Table 1 Comparison of training, validation, and unseen material dataset performance for the DNN and CDNN methods.

From: Deep learning-based inverse design of microstructured materials for optical optimization and thermal radiation control

Method (% Data)

Channels

Train

Validation

Ti

Al2O3

Library

DNN (100%)

N/A

0.0032

0.0033

0.0175

0.0141

0.0279

CDNN (20%)

n, k, 0

0.0031

0.0036

0.0171

0.0225

0.0373

CDNN (20%)

n, k, εim

0.0019

0.0032

0.0266

0.0176

0.0378

CDNN (20%)

n, k, εreal

0.0029

0.0031

0.0233

0.0184

0.0346

CDNN (20%)

n, k, λ

0.0025

0.0031

0.0185

0.0163

0.0314

  1. To greatly speed up the training/validation time, we only use 20% of the available data to train the CDNN (~ 710,000 images of 3.55 million).
Back to article page