|
|
|
Figure 4
Case study: first-pass fit of the integrated spectrum from a plant sample. Both methods start from random parameter and amplitude initialization (within reasonable ranges) with no manual adjustments and use the same target element list. Top: NLopt result (finite-difference gradients with a traditional nonlinear optimizer) stalls far from the data (R2 = −0.050). Bottom: MapsTorch (automatic differentiation with Adam optimizer) reaches a substantially better first-pass fit (R2 = 0.872). Spectra are shown as log10(counts+1) versus energy (keV). |
Open
access
access
journal menu![[Figure 4]](gy5077fig4.jpg)



