Left: Machine Learning based on artificial neural network constitutive laws to predict atomic stress. Right: Quantifying local stress levels at grain boundaries using atomic coordinate information. Credit: University of Illinois Dept. Aerospace Engineering
When designing spacecraft and aircraft, it is important to know how much stress a material can withstand without cracking. Machine learning was used by University of Illinois Urbana-Champaign to predict copper stress at the atomic level for the first time.
Huck Beng Chew, his doctoral student Yue Cui and others believe that materials such as copper are quite different at very small scales.
Chew stated that metals are often polycrystalline because they have many grains. Each grain has a single crystal structure, where all the atoms are neatly arranged and very well-ordered. The atomic structure at the border where the grains meet can be complex and subject to very high stresses.
These grain boundary stresses are responsible to the fatigue and fracture properties of the metal. However, detailed atomic-scale stress measurements have been limited to molecular dynamics simulation models. Data-driven methods based on machine learning allow the study of grain boundary stresses in real metal specimens that have been imaged using electron microscopy.
Cui stated that they used molecular dynamics simulations to simulate copper grain boundaries in order to train their machine learning algorithm to recognize the arrangement of atoms along these boundaries and identify patterns within the stress distributions within various grain boundary structures.
The algorithm eventually was able predict grain boundary stresses accurately from both experimental and simulation data.
Cui stated that they had tested the machine learning algorithm's accuracy with many different grain boundary structures, until they were certain the approach was reliable.
Cui stated that they found the task more difficult than they expected and had to incorporate physics-based constraints into their algorithms in order to make accurate predictions with very limited data.
Chew stated that "if you train the machine-learning algorithm on grain boundaries, you'll get extremely high accuracy with stress predictions of those same boundaries." But, the most important question is: can the algorithm then predict stress at a new boundary it hasn't seen before?
Chew stated that the answer was yes and it is very true.
Machine learning is a breakthrough in the field of mechanics and materials. It allows us to quickly and autonomously make predictions using data. Chew stated that this is a major improvement over the use of complex and specific physics-based models to predict failures.
This is the first step in designing aerospace materials for extreme environments.
Chew stated that scientists will be able to determine quantitative descriptors for grain boundaries in order to engineer stronger, more resistant to heat and corrosion.
Cui stressed the fact that their algorithm is general enough to be used for quantifying the atomic-scale stresses that govern fracture and failure in other materials systems.
Acta Materialia published Yue Cui's study "Machine-Learning Predictions of Atomistic Stress Along Grain Boundaries", by Huck Beng Chew.
Learn more Machine learning predicts the behavior of stainless steel at microstructural levels
Further information: Y. Cui et al., Machine-Learning Predictions of Atomistic Stress Along Grain Boundaries. Acta Materialia (2021). Machine-Learning Predictions of Atomistic Stress Along Grain Boundaries by Y. Cui et. al., (2021). DOI: 10.1016/j.actamat.2021.117387
