Screening 2.8 million ordered metals for conductivity: machine learning ranks, electron-phonon calculations check
Synopsis
Computing metallic conductivity from first principles is expensive, so large chemical spaces have gone largely unscreened. This workflow ranks about 2.8 million ordered compounds using two machine-learning proxies, a BoltzTraP2 conductivity under the constant relaxation-time approximation and the Eliashberg electron-phonon coupling λ, combined by symbolic regression into σSR=0.4σBT/λ, then runs full electron-phonon transport on the candidates. Be2IrPt, H3NaNi, LuIr and LiBePt2 reach room-temperature conductivities of roughly 34 to 38×106 S/m, comparable to aluminum's 36.59×106 S/m but still below Ag, Au and Cu.
Interpretation
High-throughput screening of metallic conductivity can now cover about 2.8 million ordered compounds instead of a small set computed one at a time. First-principles transport needs band structures, scattering rates and transport integrals, which confined earlier work to small material sets, and conductivity has remained largely absent from high-throughput databases. Against first-principles validation on about 14,000 compounds, machine-learning predictions show root-mean-square deviations of 4.14×106 S/m for conductivity and 0.18 for λ on overlapping points, and anomalously high false positives were removed by the validation step.
Ranking rests on an empirical expression with two inputs that combines the electronic-structure conductivity with scattering strength. The BoltzTraP2 σBT depends only on the band structure and uses the constant relaxation-time approximation, so it omits electron-phonon scattering and cannot be compared directly with measured conductivity. Symbolic regression on 43 materials with experimental conductivity yields σSR=0.4σBT/λ, which describes noble, alkali and transition metals well, with Sr and Ca the most noticeable outliers.
Top-ranked candidates keep high conductivity after full iterative Boltzmann transport calculations. The screening threshold was σSR>15×106 S/m, and after validation one representative with the highest predicted conductivity was kept per elemental composition, leaving 52 systems. EPW gives Be2IrPt 38.02, H3NaNi 34.82, LuIr 34.41 and LiBePt2 34.35×106 S/m, comparable to aluminum's 36.59×106 S/m, with a mean absolute error of 3.82×106 S/m between σEPW and σSR.
The high conductivity of noble metals comes from a combination of band structure and lattice vibrations rather than one factor. Silver has a closed 4d shell with d-bands about 4–7 eV below the Fermi level, a steeply dispersive 5s band at the Fermi level, low phonon frequencies and very weak coupling to conduction electrons, with λ about 0.1. Aluminum has a multi-sheet Fermi surface and λ about 0.45, superconducts at 1.2 K, and the same coupling appears at room temperature as a higher scattering rate that limits its conductivity relative to silver.
Perspective
The workflow targets ordered inorganic compounds in the Alexandria database within 25 meV/atom of the convex hull, with transport computed at the scalar-relativistic level and spin-orbit coupling tested only for Pb, Be2IrPt and LiBePt2. The experimental benchmark is a set of elemental metals with established conductivity data, not compositionally complex compounds.
Spin-orbit coupling is neglected, and including it lowers the room-temperature conductivity of Be2IrPt and LiBePt2 by 8.1% and 9.3% respectively while bringing Pb closer to experiment. Ca and Sr show the largest deviations from experiment, which the authors attribute to d-state hybridization near the Fermi level and quasiparticle corrections that may alter Fermi-surface topology. The σBT surrogate is also sensitive to k-point sampling and less accurate than the λ model.
