The failure didn't make sense at first.
The copper busbar was oversized. RMS current stayed within rating. Steady-state thermal simulation predicted an acceptable bulk temperature. On paper, the design was conservative by every conventional metric. And yet, after repeated fast switching events, edge discoloration appeared. Weeks later, microscopic cracking showed up near the surface. Eventually, contact resistance climbed and the joint degraded.
No design limit had technically been violated. The system had simply crossed into a regime the design model never accounted for.
The comfortable starting point
Every engineer learns Joule heating the same way. Power dissipation scales with the square of current: P = I²R. Locally, volumetric heat generation follows q = J²ρ, where J is current density and ρ is resistivity. Double the current, quadruple the heating. Clean, predictable, and true, provided one assumption holds: current density is uniform across the cross-section.
That assumption is only valid at low frequency. A busbar switching in tens of nanoseconds is not a low-frequency problem, even if the underlying system reads as "DC" on a multimeter.
When rise time becomes frequency
A fast switching edge behaves electromagnetically like high-frequency excitation, whether or not the waveform is technically periodic. Rise time sets an effective frequency: f_eff ≈ 1/t_r. A ten-nanosecond edge corresponds to spectral content well into the megahertz range, and megahertz-range currents don't distribute the way power-frequency currents do.
This is the skin effect. Time-varying current induces internal magnetic fields that oppose penetration into the conductor's interior, pushing current toward the surface. Current density with depth follows J(x) = J₀·exp(−x/δ), where δ is the skin depth: δ = √(2/(ωμσ)). Skin depth shrinks as frequency, permeability, or conductivity increase.
For copper, that shrinkage is dramatic. At 50 Hz, skin depth is roughly 9 mm, comparable to the whole cross-section of a modest busbar. At 10 kHz, it drops to around 0.66 mm. At 1 MHz, it's down to tens of micrometers, a layer thinner than a sheet of paper. At that point, effective conducting area collapses to roughly the conductor's perimeter multiplied by skin depth, not its full cross-section. The same total current now has to squeeze through a fraction of the available material, and because local heating scales with the square of current density, that squeeze produces a sharp, disproportionate rise in surface heating.
A second clock the design missed
There's a related and easy to miss parameter: electromagnetic diffusion time, τ_em ≈ μσL², which describes how long it takes current to penetrate fully into the bulk of a conductor of thickness L. If a pulse is shorter than this diffusion time, current simply doesn't have time to reach the interior before the excitation changes. The bulk sits electromagnetically idle while the surface absorbs nearly the full stress of the event.
So a system can show a modest average current on paper while its peak surface current density during each switching transient is extreme. Two very different pictures of the same event, depending on which clock is being read.
Heat that doesn't have time to spread
Temperature evolution follows the heat equation: ρ_m c_p (∂T/∂t) = k∇²T + J²ρ, balancing volumetric heat generation against thermal diffusion. Under steady, continuous excitation, gradients smooth out because diffusion has time to redistribute the energy. Under short pulses, energy deposits faster than it can spread. The result is a steep temperature gradient confined to a thin surface layer: high local thermal expansion, high interfacial shear stress, all while a bulk-averaged sensor reads a perfectly comfortable temperature.
This is precisely why the failure looked invisible from the outside. Bulk sensors were telling the truth about the bulk. They were never positioned to see what was happening in a layer tens of micrometers deep.
The feedback loop hiding inside "simple" copper
Copper's resistivity rises with temperature at roughly 0.39% per kelvin. That single fact turns this from a one-way problem into a coupled one. Surface current concentrates due to skin effect, which raises local temperature, which raises local resistivity, which further concentrates current toward the surface, which raises temperature again. Each step feeds the next.
The conductor stops behaving like a static, well-characterized resistor and starts behaving like a dynamically evolving electromagnetic-thermal system, one where temperature rise is no longer simply proportional to current squared. It becomes regime-dependent: shaped by frequency content, rise time, geometry, and thermal time constants working together rather than any single variable acting alone. A modest increase in peak current, or a modest reduction in switching rise time, can produce a disproportionately large jump in surface temperature, gradient severity, and fatigue accumulation.
Why the original calculations looked completely defensible
The original design used DC resistance, uniform current density, steady-state thermal modelling, and bulk temperature limits: a textbook-correct approach for the regime it assumes. Nothing about those calculations was wrong. They simply described a different physical situation than the one the busbar actually experienced once switching time scales approached electromagnetic and thermal diffusion times.
Maxwell's equations still hold. Ohm's law still holds. Fourier's law still holds. What shifted wasn't the physics, but the hierarchy of time scales the physics was operating under, and that shift moved the entire problem out of the regime where the original assumptions were valid.
The real lesson
Once switching rise time approaches or drops below electromagnetic diffusion time, a conductor stops acting like a uniform resistive volume and starts acting like a spatially evolving system where current, temperature, and material properties interact in real time. Past that boundary, "more current" isn't a bigger version of the same linear problem. It's a different physical regime entirely, one governed by surface physics rather than the RMS number sitting comfortably within spec on a datasheet.
For anyone designing pulsed power systems, fast switching contacts, or high-density busbars, the practical takeaway isn't that classical Joule heating is wrong. It's that classical Joule heating has a domain of validity, and knowing where that domain ends is the difference between a design margin that actually protects the part and one that only protects the paperwork.