Human–Machine Interface

摘要

The long-practiced traditional process of materials discovery can be improved. The evolution of this process is discussed from an historical perspective, up to the current paradigm of the machine-assisted discovery processes. Prior analogue attempts at expanding the scope and increasing the efficiency of materials discovery are shown, exemplified by parallel synthesis of combinatorial composition libraries and improved visualization of property–material–class by using a Circos layout, inspired by the Human Genome Project. Atomic and molecular simulations have evolved from structure–property calculation and crystal structure prediction towards statistically driven design processes with the aid of machine learning. Various successful examples of optimizing crystallography and compositions in super conduction, quantum charge interface transfer, and ultra-hard materials via density functional theory calculations are shown. However, the machine throughput of possible materials compositions and permutations far outpaces the experimentation validation of these designs, resulting in a need to further refine materials informatics and databases. Variational autoencoders and graphical neural networks are emerging as methods to self-generate shared material properties and material candidates. A different, high-level approach is to speed up experimental validation by linking simulation design and robotic manipulation to create a self-learning robot chemist. Yet, with all these current and future advances, the question of reproducing the intangible qualities of human creativity and serendipity remains.