Discovering the Building Blocks and Structure Property Relationships of Grain Boundaries Using Machine Learning

Brigham Young University, Provo UT

Investigators

Abstract

NONTECHNICAL SUMMARY This award supports computational, data-centric, theoretical research and education aimed to include the effects of defects in optimizing the properties of metal alloys. Defects in the structure of materials, such as interfaces between constituent crystallites or grains, can strongly influence their strength or resistance to corrosion and cracking. Scientists and engineers have been able to create a limited number of materials with enhanced corrosion and crack resistance by finding and increasing the fraction of interfaces, or grain boundaries, that were correlated with these properties. Unfortunately, these successes were limited to a small fraction of real-world materials. The PIs' preliminary work suggests that through artificial intelligence techniques - machine learning - grain boundary structures can be identified that correlate with specific material behaviors. This award supports the application of machine learning and other data intensive techniques to correlate microscopic grain boundary structure with different macroscopic properties. Specific attention will be paid to which methods provide the most useful insight and which methods allow one to identify the physical processes controlling grain boundary behaviors of interest. The work will enable the next generation of materials to be tailored with properties unique to specific applications. This could enable a range of optimized materials such as alloys with superior corrosion-resistance and deformation-resistance, high strength-high ductility materials, and enhanced fracture-resistant materials. This project supports education and training of the future workforce in data-intensive materials research. Software developed through research and data created through research will be made available to the broader community. TECHNICAL SUMMARY This award supports computational, data-centric, theoretical research and education to combine machine learning and multi-resolution representations to discover grain-boundary structure-property relationships and advance materials design. Identifying the microscopic features that affect macroscopic materials properties is essential to materials design. Optimizing alloy performance where grain boundaries play an important role is limited by current understanding of the microscopic structures that affect macroscopic properties. The PI will combine machine learning and multi-resolution representations to discover grain boundary structure-property relationships. Preliminary results led to representations that retain physical interpretability and suggest that the physical mechanisms that control grain boundary behavior can be identified. The PI's aim to obtain grain boundary structure-property relationships that incorporate both atomistic and crystallographic structure. To pursue this goal, the PIs will: 1. Apply Statistical Mechanics to Grain Boundaries: Grain boundary potential energy landscapes will be examined for their influence on both static and dynamical properties; 2. Determine Long-Range Effects of Structures on Properties: A multi-scale representation will be used to predict properties, such as shear coupling, which cannot be predicted well using only short-range environments. The PIs aim to identify the short- and long-range structures that correlate with various properties and verify their role using another representation, scattering convolutional networks. 3. Connect Crystallographic Property Trends to Atomic Structure: Translate atomistic property trends learned in this project into experimental, or crystallographic, coordinates. This will connect microscopic structure to macroscopic properties and advance toward grain boundary engineering. The research combines modeling, theory, machine learning, and simulations to discover grain boundary structure-property relationships and their governing physics. Two graduate students will work closely with both PIs to form a cross-disciplinary team to tackle this challenging problem and will be trained in data-intensive materials research and other advanced methods of materials research. Software developed through research and data created through research will be made available to the broader community. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.