MIT researchers have developed an artificial intelligence model capable of detecting and quantifying atomic-scale defects in materials without causing damage. While defects in biology are often detrimental, materials science utilizes them to enhance properties such as strength and electrical conductivity in products ranging from semiconductors to solar cells. However, accurately measuring these defects in finished goods has long been a challenge, as traditional methods often require destructive sampling or only identify limited defect types. The new AI system addresses this by analyzing data from noninvasive neutron-scattering techniques. The model, trained on a database of 2,000 semiconductor materials, can simultaneously detect and quantify up to six distinct types of point defects, a task described as impossible for conventional techniques alone. Lead author Mouyang Cheng, a PhD candidate at MIT, emphasized that existing methods lack the universal and quantitative precision required to characterize multiple defects without destroying the sample. The study, published today in the journal Matter, utilized a machine learning architecture similar to the multihead attention mechanism used in large language models. This approach allows the model to extract subtle differences in vibrational frequencies between defective and non-defective materials to predict the specific dopants used and their concentrations. The team verified the model's accuracy on an alloy common in electronics and a superconductor material, successfully identifying defect concentrations as low as 0.2 percent even when multiple defects were present simultaneously. Senior author Mingda Li compared current defect detection methods to the parable of blind men describing an elephant, where each technique only perceives a single part of the whole. This AI model provides a comprehensive view, offering a ground truth that enables engineers to better harness defects for improved performance. The research team includes postdoc Chu-Liang Fu, undergraduate researcher Bowen Yu, and collaborators from Oak Ridge National Laboratory. Despite the success, the researchers acknowledge limitations regarding immediate industrial deployment. The neutron-scattering technique used to train the model is complex and not yet widely available for routine quality control in manufacturing facilities. However, the team is already working on adapting the AI approach to Raman spectroscopy, a more accessible technique that measures light scattering. Industry partners have expressed strong interest in this potential integration. The researchers plan to expand their work beyond point defects to analyze larger structural features like grains and dislocations. By leveraging AI's pattern recognition capabilities to interpret complex signals that are indistinguishable to the human eye, this study establishes a new paradigm in defect science. The findings suggest that AI can transform the management of defects from a guessing game into a precise science, optimizing material performance while mitigating the risks of unintended properties. This work was supported by the Department of Energy and the National Science Foundation.