Abstract
The growth of non-polar GaN on Si substrates is a grand challenge in developing light-emitting diodes, where AlN layers serve as buffers to mitigate the lattice mismatch between GaN and Si in practice. Even though, the primary difficulty aries from the predominant growth of polar AlN(0001) layers. Here we demonstrate that stepped Si(320) can establish a high-quality interface with the semi-polar AlN(22 4¯1 ) as identified through machine-learning-based structure predictions that explored millions of potential interface configurations. This interface exhibits atomic-matching with low interface energy, and importantly, features reduced polarization (0.20 C/m(2)) along with superior interfacial thermal conductance (0.47 GWm(-2)K(-1)).