Idea: Generating Exploration Targets for Industrially Niche Elements

This is an idea I had a few years back which I did a fair bit of reconnaissance work into and never really followed up on it. This article will be something of a structured notes dump with revisionist notes and I’ll keep it pretty short. As far as I’m aware nobody has written a paper on this exact idea (which was my origin plan — before I got inevitably side tracked).

Great thing about Substack is I can dump these niche interest articles even if they remain in a semi-complete state.

Now the idea is that there are a lot of mineral elements that have previously had very little interest from the mineral exploration community but are now very much in vogue due to some niche military/semiconductor/other commercial purpose. As I write this America is dumping funding into anything remotely promising China-free critical minerals and so there is a lot of buzz about metals nobody previously cared about like antimony, gallium and scandium.

On occasion these metals have been assayed for (sometimes as pathfinders/academic interest/the assay package included them/recent commercial interest), but generally in geochemical databases they are excluded.

The idea is to use other minerals that are commonly assayed to predict the values for these elements of newfound interest.

Dataset

For the dataset I used Mineral Resource Tasmania’s (MRT) log of all X-ray Fluorescence (XRF) readings across the state (a very geologically diverse area). XRF readings are so inherently noisy that are only used for exploration purposes only (they are not allowed in place of a proper chemical assay), but are quick and often assay a large suite of elements at once.

My methodology was pretty simple. For every target element (I used Scandium, Gallium, Neodymium and Rubidium), columns (containing a parts per million assay value) were dropped if they were not present for at least 75% of the rows containing a valid assay for the target element. Rows were also removed if they did not have at least 50% non-empty values (excluding the target element).

Predictor elements used for each target element

The dataset was then split into three partitions. The training and test partitions were a 2:1 split of the dataset that contained a valid value for the target element and the prospect section of the dataset which meet the above conditions but lacked a valid reading for the target element. This is the part of the data that was used to flag potential new targets for critical minerals.

Number of samples in each partition of the dataset

Now for the same reasons I discuss in Part 1: Building a Transformer to Model Mineral Deposits, these samples are hardly random. There is inherent bias as somebody took the time and money to take a sample from a location and assay it. They also made (the presumably somewhat informed choice) not to assay for these minerals. This idea only works if they chose not to assay for them as they were of no commercial interest of the time. If they chose not to assay for them as they were confident of low levels of these elements than the train/test set and prospect set represent very different things.

I ended up making these slightly awkward visuals that show that the train/test dataset has rather different co-occurrences of assays than the prospect set. With the prospect set unsurprisingly having generally more missing values.

Co-occurrence of assays in train/test dataset

Co-occurrence of assays in prospect dataset

Models & Training

I chose to use a random forest due to the relatively small dataset sizes. They are also reasonably common in papers exploring the use of ML by domain rather than ML experts.

In hindsight I believe XGBoost would probably have been better, as I am yet to work on a dataset where random forests are not outperformed. If I ever come back to this looks like the sort of problem where a TabPFN model could really shine. I’ll still go through my random forest results, since I put a fair bit of effort into them at the time.

When I wrote this nobody had introduced me to Optuna yet, so I performed a old-fashioned grid search, using r-squared as my evaluation metric. (r-squared is imperfect here given the detection-limit censoring visible in the plots below)

Hyper-parameter Search Results (median result of 3-fold cross-validation)

Small datasets predictably lead to overfitting at higher max depth parameters across all elements. Regrettably I did not touch the regularisation parameters.

Predictions for models retrained using best hyper-parameters

Vertical lines in the actual values for Neodymium and Scandium are caused by detection limits of 1 PPM on XRF machines.

Being random forests I obviously had to make some pretty feature importance plots. (Importance done using Mean decrease in impurity).

Feature importance for the retrained models

In my notes I wrote in relation to this chart: “Barium was a very important predictive feature […] likely attributable to that it is assayed in a significant proportion of samples for all target elements […] However chromium concentration was another predictive feature that was disproportionately important despite its lower assay rate in the dataset. It is hypothesised that the significance of chromium as a predictive feature is due to its utility in distinguishing the mineralised ultramafic sequences that are prevalent in the western part of Tasmania”.

At the time I hadn’t been introduced to Shapley plots, so I sadly don’t have any of those handy, which I would use today if I were going to make these kind of arguments.

Prospect Set Evaluation

Summary of predictions made after applying models to prospect dataset

The predicted values on the prospect set were universally low, with the mean predicted value being both below the mean of the dataset and the crustal abundance. I would be very worried if we had found otherwise as geologists are very skilled professionals who make informed assay choices which somewhat undermines the assumption that assays are missed due to lack of commercial interest The other distinct possibility is that the model is failing to extrapolate to the prospect dataset which is (as previously discussed) fundamentally different to the train/test set.

In terms of evaluating geochemical anomalies I used multiple of crustal abundance, which is naive as it ignores the variance. By this (albeit crude) method the greatest anomaly predicted was Neodymium at 4.6x crustal abundance. Going back into the dataset this sample came from a roadside granite sample in 2006, that was reasonably enriched in the other rare earths (REEs) assayed (cerium at 5.6x crustal abundance and lanthanum at 4.7x). Given the almost invariable co-occurrence of REEs I would argue that this particular value is likely very reasonable but in a similar vein the model is not adding a lot of utility.

I noted that “Significantly, only six other samples have been assayed (by any means) for neodymium from granites in the surrounding thirty square kilometres, with a maximum assay of 57.1 ppm. This illustrates that this approach can potentially generate new geochemical anomalies. While the model does not predict economic enrichment of this sample, it is sufficiently enriched according to the model that the clays of this area could be considered prospective for secondary REE deposits”.

Conclusion

A bit more of a retrospective one. Feels good to have an idea I put quite a lot of effort into a while back in a somewhat presentable state.

Even though I don’t plan on actively working on this further, ideas and feedback are always welcome!Originally published at https://rocksml.substack.com.