Detectors hunting for dark matter lighter than a proton cannot see the tiny kick such a particle gives an atomic nucleus. They rely instead on the Migdal effect: when the nucleus is jolted, its electron cloud lags behind, and that lag can lift an electron across the semiconductor's band gap, where it can be counted. The predicted signal depends on the material's energy loss function (ELF), the quantity that says how strongly the crystal's electrons absorb energy at a given energy and momentum. In the standard formula this function is weighted by roughly $\omega^{-4}$, where $\omega$ is the energy deposited. That weighting puts almost all of the predicted rate in the first electron-volt or two above the band gap.
This region is where current first-principles calculations are weakest. They compute the ELF from density functional theory (DFT, which describes a crystal's electrons through their ground-state density) on coarse momentum grids. They then smear each electronic transition with an artificial broadening width chosen by hand, and the near-threshold rates are known to depend on that choice. We propose to split the electronic response into two parts. The first is the smooth screening from high-energy transitions, computed once on a coarse grid with full crystal detail. The second is the few band-edge transitions that dominate the rate, computed on an extremely dense grid through a compact localized-orbital model. The two are then recombined exactly. The goal is near-threshold Migdal and electron-scattering rates that converge as the grid is refined rather than depending on an arbitrary broadening width, for silicon, germanium, and the compound targets already tabulated in public tools.
Core Claim
The partitioning used in constrained random phase approximation (cRPA) can be turned into a numerical tool for recoil physics. In condensed matter, cRPA separates the polarization of a small set of "target" bands from all other transitions, in order to build effective models; Davydov et al. and Li et al. on K3C60 are examples. Our use keeps the full response and simply computes its two pieces at different resolutions. Writing $\chi^0_{\mathbf{G}\mathbf{G}'}(\mathbf{q},\omega)$ for the independent-particle polarizability between reciprocal lattice vectors $\mathbf{G},\mathbf{G}'$ and $v(\mathbf{q}+\mathbf{G})$ for the bare Coulomb interaction, the dielectric matrix is
$$ \epsilon_{\mathbf{G}\mathbf{G}'}(\mathbf{q},\omega)=\delta_{\mathbf{G}\mathbf{G}'}-v(\mathbf{q}+\mathbf{G})\left[\chi^{0,\mathrm{high}}_{\mathbf{G}\mathbf{G}'}(\mathbf{q},\omega)+\chi^{0,\mathrm{edge}}_{\mathbf{G}\mathbf{G}'}(\mathbf{q},\omega)\right]. $$
Here $\chi^{0,\mathrm{high}}$ contains every transition outside a chosen band-edge window. It is smooth in $\mathbf{q}$ and $\omega$ near threshold, so it can be computed on a coarse grid with full off-diagonal terms (the local field effects that encode how the lattice shapes the screening cloud). $\chi^{0,\mathrm{edge}}$ contains only transitions between the top valence and bottom conduction bands. It is evaluated on a dense grid via Wannier interpolation, which represents the bands with localized orbitals so they can be recomputed cheaply at any momentum.
Because no transition is counted twice, this split is exact in principle. The choice of window changes the numerical cost, not the physics. That gives the method a built-in self-test: if the rate changes when the window changes, the numerics are wrong. This double-counting ambiguity is a known problem when cRPA is used to build models; here it becomes a diagnostic.
The central empirical claim is that the Migdal rate integrated above threshold will converge as the edge grid is refined, with a residual dependence on broadening much smaller than in coarse-grid calculations. The claim also includes a physical question. If the resulting spectra differ from current coarse-grid predictions in Ge and GaAs, where the conduction bands have sharp, low-lying minima, the difference measures how much the near-threshold Migdal predictions currently in use are shaped by the broadening width.
Relation to Prior Work
Two papers supply the physics we reuse. Knapen, Kozaczuk and Lin derived the semiconductor Migdal rate in terms of the ELF, using DFT inputs for Si and Ge. They kept only the diagonal of the dielectric matrix, after estimating that off-diagonal terms change the rate by at most an $\mathcal{O}(1)$ factor. Liang et al. recast the same process as bremsstrahlung, computed the random phase approximation (RPA) dielectric function with Quantum ESPRESSO and YAMBO, and obtained the $\omega^{-4}$ spectral weighting. They found rates far above those from atom-centered tight-binding orbital models, again without full local field effects. We keep both rate formulas unchanged and replace only how the ELF is computed near threshold.
Two papers motivate the proposal, because they show the near-threshold numerics are the weak link. In-medium screening effects for halo and solar-reflected dark matter found that local field effects move electron-scattering rates by up to a factor of about 0.5. It also found that low-mass results depend sensitively on the empirical broadening $\eta$. DarkELF found large systematic differences between its Lindhard, Mermin, and GPAW time-dependent DFT ELFs, and recommends the first-principles GPAW option for single-ionization signals, which are exactly the band-edge signals.
One paper constrains the proposal. Phonon-mediated Migdal effect in semiconductor detectors found that its lattice-phonon treatment agrees with the free-recoil bremsstrahlung treatment in Si down to 10 MeV dark matter. We therefore stay with the free-recoil kinematics and do not claim to fix the impulse approximation. Our improvement concerns only the electronic side.
The cRPA papers (Davydov et al., Li et al., kagome FeGe models) supply the idea of partitioning the polarization and the symmetry-constrained Wannier construction. None of them applies it to external scattering.
Research Program
The pipeline starts from Quantum ESPRESSO ground states for Si, Ge, GaAs and GaN. These are the Si and Ge used by Knapen et al., plus two compound targets tabulated in DarkELF. GaN adds an anisotropic crystal structure.
We then build symmetry-preserving Wannier orbitals spanning the top valence and bottom conduction manifolds, disentangling them where bands cross, as in the modified Wannier90 route of Davydov et al..
Next, YAMBO computes $\chi^{0,\mathrm{high}}$ with all remaining bands and a converged set of reciprocal vectors on a coarse grid. We interpolate it in $\mathbf{q}$, which is safe because it has no sharp features near threshold.
The band-edge term $\chi^{0,\mathrm{edge}}$ is evaluated on grids one to two orders of magnitude denser. The transition matrix elements are expanded in Wannier orbitals, and we keep the few off-diagonal $\mathbf{G}$ components that the band-edge states actually require. This is an approximation, and we test it directly against full calculations. We apply a scissor correction for the DFT band-gap underestimate identically in every method, so that gap error does not masquerade as a method difference.
The recombined ELF feeds the unchanged Migdal integrand of Knapen et al. and the DM-electron rate used in DarkELF.
The same decomposition answers the diagnostic question raised by Liang et al. We compute the rate three ways. The first uses band-edge transitions alone with bare Coulomb coupling, which is a tight-binding-like model. The second is band-edge transitions screened by the high-energy background. The third is the full response. The differences between them attribute the gap between the tight-binding and bremsstrahlung predictions to specific missing physics.
Baselines
The naive reference point is the same Migdal integrand evaluated with the DarkELF Lindhard ELF for each target. This is the cheapest option and shows how much crystal detail matters at all.
The strongest existing competitor is a coarse-grid, all-band YAMBO or GPAW ELF with full local field effects, run the way current Si and diamond studies are run, at their typical broadening. This is the method of In-medium screening effects and the GPAW path in DarkELF. No reported result found on Lacuna for an absolute near-threshold Migdal rate at fixed dark matter mass in Si or Ge. The only quantitative figures the file gives for this baseline are that local field effects move rates "within a factor of 0.5" and that screening suppresses solar-reflected rates by "a factor of 0.1-1" (In-medium screening effects).
We run this baseline on identical ground states, pseudopotentials, scissor shift and dark matter halo model, using the SHM conventions of Recommended conventions: $\rho_\chi=0.3$ GeV/cm³, $v_0=238$ km/s, $v_{\mathrm{esc}}=544$ km/s. It gets the same core-hour budget as our method, spent on its densest affordable grid.
For Si we additionally run an expensive brute-force all-band calculation on the densest grid we can afford. It serves as ground truth rather than as a competitor. Beating the strongest baseline means two things. At matched compute, the split method's threshold-integrated rate must sit closer to the brute-force Si reference. And across a tenfold reduction of $\eta$, its rate must change by several times less than the baseline's.
Two ablations isolate the key ingredient. The first computes $\chi^{0,\mathrm{edge}}$ on the coarse grid only, which removes the resolution gain but keeps the split. The second drops off-diagonal terms from $\chi^{0,\mathrm{high}}$, which tests whether the local field effects live in the background part as expected.
Why not just…?
Why not just run the all-band calculation on a denser grid with smaller broadening? For silicon this is a real threat. With enough compute it may converge, which is why we use it as the reference. The cost of this approach grows with the number of grid points, the number of bands, and the square of the number of reciprocal vectors. That growth becomes prohibitive for Ge and GaAs with spin-orbit coupling and for low-symmetry cells. The file lists poor scaling of full dielectric matrices with cell size and cutoff as an open limitation of DFT response methods. If the Si comparison shows brute force reaching the same accuracy at similar cost, the proposal's value narrows to the heavier targets. We would report that plainly.
Why not just use the Mermin or optical-data ELF in DarkELF? Optical data measure the true band edge, but only at vanishing momentum. The Migdal integrand needs finite momentum transfer, and Mermin extrapolates to it with electron-gas assumptions. DarkELF itself reports large differences from the first-principles result. We include Mermin in the comparison so readers can see where in $(\mathbf{q},\omega)$ it fails.
Why not just use a Wannier tight-binding model of the band edge alone? Liang et al. found that such models give much lower low-energy rates than screened calculations. Our band-edge-only ablation reproduces this class of model on identical inputs. It tests whether the missing ingredient is background screening, rather than assuming it.
Proposed Experiments and Decisive Checks
The first check is window independence. On Si and Ge we vary the band-edge window from the minimal set of bands to one that includes several additional bands. The threshold-integrated Migdal rate for 20 to 100 MeV dark matter should stay within a few percent. A larger drift means the split is being implemented incorrectly, and it would falsify the approach as built.
The second check is agreement with the brute-force Si reference. We compare the ELF point-by-point for energies from the gap to 10 eV and several momenta, and compare the rate spectrum weighted by $\omega^{-4}$.
The third check is broadening convergence. For each target and method we plot the rate in the first 2 eV above the gap against $\eta$. The split method should approach a limit as the grid is refined. The baseline is expected to keep drifting.
The fourth check is attribution. Using the three-way decomposition, we report what fraction of the difference between bare band-edge rates and fully screened rates comes from background screening versus band-edge local fields, at each dark matter mass. This bears directly on the tension between the bremsstrahlung and tight-binding pictures.
Finally, for GaN we compute the ELF along crystal axes. We report whether the near-threshold rate depends on direction by more than the method's convergence error. This is a precondition for any later daily-modulation study and is flagged as open in Phonon-mediated Migdal effect and Li and Xia.
We will release the converged ELF tables in the DarkELF format so that the SENSEI-, CDEX- and SuperCDMS-style analyses cited in the file can adopt them directly.
References
- Construction of low-energy symmetric Hamiltonians and Hubbard parameters for twisted multilayer systems using ab initio input
- Ab initio studies on interactions in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mrow><mml:msub><mml:mi mathvariant="normal">K</mml:mi><mml:mn>3</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">C</mml:mi><mml:mn>60</mml:mn></mml:msub></mml:mrow></mml:math> under high pressure
- Migdal Effect in Semiconductors
- Describing the Migdal effect with a bremsstrahlung-like process and many-body effects
- In-medium screening effects for the Galactic halo and solar-reflected dark matter detection in semiconductor targets
- python package for dark matter scattering in dielectric targets
- Phonon-mediated Migdal effect in semiconductor detectors
- FeGe as a building block for the kagome 1:1, 1:6:6, and 1:3:5 families: Hidden <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mi>d</mml:mi></mml:math>-orbital decoupling of flat band sectors, effective models, and interaction Hamiltonians
- Recommended conventions for reporting results from direct dark matter searches
- Enhancement of the screening effect in semiconductor detectors in the presence of the neutrino magnetic moment