A data screening approach to confirming a target mineral is chlorite using EPMA and LA‐ICPMS data

Applying machine learning techniques to large datasets of in situ analyses has been proven to be a powerful tool in Earth Sciences. However, problems may arise when dealing with minerals such as chlorite, that exist as a solid solution rather than a single, stoichiometric ideal. It can be difficult to determine whether the variations in major element concentrations are due to compositional difference in the mineral of interest or due to sampling of the surrounding mineral phases in addition to the mineral of interest during the analyses. If the latter, interpretations of the results would be complicated, misled or even spurious. Here we present a method to identify chlorite based on the major and minor element content, from both LA‐ICPMS and EPMA data. Further we present a dataset of 3,317 analyses of chlorite and have shown that 7.4% of these analyses include significant quantities of non‐chlorite material.


| INTRODUCTION
The trace element content of minerals, as determined by electron probe microanalyser (EPMA) and laser ablation inductively coupled plasma mass spectrometry (LA-ICPMS), has been widely used to better understand the evolution of different geologic systems, including depth of formation of igneous rocks (Tang et al., 2021), understanding past ocean chemistry (Large et al., 2014) and identifying characteristics of mineral deposits (Cook et al., 2016).More recently, the use of machine learning to extract additional data from mineral ecology datasets has become more common.These have been used for several purposes including identifying how carbon bearing minerals evolved through time, with implications in understanding the evolution of the atmosphere and life it supported on Earth (Morrison et al., 2020) and identifying the ecological conditions present during poorly understood periods of Earth history (e.g.Hadean; Morrison et al., 2018).This has led to the combination of these techniques in investigating individual minerals major, minor and trace element chemistry.Thus far, a key focus of this research has been on changes in mineral chemistry around or between ore deposits.Gregory et al. (2019) used a combination of pyrite trace element chemistry and random forest algorithm to identify the type of ore deposit an unknown sample is from based on its trace element chemistry.O'Brien et al. (2015) applied the same technique (random forests) to gahnite trace element chemistry and developed a vector towards mineralisation at the Broken Hill deposit, Australia.While these techniques have shown promise, issues may arise especially when combining different datasets.In situ analyses may include partial analysis of intergrown matrix minerals that are not part of the mineral of interest.This can lead to spurious results, depending on the size of the minerals of interest and the amount and chemistry of the intergrown minerals.In the case of pyrite, which varies little from the stoichiometric ideal, these spurious analyses can be relatively easily avoided, or even corrected (Stepanov et al., 2020).However, for minerals like chlorite that exist as a solid solution and are often intergrown with inclusions of phyllosilicate minerals and other non-chlorite material, it can be rather difficult to avoid these inclusions during regular data collections.Chlorite chemistry has become increasingly of interest for mineral exploration (Cooke et al., 2020;Wilkinson et al., 2015), especially for porphyry systems.The purpose of this contribution is to provide an easy to use tool to identify if a given in situ chlorite analysis has a significant amount of additional material or is not chlorite at all.
The filtering process consists of three major steps.The first checks to see if the total weight percent of all the major oxides, assuming ideal chlorite concentration for the hydrous components, is between 98.5% and 101.5% inclusively.If outside the acceptable range, for EPMA data, it may indicate that an important element was not determined, which would lead to the rejection of the data being used as chlorite, and ideally, the analysis should be repeated with the missing element included.For LA-ICPMS data, it is also possible that the concentration(s) of internal standard was used during data reduction.If this is the case, the dataset can be remedied by renormalizing to 100% abundance, and the resulting data can be checked against steps two and three of this classifier.
Compositionally chlorite can be confused with amphibole, mica or clay.However, amphiboles should have 2~3 wt% H 2 O on average while micas typically have 4~5 wt% H 2 O. Chlorite can have much more water content.Thus, the second step checks to see if the H 2 O% is between 9% and 18% inclusively.The H 2 O% is calculated by subtracting the oxide weight percent of the summed cations from the total weight %, assuming an idealized chlorite formula.
The third and last step is through site occupancy based on chlorite's general structural formula (A 5-6 T 4 Z 18 ).The test checks to see if the sum of cations can fit the idealized cation abundance for the 'A' (between 4.5 and 6.5), and if the sum of cations that can fit into the idealized cation abundance for the 'T' (between 3.5 and 4.5).In the case of Al, which can substitute into both the A and T sites the Al was used to fill the T site first (combined with Si) and the remainder was used in the A site.If all conditions are met and satisfied, the analysis is classified as a chlorite.For the purposes of this classifier, Fe is assumed to be held as FeO.
The chlorite classifier is available in the electronic Appendices S1 and S2.Instructions on how to use it, and how the data presented here were entered are as follows: LA-ICP-MS data filter • The table has columns for Al, Ca, Cl, Cr, F, Fe, K, Mg, Mn, Na, Si and Ti.In order for the filter to process the data, there must be data for at least Al, Fe, Mg, and Si.• Paste the chlorite major element data in ppm into the table.• Leave cells blank for elements that were not determined.
• Make sure all of your pasted data are numbers.If an element concentration is below the detection limit, or if 'NA' or 'bdl' is reported, leave the cell blank.(plus F and Cl) in weight percent into the table.• Leave cells blank for components that were not determined.
• Make sure all of your pasted data are numbers.If a concentration is below the detection limit, or if 'NA' or 'bdl' is reported, leave the cell blank.• The last column in the excel sheet, 'Chlorite?',will output 'yes' if the analysis classifies as chlorite, and 'no' if it does not.

| RESULTS AND DISCUSSION
After applying the classifier, of the 3,317 analyses, 246 analyses (or 7.4%) were deemed to be 'not chlorite'.In Figure 1, Ti versus Sr, Ti versus Pb, Mg versus Sr and V versus Ni are plotted.These elements were chosen as they are all significant for vectoring towards porphyry mineralisation (Wilkinson et al., 2015) and are likely to be employed or modelled when investigating chlorite chemistry.However, it should be noted that not all these analyses are from porphyry systems and other investigators may be interested in elements other than those chosen here.However, these plots demonstrate how trace element results would appear with large components of non-chlorite minerals when plotting with other chlorite data.Figure 1 shows that the nonchlorite data and the chlorite data tend to plot within the same general areas of the x-y plots in most cases.However, not doing a check if the analyses were indeed predominantly of chlorite can cause issues in interpretation.The 'nonchlorite' analyses tend to be more enriched in Ni.Because V/Ni ratios are lower away from mineralisation (Wilkinson et al., 2015) these analyses could be falsely interpreted as  being from a porphyry system and distal to mineralisation when the reason for elevated Ni may just be what other components were ablated during analysis.This shows that it is important to screen the data based on classifiers built from first principles rather than inspection of element plots after data is already being used for interpretations.

| CONCLUSIONS
Screening methods based on first principles are an important aspect of robust analysis of data obtained from in situ analyses, especially with minerals that occur as solid solutions.This paper shows how literature data that have been peer reviewed can still have some analyses which do not have the correct element components to be the suspected mineral and thus must contain a significant quantity of other components.This can lead to errors interpreting the end dataset.Thus, both for single location studies and for large data compilations a screening tool, such as the one provided in this contribution, should be applied to the data prior to interpretation.

F
I G U R E 1 Trace element plots of chlorite elemental content from the chlorite chemistry database (Appendix S1).The blue spots are determined by the classifier to be not chlorite and the orange spots are determined by the classifier to be chlorite.(a) Ti versus Sr; (b) Ti versus Pb; (c) Mg versus Sr and (d) V versus Ni. al.
• The last column in the excel sheet, 'Chlorite?',will output 'yes' if the analysis classifies as chlorite, and 'no' if it does not.EPMA data filter • The table has columns for Al 2 O 3 , CaO, Cl, Cr 2 O 3 , F, FeO, Fe 2 O 3 , H 2 O, K 2 O, MgO, MnO, Na 2 O, SiO 2 and TiO 2 .In order for the filter to properly function, there must be data for at least Al 2 O 3 , FeO/Fe 2 O 3 , H 2 O, MgO and SiO 2 .• Paste your chlorite major element oxides