
Now Oak Ridge National Labs (ORNL) and its Titan supercomputer — the second fastest in the world — is working with superconductor detection algorithms from Rutgers University to turn the Titan supercomputer into a giant virtual test tube for evaluating new materials formulations before bothering to fabricate them.
"The ability to predict important material properties as a function of the chemistry of the material is a big deal," Jack Wells, director of science for the National Center for Computational Sciences at ORNL, told EE Times. "The ability to predict these materials trends should enable new predictions for experimentalists to pursue."

Each graph shows the simulated intensity of spin excitations in 15 iron-based materials, including known high-temperature superconductors (images d–h). The x axis shows the momentum of the spin excitation, and the y axis shows the energy measured in electron volts (eV). The color code indicates the intensity of spin excitations, which is compared with available experimental results (shown in black bars in images f, g, l, and m). The locations with the greatest number of spin excitations are red with decreasing frequency shown from orange to blue, allowing the researchers to predict which materials are likely to be superconducting.(Source: ORNL)
Trial-and-error has been an element of the scientific method since the alchemists of antiquity, but the Rutgers researchers, including professors Kristjan Haule, Gabriel Kotliar, and post-doctoral fellow Zhiping Yin, think it about time we used more informed methods.
Their first step was to evaluate the spin dynamics of known superconductors to identify the similarities that the researchers believe set up the conditions allowing a metal to conduct electricity without resistance. By identifying the way electrons orient and correlate their spins in a superconducting material, they can look for those same characteristics in designer materials before going to the time and expense of fabricating and testing them in the lab.
The devil is in the details, according to Kotliar’s group, which picked 15 materials — four known to be superconductors. By comparing the supercomputer simulations with the predicted spin dynamics measured in the known superconductors, they were able to identify that common characteristics that, supposedly, permit an electron to conduct electricity without resistance.
"Within a broad class of iron pnictides and chalcogenides, we have indeed validated the approach against available experimental data," Kotliar told EE Times. "Futhermore, we have made predictions for the spectra of materials such as MgFe2 where the experiments have not been carried out."
The strategy the experimenters are using is theoretically predicting neutron scattering measurements, which could be used in many different classes of compounds, such as plutonium-based compounds, which are very different from iron compounds, according to Kotliar.
Kotliar’s group would first compute some properties of those compounds for which some measurements exist. This would allow them to tune a few knobs in the computational machinery; and then with the knobs tuned, they could predict the neutron scattering spectra.
Next the researchers would perform experimental validation, or encourage other scientists to, in order to increase the accuracy of their computational machinery, according to Kotliar.
"Predictive theory in the field of correlated electrons materials is still in its infancy, and I would like to see this feedback of theory computation and experiment to continue for a while," Kotliar told us. "What we have demonstrated in this paper, is proof of principle."
Once the method is proven with other compounds, the interplay between simulations and experimentation could become a two-way street. "This type of calculation demonstrates that theory and computation are becoming equal partners with experimentation in the area of materials research."
The team used Dynamical Mean Field Theory to reduce the vast number of interactions and the Monte Carlo method to statistically select the best solutions. Even with those shortcuts, the team used 20 million processor-hours on the 27 petaFLOP Titan to prove its concept.
Next, the researchers plan to optimize their techniques, including the algorithms; find cooperative methods for humans and supercomputers to search for and find interesting materials; and study ones that are difficult to study experimentally, such as radioactive materials. "Given these elements and some patience, we are going to find very interesting things. We have not yet hit the jackpot, but there are good signs all along," Kotliar told us.
Funding was provided by the National Science Foundation, the Oak Ridge Leadership Computing Facility, and the Department of Energy Office of Science User Facility at ORNL.
About the author:
R. Colin Johnson is Advanced Technology Editor at EE Times
