AMP_06_September_2021

A D V A N C E D M A T E R I A L S & P R O C E S S E S | S E P T E M B E R 2 0 2 1 6 MACHINE LEARNING | AI SMART, SPEEDY NEURAL NETWORKS Scientists from the Max Planck Institute for Iron Research (MPIE) and DeepMetis, an artificial intelligence (AI) firm, both in Germany, used deep neu- ral networks to calculate local stress in complex materials—up to 8300 times faster than a standard solver could. “Our latest work shows how all these calculations can be replaced by ma- chine learning. Instead of solving the equations directly, we developed a neu- ral network that can learn the physics and predict correct answers to complex and nonlinear mechanics questions simply by looking at a large set of data,” explains Jaber Rezaei Mianroodi, head of the MPIE’s group on computational sustainable metallurgy. After training with precomputed physical responses, the neural network is able to predict solutions to problems and configurations it never encountered during the teaching period. Much like an experienced engineer who develops an intuition for complex mechanical problems and is able to make ed- ucated guesses in seconds, the network appears to learn the un- derlying physics and predict solu- tions in microsec- onds. Unlike con- ventional solvers, which require a trial and error approach to solving non- linear problems, the trained machine learning solver is not iterative. “Our solver consumes orders of magnitude less computation time, opening up new possibilities for advanced materials models,” explains Nima Siboni of Deep- Metis. www.mpie.de/2281/en. AI TOOL FOR STRESS AND STRAIN Researchers at MIT, Cambridge, Mass., developed a technique to quick- ly determine certain properties of a ma- terial, like stress and strain, based on an image of the material showing its internal structure. The approach could one day eliminate the need for difficult physics-based calculations, instead re- lying on computer vision and machine learning to generate estimates in real time. The team believes the advance could enable faster design prototyping and material inspections. “It’s a brand new approach,” says Ph.D. student Zhenze Yang, adding that the algorithm “completes the whole process without any domain knowledge of physics.” Researchers used a machine learn- ing approach that included a genera- tive adversarial neural network. They trained the network with thousands of paired images—one depicting a mate- rial’s internal microstructure subject to mechanical forces, and the other showing color-coded stress and strain values. With these examples, the net- work uses principles of game theory to iteratively determine the relationships between a material’s geometry and the resulting stresses. Once trained, the network runs al- most instantly on consumer-grade com- puter processors, which could enable mechanics and inspectors to diagnose potential problems with machinery simply by taking a picture. In the study, researchers worked primarily with com- posite materials that included both soft and brittle components with random geometrical arrangements. In future work, the team plans to use a wider va- riety of material types. mit.edu . Machine learning solver for material mechanics. Courtesy of J.R. Mianroodi. Airwayz Drones Ltd., Israel, a specialist in artificial intelligence (AI)-based flight systems, formed a strategic partnership with Digiland Pte. Ltd., Singapore, a financing, technology, and security provider. The team plans to develop and offer ready-to-deploy drone fleet management solutions—reportedly the first in the world with AI— to governments, companies, and individuals. www.airwayz.co. BRIEF MIT researchers developed a machine learning technique that uses an image of a material’s internal structure to estimate the stresses and strains acting on the material.