FWBP: A New Generative Modeling Scheme

Generative models are often used to generate differentiated samples for images, audio and video, focusing on the diversity of the generated samples. For materials, molecules and other fields, it is more important to generate samples with targeted properties. More attention should be paid to the accuracy of the generated samples. In the spirit of exploring conditional generative neural networks, this manuscript proposes a new generative modeling scheme-FWBP, i.e., Fixed Weight Backpropagation. FWBP can effectively utilize trained prediction models and generate samples with targeted attributes by fixing the weights in prediction models and then backpropagate the properties. High-entropy alloys (HEAs) are alloys composed of multiple elements. The property design of HEAs is of great significance for practical applications. Firstly, a set of function is used to illustrate the general idea and effectiveness of FWBP. Then, FWBP is applied to the generation of HEAs with targeted properties. The accuracy of FWBP is verified in four datasets with various networks. FWBP's efficiency is also proved in the field of image generation.