Constructing Structure-Property Maps using High-Throughput Experiments and Machine Learning

A fundamental challenge in materials science is to map structures to properties. Physics-based models are limited by theoretical idealizations, whereas data-driven models are limited by the size of available experimental datasets. Here, we demonstrate an approach to construct structure-property maps through a combination of a high-throughput experiment and machine learning. We print thousands of samples, the structure of each being defined by a pixel arrangement. We then measure the stress–strain curves of these samples by developing a high-throughput experiment. From these, we derived stiffness, fracture strain, fracture stress, and work of fracture, forming a dataset linking pixelated structures to properties. A convolutional neural network, initialized on ImageNet and fine-tuned by transfer learning, learned maps that generalize: out-of-sample test errors were 2.44%, 7.05%, 4.65%, and 12.16% for the four properties, despite the design space (∼10³⁵) far exceeding the fabricated set (∼10³). Embedding the learned map within an active learning closed loop to optimize fracture stress and fracture strain. Within a few iterations, it discovered a bar-like topology for fracture stress (94% improvement) and a zig-zag topology for fracture strain (282% improvement). These results demonstrate that high-throughput experiments combined with machine learning can construct reliable structure–property maps and effectively optimize fracture properties.