PMC

This figure shows an example amalgamation procedure, from the input compositions (Panel A) and the amalgamation matrix (Panel B) to the resultant amalgams (Panel C). The purpose of data-driven amalgamation is find the best amalgamation matrix for a given task. One could conceptualize amalgamation as a type of feed-forward network where each component has only one outgoing connection (Panel D). In this study, we search for an amalgamation matrix that maximizes an objective function; since this is equivalent to minimizing a loss, one could further conceptualize amalgams as a hidden layer in a (linearly activated) neural network. Although this suggests that we could use gradient descent, we choose to minimize the loss with a genetic algorithm because the amalgamation matrix is binary.

(A) Schematic illustration of the amalgamation reaction, converting monometallic seeds into intermetallic NCs. As-synthesized PdGa, PdIn, PdZn, AuGa₂, Cu₂Ga, Ni₂Ga_3, and Ag₃Ga IMC NCs are shown in TEM images (B to H) and STEM EDX mappings (B to E and G). Line profile scans for AuGa₂ and Ni₂Ga₃ NCs (E and G) highlight the compositional uniformity of the IMC NCs. a.u., arbitrary units.

 Representation of data amalgamation.

(A) XRD patterns measured from aliquots taken during the synthesis of Ag₃Ga NCs reveal the gradual conversion of Ag seeds to the intermetallic phase. Ga shows surface wetting and island growth behavior before intermetallic phase conversion as seen from a STEM EDX map of an aliquot taken after 10 s of reaction time (B) and as illustrated in amalgamation schematics (C). (D to F) Aliquots taken after 10 s, 2 min, and 10 min of reaction time are shown in TEM and STEM EDX images.