. The data are classified in 10 quantile (decile) classes and coloured as per the mapping convention of Smith et al. (2014, 2019). The spatial distributions (circles in Fig. 5a, b) show strong similarities with the interpolated quartz distribution maps in the A and C horizons (rasters in Fig. 5a, b from figures 140 and 141 in Smith et al. 2014, respectively), reflecting a dominant mineralogical control on the SiO₂ concentrations.

The proposed method to estimate the missing SiO₂ and LOI data was validated by applying it to an Exploring for the Future (EFTF; http://www.ga.gov.au/eftf) dataset (as yet unpublished) from Australia, which comprises geochemistry by XRF (including SiO₂) and ICP-MS, and mineralogy by XRD. The dataset of 260 samples from the National Geochemical Survey of Australia (NGSA; Caritat and Cooper 2011a; Caritat 2022) crosses the continent from the temperate coast of South Australia (SA), through semi-arid parts of the Northern Territory (NT) and Queensland (Qld), to the tropical Gulf of Carpentaria coast, defining the SA–Qld–NT study area (large crosses in Fig. 6). Thus a wide range of geological, geomorphological and climate conditions are intersected by this dataset, making it a suitable comparison to the NASGL dataset.

Figure 7 shows the correlation between measured and estimated SiO₂ as per the recursive inversion method described herein for the SA–Qld–NT dataset. The correlation is linear, strong (R² = 0.91), with a slope close to unity (0.96) and a small intercept (1.7 wt%). The distribution is also fairly homoscedastic. We interpret this to mean that the method to estimate SiO₂ in NASGL should be robust and widely applicable.

A second validation approach was tested whereby the spatial distribution obtained for the estimated SiO₂ in the C horizon of the NASGL samples was compared with measured SiO₂ in an independent dataset. The most extensive available dataset with SiO₂ concentrations in the USA to our knowledge is that of Shacklette and Boerngen (1984), albeit at a much lower spatial density than the NASGL project. They reported inorganic chemical analyses of soil and other regolith collected across the conterminous USA mostly during the 1960s and 1970s (no mineralogy is reported). The target medium was the subsoil at ∼20 cm below surface to avoid any surface contamination; this depth commonly is within the range of a soil's B horizon, a zone of element accumulation (Boerngen and Shacklette 1981). Although more than 1300 sites were sampled in total, only 407 were analysed for Si (by emission spectrography of the <2 mm grainsize fraction) in ‘Phase two’ (∼1969–1975) of the project (Shacklette and Boerngen 1984). The Si (wt%) concentrations were converted to SiO₂ (wt%) before use here.

Firstly the empirical distribution functions of the ‘Shacklette and Boerngen’ and NASGL datasets were compared using a Kolmogorov–Smirnov (K-S) test of distribution similarity (Kolmogorov 1933) (Fig. 8). This non-parametric test quantifies the distance D separating an empirical distribution function from the cumulative distribution function of a reference distribution and an n-scaled critical value (CV). The null hypothesis (H₀) being tested is that the two populations are indistinguishable and is quantified at a given probability p. The K-S test applied to the ‘Shacklette and Boerngen’ measured SiO₂ concentrations and NASGL C horizon-estimated SiO₂ concentrations yields D = 0.0557, which is smaller than CV = 0.0703, therefore justifying accepting the H₀ at p < 0.05 (AAT Bioquest 2023).

Secondly a more rigorous test than comparing the general distributions, namely checking spatial consistency of SiO₂ values, was applied. It has to be cautioned that (1) we are comparing two different soil horizons, B horizon in the ‘Shacklette and Boerngen’ dataset v. C horizon in the NAGSL dataset; and (2) soil heterogeneity is present at all scales (e.g. Pedersen et al. 2015), implying that comparing sample pairs distant by even a few metres can give substantially different concentration values. Nonetheless we extracted the closest NASGL site to each ‘Shacklette and Boerngen’ site and filtered out those pairs where the distance between the two was 0.04 degrees of latitude/longitude (∼4 km) or greater. The scatterplot and linear regression for the resulting subset are shown in Figure 9. The regression is surprisingly strong (R² = 0.79) and with a slope close to unity (0.96). The relative standard deviation (RSD) on these pairs of samples (5%) is only marginally greater than the RSD on field duplicates obtained in the NGSA for SiO₂ (4%; table 1 of Caritat and Cooper 2011b), which is remarkable given the spatial distance between these ‘Shacklette and Boerngen’ and NASGL sites (∼1 to 4 km) and their different sample media (B v. C horizons).

Overall, the above validation assessments provide confidence that the SiO₂ estimates for the NASGL dataset, and by inference the recursive inversion methodology in general, compare favourably with ground-based measurements (‘Shacklette and Boerngen’ dataset) and mineralogical/geochemical data from a large independent study area (SA–Qld–NT dataset).

The current method of estimating SiO₂ when missing from a dataset can be applied to other situations, e.g. where ICP analysis has been used and Si not determined. However, the methodology currently relies on mineralogical data being available for the same samples. Another limitation is that total or near-total geochemical analytical methods have to be used to ensure internal consistency with the mineralogical data, as such weak or partial digestion/leach techniques for sample preparation do not lend themselves directly to being compared with bulk mineralogy. Despite these limitations, there are many cases where (near-)total geochemical analysis and mineralogy have been determined, for instance in industry and government datasets.

In a complementary approach in progress, we are developing a machine-learning approach using linear regression and random forest algorithms to estimate SiO₂ where it is missing, based on geochemical information, mineralogical information, and both geochemical and mineralogical information. This method will be tested on various datasets, including the NASGL and SA–Qld–NT datasets, to ensure its universal applicability and will be reported separately (Grunsky et al. 2018)

The original geochemical and mineralogical data for soils of the conterminous USA (A and C horizon datasets) were downloaded from https://mrdata.usgs.gov/ds-801/. The ‘Shacklette and Boerngen’ dataset was downloaded from https://mrdata.usgs.gov/ussoils/. A worked example for the five selected samples of Figure 5 is available as a Microsoft Excel™ spreadsheet (NALG_Ch_oxides_with_estimated_SiO₂_LOI_worked example.xlsx) here: https://doi.org/10.5281/zenodo.8191287. The new datasets including sample identification, coordinates, converted major oxide concentrations, and the concentration estimates for SiO₂ and LOI in wt% for the A and C horizon datasets from the NASGL project are available as comma-separated value files (NALG_Ah_oxides_with_estimated_SiO₂_LOI.csv and NALG_Ch_oxides_with_estimated_SiO₂_LOI.csv) here: https://doi.org/10.5281/zenodo.8191287.

We provide a novel method for estimating the concentrations of silica (SiO₂ wt%) and loss-on-ignition (LOI wt%) in the NASGL project datasets. These datasets include comprehensive elemental and mineralogical compositions, determined mostly by four-acid digestion ICP-AES or ICP-MS, depending on the element, and Rietveld refinement XRD, respectively. Unfortunately, neither Si/SiO₂ nor LOI are quantified, both of which are significant components of most soils. Our estimation method combines the precision of the ICP determinations with the completeness of the XRD data. As the NASGL samples contain up to 95 wt% amorphous material of unknown geochemical or mineralogical composition, it is not possible to directly calculate SiO₂ or LOI contents from mineralogy alone. However, a recursive inversion approach, i.e. calculating geochemistry from mineralogy, can be invoked to calculate minimum SiO₂, H₂O and CO₂ concentrations. Thus, we inverted an estimate for SiO₂ by adding up the SiO₂ contributions from all Si-bearing minerals (silicates). This ‘normative’ SiO₂ represents a minimum estimation of the total SiO₂ in each sample. Similarly, we inverted estimates for H₂O by adding up the H₂O contributions from all O–H-bearing minerals (hydrates), and for CO₂ by adding up the CO₂ contributions from all C-bearing minerals (carbonates). Combining the latter two components gives a minimum estimate for LOI. Thus, 100 wt% – (all major oxides from ICP + TEs from ICP + ‘normative’ H₂O + ‘normative’ CO₂) yields a maximum estimate of the total SiO₂ in each sample. The final or ‘consensus’ SiO₂ estimate is then calculated as the average between the two aforementioned estimates, trimmed as necessary to yield a total composition (all major oxides from ICP + estimated SiO₂ + TEs from ICP + ‘normative’ H₂O + ‘normative’ CO₂) of no more than 100 wt%. For most samples, the above sum falls below 100 wt% and the difference is taken to represent LOI not otherwise accounted for in the quantified hydrate and carbonate minerals. The source of this LOI contribution likely includes H₂O and CO₂ in the amorphous phase as well as other volatile components present in soil. We examine the statistical distributions of the SiO₂ and LOI estimates and validate the technique against a separate dataset from Australia where XRF, ICP and XRD data on the same samples exist. The correlation between predicted and observed SiO₂ is deemed strong (R² = 0.91). Further, we compared the estimated NASGL C horizon SiO₂ estimates with an independent dataset covering the conterminous USA, the ‘Shacklette and Boerngen’ dataset. The distributions of these two datasets are shown by a Kolmogorov–Smirnov test not to be statistically different. Spatially we demonstrate that the closest NAGSL sites and ‘Shacklette and Boerngen’ sites have highly correlated SiO₂ concentrations (R² = 0.79). Together, these validation assessments give us the confidence to recommend the approach of combining geochemical and mineralogical datasets to estimate missing SiO₂ and LOI in datasets elsewhere. However, as each situation is different, any estimation results ideally should be ground-truthed.

The estimation method and Australian validation study were conducted as part of Geoscience Australia's Exploring for the Future (EFTF; https://www.eftf.ga.gov.au/) programme, which provides precompetitive information to inform decision-making by government, community and industry on the sustainable development of Australia's mineral, energy and groundwater resources. By gathering, analysing and interpreting new and existing precompetitive geoscience data and knowledge, the EFTF is building a national picture of Australia's geology and resource potential. This leads to a strong economy, resilient society and sustainable environment for the benefit of all Australians. This includes supporting Australia's transition to net zero emissions, strong, sustainable resources and agriculture sectors, and economic opportunities and social benefits for Australia's regional and remote communities. The EFTF programme, which commenced in 2016, is an eight year, $225 m investment by the Australian Government. The authors appreciate comments on a draft version of this manuscript by Laurel G. Woodruff (United States Geological Survey) and Clemens Reimann (Geological Survey of Norway, retired), as well as internal reviews by Philip Main, Tara Webster and Anthony Schofield (Geoscience Australia). We thank journal referees Alecos Demetriades and Ryan Noble, as well as Editor-in-Chief Scott Wood, for their detailed and constructive reviews of the paper. PdC publishes with permission from the Chief Executive Officer, Geoscience Australia.

PdeC: conceptualization (lead), data curation (lead), formal analysis (lead), investigation (lead), methodology (lead), validation (lead), visualization (lead), writing – original draft (lead), writing – review & editing (lead); ECG: data curation (supporting), formal analysis (supporting), methodology (supporting), writing – review & editing (supporting); DBS: data curation (supporting), formal analysis (supporting), methodology (supporting), writing – review & editing (supporting)

This study was funded as part of Geoscience Australia's Exploring for the Future (EFTF; https://www.eftf.ga.gov.au/) programme.

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The datasets generated during and/or analysed during the current study are available in the Zenodo repository, https://doi.org/10.5281/zenodo.8191287