ScGAN: a generative adversarial network to predict hypothetical superconductors.

Area of Science:

• Materials Science
• Condensed Matter Physics
• Artificial Intelligence

Background:

• High-temperature superconductors (HTSs) were discovered over 30 years ago but lack mechanistic understanding and systematic search methods.
• The discovery of new superconductors is crucial for technological advancements and fundamental scientific insight.

Purpose of the Study:

• To develop an efficient computational method for predicting novel superconducting materials.
• To accelerate the search for high-temperature superconductors (HTSs) using artificial intelligence.

Main Methods:

• A generative adversarial network (GAN), named ScGAN, was developed for superconductor prediction.
• ScGAN was trained on the Open Quantum Materials Database and fine-tuned using the SuperCon database.
• A classification model was used to validate the predicted superconducting candidates.

Main Results:

• ScGAN predicted superconducting candidates with a 70% success rate when validated by a classification model.
• This represents a 23-fold increase in discovery rate compared to traditional manual search methods.
• Over 99% of ScGAN predictions were novel materials, including promising high-temperature superconductor candidates.

Conclusions:

• ScGAN offers a novel and efficient approach to discovering new superconductors.
• The method has the potential to yield materials for technological applications and advance the understanding of high-temperature superconductivity.