Two-phase cooling: lessons from 8 years of multiphase pipelines

Two-phase direct-to-chip cooling is being evaluated across the industry as a way to remove more heat per unit of coolant mass flow, which matters as rack power climbs past 100 kW. The physics that makes it attractive, latent heat absorption at a near-constant temperature, is also the physics that makes it unforgiving to design with steady-state methods.

Multiphase pipeline engineering has been dealing with this exact tension for decades. A few lessons transfer directly.

Flow regime is not optional to model

In a single-phase loop, flow regime is not a meaningful concept: liquid moves through a pipe and that is the whole story. In a two-phase loop, the same mass flow rate can produce a bubbly, slug, churn, or annular flow regime depending on local geometry, orientation, and heat flux, and each regime has different pressure drop and heat transfer characteristics. A cold plate designed around an assumed annular regime will underperform, sometimes badly, if local conditions produce slug flow instead.

This is precisely why steady-state, single-point pressure drop calculations fail for two-phase networks in a way they do not for single-phase ones: the regime itself is a function of the transient operating point, not just the design point.

Density wave oscillations are a known failure mode, not an edge case

Boiling channels connected to a common manifold are susceptible to density wave oscillations: a self-sustaining instability where a perturbation in vapor quality changes local pressure drop, which changes flow, which changes vapor quality, and the cycle reinforces itself rather than damping out. This is well documented in nuclear and process engineering literature going back to the 1960s. It is not a hypothetical for a data center two-phase loop; it is a predictable consequence of connecting multiple boiling channels to shared inlet and outlet headers, and it needs to be checked explicitly at the network design stage.

Pressure drop reversal changes how manifolds should be sized

In single-phase flow, pressure drop increases monotonically with flow rate, which makes manifold balancing relatively intuitive. In two-phase flow, pressure drop against flow rate is not always monotonic. Under specific conditions, a channel carrying less flow can see a lower pressure drop than one carrying more, which is destabilizing for parallel-channel balancing and can drive uneven flow distribution across cold plates fed from the same manifold, independent of any pump or valve transient.

What this means for design practice

None of these phenomena are unknown. They are extensively characterized in multiphase pipeline and boiling heat transfer literature. What is missing in most data center two-phase cooling designs today is a transient, network-level model that carries this physics through to the manifold and cold plate level, rather than relying on component-level vendor testing that assumes ideal, evenly distributed flow.

Eight years of multiphase flow assurance work on oil and gas networks is, in large part, eight years of learning to distrust steady-state, single-point design for exactly this class of problem. The same discipline applies here.