Origins of hole traps in hydrogenated nanocrystalline and amorphous silicon revealed through machine learning

Abstract

Genetic programming is used to identify the structural features most strongly associated with hole traps in hydrogenated nanocrystalline silicon with very low crystalline volume fraction. The genetic programming algorithm reveals that hole traps are most strongly associated with local structures within the amorphous region in which a single hydrogen atom is bound to two silicon atoms (bridge bonds), near fivefold coordinated silicon (floating bonds), or where there is a particularly dense cluster of many silicon atoms. Based on these results, we propose a mechanism by which deep hole traps associated with bridge bonds may contribute to the Staebler-Wronski effect.

Learning physical descriptors for materials science by compressed sensing

“…Furthermore, the CS analysis sets forth that the NSP is surely violated already for W = 2 if any two columns of D are a multiple of each other. In general, one can expect the solution of equation (8) to differ significantly from the solution of equation (7) if any two columns of D are highly correlated or, in general, if any of the columns of D lies close to the span of a small number of any other its columns.…”

“…The reason for the large coefficients for the former is that considering the 82-dimensional vectors whose components are the values of ( ) r B s and ( ) r B p for all the materials (in the same order), these vectors are almost parallel (their Pearson correlation coefficient is 0.996). 7  In order to understand this, let us have a look at both least square models: to be correlated to r p . This is true for all the atoms considered in this work, but it could well be that such strong correlation is not valid for all atoms.…”

“…In the optimal ℓ 0 model, the highly correlated vectors ( ) their coefficients in the linear expansion have almost the same magnitude, ∼1.3, and opposite7 Differently from H − L and IP − EA, which are expected, and are, correlated on physical grounds for all atoms, r s is not necessairly expected…”

“…It is typically considered a universal approach for learning (fitting) a complex relationship ( ) = y f x . Some, though few, machine-learning (ML) based works have been done in materials science, e.g., [2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17]. Most of them use kernel ridge regression (KRR) or Gaussian processes, where in both cases the fitted/learned property is expressed as a weighted sum over all or selected data points.…”

Hole-mobility-limiting atomic structures in hydrogenated amorphous silicon

“…We see that in both hydrogen content ensembles the application of stress reduces the incidence of moderately high HTD ID defects (HTD > 0.1 eV), but also that decreasing the hydrogen in the samples causes the expression of fewer IDs. We believe that this is due to the less-constrained Si-H bonds (hydrogen is usually bonded to only one, or at most two silicon atoms [34]) allowing a higher degree of bond rearrangement under the added potential of a hole being introduced into the system, and thus allowing stronger hole trapping. In the tensile stress regime, it is interesting to note that the decrease in occurrence of IDs and FBs does not explain the overall increased tails in the full hole trap distributions (seen in the generally higher values in the positive stress columns of Fig.…”