Adaptive machine learning for efficient materials design
- PDF / 89,106,992 Bytes
- 8 Pages / 585 x 783 pts Page_size
- 63 Downloads / 217 Views
troduction Materials that enable critical technological applications, including energy, electronics, security, and environment, are multicomponent by design and have enormous complexities at the electronic, atomic, mesoscale, and continuum length scales. Historically, the value of developing predictive computational strategies and tools that enable promising discoveries in the form of new materials and science have been pivotal in driving key materials innovations. Traditionally, new materials are discovered by trial-and-error or intuition-based approaches. However, these approaches can be inefficient, especially when the search space becomes vast because of increased complexity in chemistry, crystal structure, or microstructure. In materials science, we have a combinatorially large search space spanning millions of possibilities, but only a tiny fraction is experimentally explored. In addition, each experiment can be expensive and time-consuming, requiring rigorous synthesis efforts, heat-treatment protocols (if any), and characterization to produce one data point. Although combinatorial synthesis and characterization efforts have been successfully demonstrated, the effect has been realized only for a few materials classes.1–6 Key examples include combinatorial discovery of Ni-Ti-Cu-based ternary shape-memory alloy compositions with an extremely small hysteresis width6 and a combinatorial search for quarternary Ce-rich oxygen-evolution catalyst oxides for water electrolysis.1 Physics-based computational methods, such as density functional theory (DFT)7,8 and molecular dynamics (MD),9
have successfully uncovered insights10–12 and have guided experiments toward promising regions in the design space. For example, Burkert et al. used DFT calculations that predicted that the tetragonal FeCo alloys have a large saturation magnetization and tunable uniaxial magnetic anisotropy energy as a function of alloy composition, showing promise for hard disk technology. Later, Andersson et al.13 experimentally confirmed these predictions and found qualitative agreement between theory and experiments. Similarly, Takenaka et al.12 used MD simulations to examine and resolve the local structure of a prototypical relaxor ferroelectric material PbMg1/3Nb2/3O3PbTiO3 with unusually large piezoelectric coefficients. Prior to this work, there was no consensus on the origin for explaining some of the functional properties of this material. Their MD simulations revealed that the unusual properties of these relaxors arise from a multidomain state with small domain sizes (2–10 nm) and not from the widely used nonpolar matrix, owing to the local dynamics. The outcome was a “slush-like” multidomain state as a novel framework for describing the relaxor structure. Takenaka et al. also proposed a new design rule, which suggested that new relaxors could be discovered by tailoring materials with low local domain-wall energies. Recent experimental work by Peters et al.14 showed evidence supporting the design rule. However, these methods have limitations
