Researchers at MIT have developed a new approach that may enable engineers to determine what’s happening inside a material simply by analyzing its surface properties. By utilizing deep learning, the team was able to compare a vast dataset of simulated information on the external force fields of materials and their corresponding internal structure, creating a system that could make accurate predictions of the interior from surface data. Markus Buehler, a professor of civil and environmental engineering, and doctoral student Zhenze Yang published their findings in the journal Advanced Materials. According to Buehler, it is a widespread problem in engineering to want to know what is inside a material, such as a piece of a car or airplane. Although X-rays and other methods may be used, they can be expensive and require bulky equipment. Therefore, the team aimed to develop an AI algorithm that could examine the surface of a material, which can be easily seen, photographed, or measured, to determine what is happening inside, such as cracks, damages, stresses, or microstructure details.
The use of this technique is not limited to engineering, as it can also be applied to biological tissues, according to Markus Buehler. The aim is to determine whether there is disease, growth, or changes in the tissue in a non-invasive manner. To address the complexity of the problem, methods were created to give all the possible options that may result in a particular surface scenario, since many problems have multiple solutions. The AI model was trained using vast amounts of data on surface measurements and the associated interior properties of various materials, including those with different components and properties. The technique can work even for materials whose complexity is not fully understood, as long as enough data is collected to train the model. The method can be broadly applied to different engineering disciplines such as fluid dynamics and other types, and can also be used to determine a variety of properties, not just stress and strain. Buehler cited the example of the magnetic fields inside a fusion reactor, indicating the universality of the approach.
The idea for the approach came to Yang when he encountered blurred imagery in his study of a material and pondered how missing data in the blurred areas could be recovered. This led him to the inverse problem of missing information recovery, which is a common challenge in material science. The iterative process of developing the technique involved comparing preliminary predictions made by the model with actual data and refining it further. Testing against materials with well-known properties showed that the predictions of the new method were accurate. The training data comprised various measurements of surface properties and even simulated data, allowing engineers to make predictions even for materials with unknown characteristics.
The approach has many potential applications. For example, it could be used in airplane inspections, where currently only a few representative areas are tested with expensive methods. This new approach provides a less expensive way of collecting data and making predictions, guiding engineers as to where to look further and potentially allowing more efficient use of more expensive equipment. The researchers expect that this method will initially be used in laboratory settings, such as for testing materials used in soft robotics, where measurements of the internal properties of hydrogels, proteins, and biomaterials are difficult to obtain. Researchers can use this technique to design better composites and grippers. The method will be freely available for use via GitHub.