Figure 3
From: Deep learning based analysis of microstructured materials for thermal radiation control
Neural-network predictions for two materials (Ti/Al2O3) that are not used in the in the training process. (a, b) The predicted optical properties vs. the FDTD computed properties, with and without 10 simulations included in training for alumina and titanium. Surface plot of the absolute error between prediction and simulation with no simulations included (c, d) and with simulations included in training. The wavelength is on the x-axis and the geometric information is visualized with the aspect ratio on the y-axis. Including 10 simulations (1% of the dataset) dramatically reduces the error in the alumina to a near zero value across all wavelengths and aspect ratios. For Ti, the resonance driven peaks in the low-aspect ratio structures are reduced and the error in all other sections becomes approximately zero.
