Answers are grounded in this paper and cite their sources.
We introduce a data-efficient neural-vision method to estimate the crystallographic orientation of a rotating substrate from reflection high-energy electron diffraction (RHEED) patterns during molecular beam epitaxy (MBE) growth. We combine single-frame azimuth predictions of a neural network with the known rotation kinematics in a physics-aware postprocessing step. This step detects the rotation direction and recovers a robust reference azimuth, replacing manual frame-by-frame inspection. Trained on RHEED patterns from only 15 independently grown CdTe samples, the model infers the crystallographic directions to within the 0.5° target for 95% of rotations. Under leave-one-structure-out evaluation on all 20 samples, the mean error is 0.18°. High accuracy stems from the postprocessing rather than from the network architecture. A 2D ResNet reaches the accuracy of a 3D ResNet benchmark trained on the same number of samples at approximately one tenth of the training cost. Our work delivers (i) a physics-constrained estimation procedure for rotating-substrate RHEED, applicable to any azimuth regressor that is unbiased over a rotation, (ii) an analysis of why few independently grown samples suffice, and (iii) a trained system validated for online orientation estimation during CdTe growth.
No takes yet. Share an insight, caveat, or question.
Turowski et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: