The virtual poster interface for the AGU fall meeting leaves something to be desired, so I thought it might be worth posting the contents of my poster here, where it is easy to access and share.

Audio recording walking through the poster.
Figure 1: Distribution of compositions in the laboratory spectral library. Orange bars represent the Training set, Green bars represent the test set.
Figure 2: (Left) All library points (grey dots) projected onto the first two PCs with selected sample types denoted by a range of colors and symbols (see legend). The dashed line represents the convex hull of all the library points. (Right) Same data without the SCCTs highlighted. Approximate groupings of samples are shown by shaded regions.
Figure 3: Test set predictions for the selected multivariate models. Perfect results would fall on the line.
ElementRMSEP wt.%Model
SiO26.1Average (GBR, PLS)
TiO20.3RF
Al2O31.8Average
FeOT3.1GBR
MgO1.1GBR
CaO1.3Blend RF+PLS
Na2O0.5Blend GBR+LASSO
K2O0.6LASSO
Summary of 3 m test set RMSEPs and selected models.
Figure 4: Histograms of Mars predictions through Sol 239 for each major element.
Figure 5: PCA plot of Mars spectra with convex hull of laboratory data superimposed (dashed line). Most Mars data plot within the space spanned by the lab data. Points that plot outside are indicated in red.
Figure 6: Local RMSEP indicates how accuracy is estimated to vary with predicted composition. Black points are the unsmoothed values calculated using the nearest 60 test set predictions. Blue curves show the result of smoothing and extrapolating.