Part 4/6:

The paper delves into the use of graph neural networks (GNNs) in the invention process. These networks are designed to understand materials at a detailed level, allowing them to generate new material structures that meet specific requirements. The three-step training process (pre-training, fine-tuning, and reinforcement) enables the AI to continuously improve its predictions.

The GNN-based tool significantly improves the materials discovery process by quickly generating suggestions that are more likely to succeed, allowing scientists to focus on evaluating promising AI-generated ideas rather than generating them from scratch. This is evident in the data, which shows a clear upward trend in the number of new materials discovered and patent filings after the integration of the AI tool.

The Human Impact of AI in Science

[...]