Researchers have found that LIBS spectroscopy models trained on compressed rock pellets produce inaccurate compositional readings when applied to fine-grained loose powders, raising concerns for planetary missions analyzing Martian soils.
A laboratory technique central to planetary geochemistry and used aboard NASA rovers to determine rock and soil composition on Mars produces significantly inaccurate results when analyzing loose, fine-grained materials, according to a recent study published in the journal Spectrochimica Acta Part B: Atomic Spectroscopy.1
In their study, the researchers, examining laser-induced breakdown spectroscopy (LIBS), found that predictive models trained on compressed rock pellets fail substantially when applied to loose mineral powders, particularly those with grain sizes below 38 micrometers.1
As a technique, LIBS works by firing a laser at a geological target, vaporizing a small amount of material and analyzing the resulting plasma's light emissions to determine elemental composition.1,2 The technique is a cornerstone of instruments like the
As part of the experimental procedure, the researchers tested three pure minerals (olivine, labradorite, and augite) across five grain-size fractions in both pressed pellet and loose powder form. Their findings revealed two compounding problems. First, minerals with compositions outside the range of a model's training data produced poor predictions even in pellet form.1 Second, and more critically, fine loose powders generated weak, noisy spectral signals that caused models to default toward the statistical average of their training data rather than reflecting actual sample chemistry.1
Physics can be used to explain why these issues are occurring. Fine powders couple poorly with the laser, generate lower-temperature plasma, and form deep, narrow ablation pits that may trap and distort the plasma signal.1 The result is spectra that the model was never trained to interpret.1
Another concern that the research team highlighted was that the standard uncertainty metrics returned by these models appeared low. To fix these issues, the researchers proposed that training data sets should be expanded to include spectra from loose powders if such materials are expected in the field.1
“Our results caution that predictions by multivariate PLS models are only as good as their training data,” the authors wrote in their study.1
