Reinforcement learning-based knowledge graph reasoning for aluminum alloy applications

The study of the interrelations between the composition and properties of materials is greatly significant for accelerating the research and development of new materials. Knowledge graph reasoning provides effective support for exploring potential materials information by structuring graph data and creating linkages. However, the nature of materials data leads to sparse graph structures that differ from those typically encountered in benchmark datasets. To understand what the implications of this are on the performance of knowledge graph reasoning algorithms, we conducted an empirical study based on an aluminum alloy dataset. The task of reasoning can be formulated as a link prediction problem where both material compositions and properties correspond to entities in a knowledge graph, and our objective is to predict the potential relations among them. To overcome the limitation of existing algorithms concerning sparse knowledge graphs, we propose a novel knowledge-graph reasoning algorithm based on reinforcement learning, which reduces space exploration using multi-agents and solves the problem of sparse graphs through a new reward-shaping mechanism. The experimental results show that our method yielded performance gains of 53.9% for Hits@1, 43.0% for Hits@3, 41.6% for Hits@5, 37.3% for Hits@10, and 39.4% for the mean reciprocal rank with respect to the traditional reinforcement learning-based knowledge graph reasoning algorithm MINERVA. Additionally, we implemented a knowledge graph querying and reasoning system for the aluminum alloy dataset to visualize the process of reasoning for materials research.