Electromigration
Electromigration is the diffusion-controlled transport of metal atoms driven by momentum exchange with conducting electrons. At the current densities carried by on-chip interconnect — routinely above $10^{6}\ \text{A/cm}^{2}$, a level that would vaporise a macroscopic wire — that momentum transfer is large enough to displace lattice atoms over a product lifetime, opening voids that fail a line and growing hillocks that short neighbouring ones.
The driving force
An atom in a current-carrying conductor feels two electrostatic contributions: the direct field acting on its ionic charge, and the far larger "electron wind" from scattering carriers. Both are collected into an effective valence $Z^{*}$, giving
For copper $Z^{*}$ is negative and of order unity to ten, meaning the wind term dominates and atoms migrate toward the anode, in the direction of electron flow. The resulting atomic flux follows the Nernst-Einstein relation
so lifetime is exponentially sensitive to temperature and linearly sensitive to current density through the flux, before any statistics are applied.
Flux divergence, not flux, causes failure
Uniform atomic flux moves mass without changing local density and damages nothing. Damage requires divergence: a location where more atoms leave than arrive, or the reverse. Mass conservation gives
Divergence sites are structural, which is why electromigration is a layout problem as much as a materials problem. The classic sites are a via transition where copper meets a refractory liner that blocks diffusion, a sharp change in line width, a grain-boundary triple point, and any interface where the dominant diffusion path terminates.
Sign-off therefore needs two mappings, not one scalar current limit. Circuit activity maps into average and RMS current over the mission profile; routed geometry maps that current into local current density and temperature near vias, line ends, bends, barriers, and thermal bottlenecks. The reliability model then maps the local history into a failure distribution. Each mapping carries assumptions, so qualified rules must state waveform class, stack, geometry, temperature model, failure criterion, percentile, and applicable Blech exemption.
Which diffusion path dominates
Copper offers several transport routes in parallel, and the fastest one sets the lifetime. Their effective activation energies depend on the material stack, microstructure and stress window; they help diagnose a likely rate-limiting path but do not identify it uniquely.
| Path | $E_a$ (eV) | Comment |
|---|---|---|
| Cu / dielectric cap surface | 0.8 – 1.0 | dominant in unclad damascene Cu |
| Grain boundary | 1.0 – 1.2 | matters in fine-grained narrow lines |
| Cu / metal liner interface | 1.2 – 1.5 | Co or Ru cladding suppresses surface path |
| Bulk lattice | often higher than interface or boundary paths | generally slower in conventional Cu stacks; dominance must be established for the actual material and microstructure |
In conventional damascene Cu, transport along the Cu/cap interface is often rate-limiting. Metal caps, alloying, surface treatment, liner selection and geometry can suppress fast paths, but their benefit is process-specific and must be qualified against resistance, integration and reliability trade-offs.
Back-stress and the immortality criterion
Accumulating atoms put the anode into compression and deplete the cathode into tension, and that stress gradient drives a counter-flux. When it cancels the wind exactly, transport stops. The balance yields the Blech threshold product
For a qualified material stack and boundary condition, segments below the measured threshold may be treated as EM-immortal within that model. The exemption is not universal: reservoirs, vias, stress relaxation, thermal gradients and alternate failure modes can invalidate a simple $jL$ test.
Lifetime projection
Accelerated stress data is fitted to Black's equation and extrapolated to use conditions:
where the fitted current exponent $n$ depends on mechanism, stress regime, geometry, and failure criterion; values near one or two can arise in common regimes but are not universal mechanism labels. Failure distributions are often modeled as lognormal over a qualified population, and sign-off targets a specified low failure percentile rather than assuming the median represents usable life.
Design and sign-off practice
EM checking extracts per-segment RMS and average current from dynamic simulation, since void growth responds to average electron flux while self-heating responds to RMS. Each segment is tested against a current-density limit derated for temperature, then exempted if it satisfies the Blech criterion. Power grids and heavily active clock networks often dominate violation lists because of sustained current and self-heating. Reversing current can reduce net directional transport in some structures, but recovery is incomplete and mechanism-dependent; bidirectional nets require waveform- and stack-qualified rules rather than an automatic relaxed limit. Self-heating couples back into the Arrhenius term, so a line that passes at nominal temperature can fail once $\Delta T_{\text{Joule}} \propto j^{2}\rho R_{\text{th}}$ is included — a common source of late-stage sign-off surprises.
Electromigration is ultimately a constraint on current density, geometry and interface quality simultaneously, and it is the reason advanced-node interconnect development is dominated by liner, cap and grain-structure engineering rather than by simple dimensional scaling.