Cobalt Interconnect Agglomeration
# Cobalt Interconnect Agglomeration
Cobalt replaced copper at the very narrowest local interconnect lines for a specific reason: below roughly 20nm linewidth, copper's own resistivity shoots up from electron scattering at the wire's grain boundaries and surfaces, and cobalt's shorter electron mean free path makes it less penalized by that same scaling. Thin cobalt lines carry a different liability instead. Heat the metal during back-end-of-line processing, and a thin, continuous cobalt film doesn't always stay a film — it can ball up into discrete islands, the same way a thin film of any metal minimizes its surface energy by breaking into droplets once it's hot enough and thin enough. That balling-up is agglomeration, and in a cobalt line a few atoms wide, it means the wire itself can pinch off into gaps with nothing conducting across them.
Agglomeration is a surface-energy problem that gets worse exactly as lines get thinner. A continuous film has more total surface area per unit volume than the same metal redistributed into isolated islands, so above some temperature threshold, the thermodynamic drive to minimize that surface energy wins and the film breaks apart — a well-known failure mode in thin-film metallurgy generally, not unique to cobalt, but cobalt interconnect hits the threshold at process-relevant temperatures specifically because the lines are thin enough that surface effects dominate over bulk behavior. A thick cobalt film barely notices the same thermal budget that can agglomerate a line only a few nanometers across, which is exactly why this became a back-end-of-line reliability concern only once local interconnect scaled down to cobalt's own sweet spot.
The electrical outcome ranges from a quiet resistance increase to an outright open. A partial agglomeration that thins the line without fully severing it still raises local resistance, since current has to crowd through a narrower remaining cross-section — a slow performance degradation that can pass initial test and only show up as a timing or IR-drop margin problem later. A full pinch-off is a dead open, catching it on wafer-level electrical test or, worse, as a field failure if the agglomeration event happens gradually under operating-temperature thermal cycling rather than during the back-end-of-line anneal itself.
Process control works the same materials-science lever in reverse — raise the energy barrier to islanding instead of lowering the temperature the line sees. A thin liner or capping layer (often a nitride or a thin ruthenium interface, the kind of barrier/liner stack work Applied Materials and Lam Research both develop for sub-20nm interconnect) anchors the cobalt film mechanically and electronically, raising the activation energy agglomeration needs to get started. Grain-boundary engineering during cobalt deposition — favoring larger, more uniform grains over a fine polycrystalline structure — reduces the density of easy nucleation sites a breaking film would start from. And back-end-of-line thermal budget itself gets tracked specifically against the cobalt module's own agglomeration threshold, not just against the usual dopant-activation and low-k-stability limits the rest of the flow already has to respect, since cobalt's threshold can sit lower than either of those.
STI divot and the bird's beak lose area at an isolation edge; silicide piping punches a hole through a junction from above. Cobalt agglomeration fails a conductor by a completely different mechanism — not contamination, not a crystal defect, just ordinary thin-film thermodynamics catching up with a line that's been scaled thin enough for surface energy to win.