Researchers have developed a machine-learning technique that detects internal structures, voids, and fractures within a material based on surface data. 

"Anything that could give rise to smarter-than-human intelligence - in the form of Artificial Intelligence, brain-computer interfaces, or neuroscience-based human intelligence enhancement - wins hands down beyond contest as doing the most to change the world. Nothing else is even in the same league." - Eliezer Yudkowsky.

Material engineering problem

Materials engineering problems are often hard to solve because there needs to be more information, like inverse problems with only border data or design issues with a simple goal but an ample search space. Several deep learning (DL) architectures forecast missing mechanical information given limited known data in part of the domain and characterize composite geometries using recovered mechanical fields for 2D and 3D complex microstructures.

2D microstructures

In 2D, a conditional generative adversarial network (GAN) is used to complete partially masked field maps and predict the composite geometry with convolutional models with high accuracy and generality. It is done by making accurate predictions on field data with mixed stress/strain components, hierarchical geometries, different material properties, different types of microstructures, and ill-posed inverse problems. 

3D microstructures

In 3D, a Transformer-based design is used to predict complete 3D mechanical fields from snapshots of the input fields. The model works well no matter how complicated the microstructure is, and it can recover the whole bulk area from a single image of the surface field. Furthermore, the internal structure can be described using only edge measurements. The entire framework makes it easy to analyze and build things when you don't have all the information you need. They also make it easy to go back and forth from properties to material structures.

Proposed work

In this work, the researchers provide AI-based frameworks for predicting whole strain and stress fields from partial field data and achieving inverse translation from mechanical areas to composite microstructures. Numerical simulations like FEA are forward solvers used to determine structure-property relationships. However, to solve inverse problems such as geometry identification, additional optimizers such as optimization algorithms are frequently required in conjunction with the forward solver to iteratively search for the best solutions. 

Evaluation

To solve this limitation, the researchers use a combination of deep learning architectures to directly link an incomplete strain or stress field to the heterogeneous structure in 2D and 3D circumstances. In 2D, they use a conditional generative adversarial network (GAN) technique and a convolutional model to recover the masked region in a field map and then identify the composite structures from recovered mechanical fields. 

The field-completion method is then validated in a variety of scenarios, including: 

• When multiple strain/stress components are intermingled in a dataset. 
• When out-of-distribution microstructures with variable material forms and grid sizes must be stretched. 
• When the mechanical properties of the constituent materials entail plasticity.
• When given continuous microstructures rather than discrete blocks, such as Cahn-Hilliard patterns. 

Conclusion

The methods suggested in this work can also be used to solve reversed design problems. The proposed methods can deal with more complex material responses compared to designing a single function. These responses can be fed into deep learning models in the same way as partial field information to get the corresponding material structures. For example, by combining our method with optimization algorithms, we can iteratively look for suitable candidates with boundary responses most like the goal. Furthermore, experimental methods like "additive manufacturing" can be used to make actual samples of designs suggested by a computer framework.

Image source: Unsplash