Beyond the Instantaneous Kick: Trajectory-Resolved Migdal Rates in Semiconductor Dark Matter Detectors

Detectors searching for light dark matter increasingly rely on the Migdal effect. When an incoming particle knocks an atomic nucleus, the surrounding electrons briefly lag behind the moving nucleus and can be excited or knocked loose. This produces a small electronic signal even when the nuclear recoil itself is too faint to register. Every current prediction of this signal in semiconductors models the nucleus as receiving a single instantaneous velocity kick and then gliding freely. Real recoiling atoms in a crystal do something else: they are pushed back by their neighbours, and at recoil energies of tens to hundreds of electron-volts they strike neighbouring atoms, bend their paths, and are kicked again. Each of these later velocity changes is another source of electron excitation, and the kicks can reinforce or cancel one another.

We propose to compute the Migdal signal from the recoiling atom's actual velocity history rather than a single step. We generate recoil trajectories with density functional theory (DFT, the standard quantum method that computes a material's electronic structure and interatomic forces from its electron density). We then represent each trajectory as a piecewise fifth-degree (quintic) polynomial, borrowed from robot trajectory planning, that matches position, velocity, and acceleration exactly at every sampled time and keeps the true discontinuity at the moment of impact. Because the piecewise polynomial has a closed-form Fourier transform, the frequency content of the motion, which is exactly what sets the electron-excitation spectrum, comes out analytically and without the aliasing errors of a discretely sampled trajectory. The deliverable is a correction factor to existing Migdal rates for silicon, germanium, and diamond, validated against direct real-time quantum simulations of kicked atoms. It includes an explicit map of where the instantaneous-kick picture holds and where it fails.

Core Claim

In the bremsstrahlung-like description of Describing the Migdal effect with a bremsstrahlung-like process and many-body effects, the recoiling nucleus acts as a moving charge. Its velocity change radiates a virtual photon that the crystal's electrons absorb, with a strength set by the energy loss function. The energy loss function is the DFT-computed quantity $\mathrm{Im}[-1/\epsilon(\mathbf{k},\omega)]$ describing how readily the material absorbs energy $\omega$ at momentum $\mathbf{k}$. The $\omega^{-4}$ scaling that paper reports comes from treating the velocity change as a step. Schematically, and keeping that paper's prefactors, we replace the step's velocity jump $\Delta\mathbf{v}$ by a trajectory form factor

$$\mathbf{A}(\omega)=\int dt\,\dot{\mathbf{v}}(t)\,e^{i\omega t},$$

where $\mathbf{v}(t)$ is the recoiling atom's velocity history and $\dot{\mathbf{v}}$ its acceleration. The excitation rate at energy transfer $\omega$ then scales as

$$\frac{dR}{d\omega}\;\propto\;\int \frac{d^3k}{k^2}\,\frac{|\hat{\mathbf{k}}\cdot\mathbf{A}(\omega)|^2}{\omega^4}\,\mathrm{Im}\!\left[\frac{-1}{\epsilon(\mathbf{k},\omega)}\right].$$

For a pure step, $\mathbf{A}(\omega)=\Delta\mathbf{v}$ and the standard impulse-approximation result is recovered exactly.

The claim has two parts. First, for recoils of roughly 10 to 500 eV, the later hard collisions with neighbouring atoms happen on timescales of a few femtoseconds. That is comparable to the inverse of the eV-scale electronic energies that matter here, so $\mathbf{A}(\omega)$ should deviate from $\Delta\mathbf{v}$ by an amount that is measurable and depends on crystal direction. Second, piecewise quintic Hermite interpolation of the trajectory makes $\mathbf{A}(\omega)$ both accurate and cheap. Hermite interpolation here means fitting each segment to the position, velocity, and DFT-computed acceleration at its two ends. The smoothness of a function controls how fast its Fourier transform decays, so a representation that is smooth everywhere except at the one real impact time does not invent high-frequency content, which a finite-difference or FFT treatment of discrete samples would.

We state the scope honestly. A classical trajectory is meaningful only when the recoil energy is well above the lattice vibration quantum (about 60 meV in silicon, about 165 meV in diamond). This proposal therefore does not address the sub-50 MeV dark matter regime where the impulse approximation is known to fail. That regime belongs to the phonon-mediated treatment. Our target is the intermediate regime: dark matter of roughly 100 MeV to a few GeV, and neutron-calibration recoils. There the impulse approximation is routinely assumed valid, but the atom's subsequent path is ignored.

Relation to Prior Work

Three papers supply the physics we build on, and all three share the assumption we relax. Migdal Effect in Semiconductors derived the semiconductor Migdal rate in terms of the DFT energy loss function under the impulse approximation. It showed that isolated-atom calculations are inaccurate for crystals, but it treats the recoil kinematically as one kick. The bremsstrahlung-like paper above uses Quantum ESPRESSO and YAMBO to compute the dielectric response with many-body screening for silicon and diamond. It supplies our formalism and our energy loss functions, and it likewise assumes an idealized instantaneous kick with no later deceleration. Phonon-mediated Migdal effect in semiconductor detectors replaces the free nucleus with virtual phonon exchange and extends the calculation to 10–50 MeV. It finds agreement with the impulse picture in silicon. That result constrains us: our correction must vanish where the lattice restoring force is the only effect, and we treat reproducing it as a required check. It is also limited to monatomic crystals at zero temperature under an incoherent approximation. Diamond's light, stiff lattice is where we most expect differences to appear.

The trajectory side is motivated by Accelerating multiscale electronic stopping power predictions with time-dependent density functional theory and machine learning and its companion Accelerating Electronic Stopping Power Predictions by 10 Million Times with a Combination of Time-Dependent Density Functional Theory and Machine Learning. Those papers integrate real-time TDDFT electronic drag along ion paths in aluminium, and they report that stopping depth varies by 14% between crystal directions. They state explicitly that their paths are straight lines, with no deflection. We take the opposite regime: slow, strongly deflected recoils in semiconductors. We also compute a different observable, the excitation spectrum produced by changes in velocity rather than the integrated energy loss.

The polynomial machinery comes from robotics. There, quintic polynomials are the unique minimum-jerk curves matching position, velocity, and acceleration at the endpoints, and they are solved in closed form: A Fast Online Omnidirectional Quadrupedal Jumping Framework Via Virtual-Model Control and Minimum Jerk Trajectory Generation reports 50 microseconds per trajectory. An Efficient Trajectory Planner for Car-like Robots on Uneven Terrain uses piecewise quintics with four continuous derivatives. The research file suggests using such polynomials to plan the recoil path itself, and we reject that. A recoiling atom does not minimize jerk. Its path is set by forces, and imposing a guessed endpoint, which that approach requires, would be unphysical. We keep only the part that transfers: boundary-matched quintic interpolation as an exact, differentiable representation of physically computed trajectories, with an analytic Fourier transform.

Research Program

The first component generates trajectories. For silicon, germanium, and diamond, we sample recoil energies from 1 eV to 1 keV and recoil directions over the irreducible part of the crystal's symmetry. We integrate each recoil in a supercell for roughly 50–100 fs. The reference trajectories use Born–Oppenheimer DFT forces, meaning forces recomputed from the electronic ground state at each atomic configuration. For large ensembles we use a MACE interatomic potential (ACEsuit/mace) fitted to DFT, including short-range repulsive configurations, and we check it against DFT forces on held-out close-collision snapshots. Electronic drag along the path is added as a velocity-dependent friction term. Where possible we calibrate it against real-time TDDFT runs, using the straight-line surrogate idea of the stopping-power work as a starting point.

The second component represents the trajectory and computes $\mathbf{A}(\omega)$. Each trajectory becomes a piecewise quintic Hermite curve with knots at the integration steps. The first segment starts from the pre-impact rest state and has a true velocity jump at $t=0$. Every segment matches position, velocity, and acceleration exactly at both ends, which gives continuity of the second derivative. The Fourier integral of a polynomial segment against $e^{i\omega t}$ is closed-form, so $\mathbf{A}(\omega)$ is an exact finite sum for the interpolant. The truncation at the end of the simulation is handled by adding the analytic tail of a harmonically bound atom.

The third component assembles the rates. We combine $\mathbf{A}(\omega)$ with energy loss functions from Quantum ESPRESSO and YAMBO computed exactly as in the bremsstrahlung-like paper, including its diagonal-only dielectric approximation, so that the only change is the trajectory. We fold the result over the standard halo velocity distribution for dark matter masses from 10 MeV to 5 GeV. For neutron-calibration geometries we fold over recoil spectra at fixed recoil energy. The outputs are a ratio $R_{\rm traj}/R_{\rm IA}$ as a function of mass, energy transfer, and crystal orientation, and a decomposition of that ratio into contributions from the first kick, the lattice restoring force, secondary collisions, and electronic friction.

The fourth component validates against an independent calculation. For about fifty recoils per material spanning the energy range, we run Ehrenfest real-time TDDFT. In that method the kicked nucleus moves classically while the electrons evolve quantum mechanically, and the number of electrons excited above the gap is read off by projecting onto unoccupied states. This directly measures excitation along the real trajectory without using our form factor. It is expensive, and the stopping-power papers report that exhaustive trajectory sampling would cost hundreds of millions of core-hours. That cost is why it serves as a spot check and not as the method.

Baselines

The naive reference point is the impulse approximation as implemented in Migdal Effect in Semiconductors and the bremsstrahlung-like paper, which amounts to $\mathbf{A}(\omega)=\Delta\mathbf{v}$. We run it with the identical energy loss function, velocity distribution, and exposure (the 1 kg·yr silicon and diamond benchmarks used in those papers). Every difference is then attributable to the trajectory. The file contains no reported numerical Migdal event rate or cross-section limit to quote; "no reported result found on Lacuna for semiconductor Migdal rates at a specific mass and exposure." We therefore regenerate the impulse-approximation curves ourselves and require that they reproduce the published trends: $\omega^{-4}$ scaling and diamond's feature near 13 eV. This check comes before any comparison.

The strongest existing competitor differs by regime. Below about 50 MeV it is the phonon-mediated calculation in Phonon-mediated Migdal effect in semiconductor detectors, evaluated at its reported masses of 10, 20, and 50 MeV in silicon. There our method is at or beyond its validity limit, and it should agree with that calculation rather than beat it. Above 50 MeV the strongest competitor is the bremsstrahlung-like impulse calculation itself. On the question of what excitation a real recoil produces, the arbiter for both is Ehrenfest real-time TDDFT on the validation set. We count our method as beating the impulse approximation only if the impulse approximation misses the TDDFT excitation yield by more than the combined numerical uncertainty and our trajectory-resolved prediction does not. "Beating" is defined as reducing the error relative to TDDFT at least twofold in the regime where they disagree, under equal energy loss inputs and equal supercell convergence.

For the energy-loss component we use the straight-line surrogate of the stopping-power work as a reference. Its reported accuracy is a mean absolute percentage error of 1.9% on unseen crystal directions for protons in aluminium (Accelerating Electronic Stopping Power Predictions by 10 Million Times with a Combination of Time-Dependent Density Functional Theory and Machine Learning). That figure is not directly transferable, because it covers fast protons in a metal. We use it only to show that a straight path, fed the same friction model, gives a different form factor from the deflected path.

Two ablations isolate the key ingredients. The first keeps only the first kick plus harmonic restoring force and removes secondary collisions; this separates the lattice-binding correction, which the phonon paper suggests is small, from the collision correction we predict. The second replaces the quintic Hermite representation with cubic Hermite, which is continuous only in the first derivative, and with a direct FFT of the sampled velocities. All three receive the same trajectories and timestep, and we measure form-factor error against a timestep-converged reference.

Why not just…?

Why not just take a small molecular-dynamics timestep and Fourier-transform the sampled velocities? This is the most serious alternative, and it may work well enough. The concern is specific. At eV energies the relevant periods are a few femtoseconds, and close collisions last a similar time. A discrete transform of sampled velocities therefore mixes the true impact discontinuity, the truncation of the simulation, and aliasing at the sampling frequency. All three produce spurious high-frequency content, and the $\omega^{-4}$ weighting then amplifies it. The timestep-convergence ablation settles this directly. If the FFT converges to within 5% of the reference at a timestep no finer than the one the quintic representation needs, the polynomial ingredient is unnecessary and we will say so. The physics result, trajectory-resolved rates, stands regardless.

Why not just run Ehrenfest real-time TDDFT for every event? It answers the question directly, but its cost puts it far outside survey scale. The stopping-power papers quantify this, needing millions-fold acceleration to make trajectory ensembles feasible for a single material. Rates must be folded over masses, directions, and velocities. TDDFT is therefore our validation tool on roughly fifty events per material, and the form factor is what allows the result to generalize across thousands.

Why not just use empirical stopping codes or Lindhard-style ionization yields? SRIM-type codes assume isotropic materials, which the 14% directional difference in stopping depth reported in Accelerating multiscale electronic stopping power predictions with time-dependent density functional theory and machine learning contradicts. More fundamentally, these codes predict average energy lost to electrons, not the spectrum of the high-energy excitation tail that makes the Migdal effect a distinct signal. We will still compare our total electronic energy deposit against such yields as a consistency check.

Why not just extend the phonon-mediated field theory to higher masses? In principle it contains all the relevant physics. In practice its incoherent, rotationally symmetric, monatomic approximations are exactly what fail when an atom collides with a specific neighbour along a specific crystal direction. Agreement with it at 10–50 MeV is a required check, and any disagreement above that range is a prediction we test against TDDFT.

Decisive Checks

Four results decide whether the idea works. First, with lattice forces switched off, the trajectory-resolved rate must reproduce the impulse-approximation rate to numerical precision. Second, with only harmonic binding switched on, it must agree with the phonon-mediated silicon results at 10–50 MeV. The third is the central test. Across the TDDFT validation set, we measure how much the secondary-collision contribution changes the excitation yield above the band gap as a function of recoil energy and direction. We expect the largest effect in diamond and for recoils along densely packed directions, and a smaller one along open channels. If the trajectory correction stays below 10% everywhere from 10 eV to 1 keV in all three materials, we report that negative result. The instantaneous-kick picture would then be vindicated for semiconductor Migdal predictions, which is itself useful to experiments planning with silicon and germanium. Fourth, we test whether the quintic form factor converges at a coarser timestep than the direct FFT.

If the correction is large, the payoff is a corrected Migdal rate for the 100 MeV to GeV window and for neutron calibrations. It would also establish a direction-dependent Migdal rate, which could produce a daily modulation as the Earth rotates the detector's crystal axes relative to the dark matter wind.