News

A machine-learning method developed at MIT detects internal structures, voids, and cracks inside

Issued at 2023-05-24



A machine-learning method developed at MIT detects internal structures, voids, and cracks inside a material based on data

Researchers at MIT, Cambridge, Mass., used deep learning to compare a vast amount of simulated data about the exterior force fields of materials with the matching internal structure, and from there, they developed a system that could reliably predict the interior from the surface data.

Currently, identifying the internal microstructure of a material is difficult because only material responses from indirect measurements at boundaries or interfaces are available. This makes inverse problems, such as failure analysis, nondestructive testing, and ultrasonic or X-ray characterization of materials, particularly a challenge.

Traditionally, numerical simulations like FEA as a forward solver have determined the relationship between structures and properties. Researchers provide AI-based frameworks to accomplish inverse translation from mechanical fields to composite microstructures, allowing for the prediction of whole strain and stress fields from partial knowledge of field data.

Multiple deep-learning architectures directly connect a 2D or 3D strain or stress field and the heterogeneous structure. In the 2D case, researchers recover the masked area in a field map using a convolutional model. They identify the composite structures using the mechanical fields retrieved from the masked regions. The field-completion method is tested in several real-world scenarios, such as:

  • When there are a variety of stress components in a dataset together
  • When applied to non-distributed microstructures with irregular forms and grid sizes, it is clear that more work is required
  • When mechanical behaviors of the components entail plasticity in the materials
  • When the microstructures provided are continuous rather than discrete blocks, such as Cahn-Hilliard patterns
  • When interior structures must be defined from indirect surface measurement, this model works well regardless of structural complexity and recovers the whole field from even a single surface field map.

For this research, scientists trained an AI model with massive data on outside metrics and the corresponding inside characteristics. The model included composites with a single material type as well as those made from a mix of components. The procedure was developed iteratively, with the model making early predictions that were then compared against real data on the substance in the issue. The resulting model was tested when a material’s inner workings were well known enough to calculate them, and its predictions agreed with the calculated values. The training data included images of the surfaces and measurements of their stresses, electric and magnetic fields, and other attributes. They ran data simulations in many situations using prior knowledge of a material’s atomic structure. The approach may provide an estimate good enough to point engineers in the right direction for future experiments, even when a novel material has numerous unknown features.

Markus Buehler, MIT professor in engineering and author of the study says the current practice of inspecting aircraft, which analyzes a small sample of a plane using costly procedures like X-rays, could be an example of the potential use of this approach.

Inverse issues that only have boundary data information and design jobs with simple goals but a huge search area are two other examples of situations where little information presents a challenge while solving materials engineering assignments.

To overcome these obstacles, several different deep learning architectures are used to both define the composite geometries from the recovered mechanical fields for 2D and 3D complex microstructures and to anticipate missing mechanical information given limited known data in part of the domain.

To predict the composite geometry with convolutional models for 2D field data with mixed stress/strain components, hierarchical geometries, different materials properties, and various types of microstructures, including ill-posed inverse problems, a conditional generative adversarial network (GAN) is used.

To make accurate predictions of whole 3D mechanical fields from 2D input field snapshots, a Transformer-based architecture is constructed in 3D. Regardless of the complexity of the microstructure, the model shows great performance and can recover the complete bulk field from a single surface field picture. This makes it possible to characterize the interior structure using just border data.

In addition to facilitating analysis and design with little data, the holistic frameworks also provide direct inverse translation from characteristics back to materials structures.

For more information:

Massachusetts Institute of Technology Department of Civil and Environmental Engineering

https://cee.mit.edu/




Source: ASM International



Related News