Fig. 7. (a) The effect of a metal atom's presence in the structure, given as the mean and standard deviation (given as vertical lines for each mean) for each metal atom's SHAP values. The SHAP values are the output of the ML-MotEx algorithm, trained with fits to 10 000 random fragments of the [Fe4(FeW9O34)2]16− cluster,20 for each PDF in the in situ series. The colour of the metal–oxygen polyhedra in the investigated structure is reflected in the colour of the plots (split into three segments for clarification). (b) The resulting cluster structures with fits for the precursor, intermediate and product phases after −0.17, 1.08 and 1.42 min of heating. Green and red polyhedra are deemed significant contributors to the fit by the ML algorithm, having a trending mean SHAP value below −0.001 or above +0.001, respectively. Grey polyhedra are deemed insignificant by the ML algorithm, with a trending mean SHAP value in the range of −0.001 to 0.001. Below each of the structures, fits to the data are shown, which includes only significant metal atoms contributing positively to the Rwp value (i.e. green atoms in the structures). The structure visualizations were performed with VESTA.24.
