A Novel Phase-Change Cooling Approach for High-Density AI Chips

Schematic of the acoustofluidic microchannel heat transfer

AI is changing the scale and intensity of modern computing. Larger models, higher accelerator power densities and denser server architectures are placing unprecedented pressure on data-centre cooling systems. The challenge is no longer only how to remove heat from a room or a rack. Increasingly, the limiting physics sits much closer to the chip.

At this scale, thermal management becomes a multiphase flow problem. Heat must be removed from compact surfaces, under high local heat flux, while maintaining stable flow, acceptable pressure drop and reliable contact between coolant and heated surfaces.

This is where Mansim’s R&D work is focused: combining advanced engineering, experimental insight and data-driven modelling to understand the physics behind next-generation cooling systems.

A recent Mansim R&D publication in Scientific Reports, “Hydrodynamic Investigation of Acoustic Bubble–Surface Interactions on Nanoarray-Coated Fins for Microelectronic Cooling with Machine Learning-Based Analysis”, explores one of the key questions facing high-power electronics: can bubble dynamics be controlled to improve phase-change chip cooling?

The study investigates how acoustic excitation, nanoarray-coated micro-pin fins and machine-learning analysis can work together to improve liquid–vapour phase-change cooling performance.

Why AI hardware needs smarter cooling

As AI infrastructure expands, chip-level heat removal is becoming a critical design constraint. High-performance processors and accelerators generate intense local heat loads, and these loads are not always uniform. Hot spots can form rapidly, affecting reliability, performance and long-term hardware stability.

Traditional cooling approaches can still play an important role, but they are increasingly being pushed towards their limits. Larger pumps, higher flow rates or bigger heat sinks are not always practical, particularly when data-centre operators are trying to improve energy efficiency and reduce cooling overhead.

A more effective route is to improve the physics of heat transfer at the surface itself.

Phase-change cooling is attractive because it uses the latent heat of vaporisation to remove large amounts of thermal energy. However, boiling and evaporation are difficult to control. Bubbles can enhance cooling when they nucleate and detach efficiently, but they can also reduce performance if they coalesce, block liquid supply or create unstable vapour regions.

For high-power AI chips, the problem is not simply “more cooling”. It is controlled cooling.

Bubble dynamics as a design variable

In phase-change cooling, bubbles are not just a visual feature of boiling. They are part of the heat-transfer mechanism.

When bubbles form on a heated surface, they can disturb the thermal boundary layer and bring fresh liquid into contact with the wall. This improves local heat transfer. But if bubbles remain attached for too long, merge into larger vapour structures or prevent liquid replenishment, they can increase thermal resistance and trigger dry-out.

The Mansim study investigates this balance by looking at acoustic bubble–surface interactions on nanoarray-coated fins. Acoustic excitation is used to influence bubble motion, while engineered surface structures are used to control where and how bubbles form.

This combination is important. Surface engineering can create favourable nucleation sites, but it is the interaction between the surface, the fluid and the bubble motion that determines whether the cooling system remains stable under load.

Nanoarray-coated fins and phase-change performance

The study examines micro-pin fin surfaces coated with ZnO nanoarrays, including nanorod and nanosheet structures. These coatings change the surface at the micro and nanoscale, increasing the effective surface area and influencing wettability, nucleation behaviour and fluid–surface interaction.

For chip cooling, this matters because the surface is where thermal resistance is often concentrated. A well-designed surface can support more effective bubble formation and removal, helping maintain liquid access to the heated wall.

One of the key results is the strong improvement observed for nanosheet-coated surfaces compared with smooth surfaces. The study reports a 71.2% increase in critical heat flux and a 160.9% improvement in heat transfer coefficient.

This is particularly relevant because critical heat flux and heat transfer coefficient are often difficult to improve together. Increasing nucleation can improve heat transfer, but if bubble behaviour becomes unstable, the system may reach dry-out earlier. The results suggest that combining nanoarray surface design with acoustic bubble control can help manage this trade-off more effectively.

Where machine learning adds value

The physics of phase-change cooling is strongly nonlinear. Heat flux, pressure drop, flow velocity, surface morphology and acoustic forcing do not act in isolation. Their effects depend on thresholds, operating windows and coupled interactions.

This is why Mansim’s R&D approach combines experimental measurements with machine-learning-based analysis. The study uses models including LASSO, Random Forest and Deep Neural Networks to predict heat transfer performance and identify the parameters that matter most.

The Deep Neural Network model achieved strong predictive accuracy, with an R² value of 0.99 and a mean absolute error of 0.01. Random Forest also performed strongly, while LASSO was less effective at capturing the nonlinear behaviour of the cooling system.

However, prediction accuracy is only part of the story. The study also uses interpretability methods such as SHAP, Partial Dependence Plots, Symbolic Metamodeling, Double Machine Learning and TCAV. These tools help connect model outputs back to physical behaviour.

This is important for engineering design. A black-box prediction may tell us what performance to expect, but interpretable modelling helps explain why the system behaves in a certain way.

Design variables that matter at chip level

The analysis shows that heat flux is the dominant driver of bubble dynamics and heat transfer performance. This is expected, but the study also shows that cooling performance depends strongly on how heat flux interacts with pressure drop, surface morphology and velocity.

Pressure drop is especially important because it reflects the hydraulic cost of the cooling strategy. A cooling design that improves heat transfer but creates excessive pressure loss may not be practical for real systems. In data-centre cooling, efficiency matters at every level, from chip thermal resistance to pumping power and system integration.

Velocity also plays a role by influencing bubble residence time and convective transport. Too little flow can allow vapour structures to remain near the surface. Too much flow may increase pressure drop or disrupt favourable boiling behaviour.

Surface morphology adds another layer of control. Nanorod and nanosheet structures influence nucleation density, bubble detachment and local fluid renewal. The study shows that these engineered surfaces are not passive enhancements; they actively shape the thermal-fluid behaviour of the system.

Mansim’s R&D direction

This work reflects the type of R&D Mansim is building around intelligent simulation and advanced engineering.

The cooling challenge facing AI infrastructure cannot be solved through conventional design rules alone. It requires a better understanding of multiphase flow, surface physics, thermal transport and data-driven model interpretation.

Mansim’s role is to connect these areas: using research-grade methods to investigate complex engineering problems, then translating that insight into practical design knowledge for industry.

In this study, that means moving from bubble-scale physics to chip-scale cooling performance. It means using experimental data not only to measure what happens, but to build predictive and interpretable models that support better engineering decisions.

From micro-scale physics to data-centre impact

The AI compute boom is often discussed in terms of energy demand, GPU availability and data-centre capacity. But behind these headline issues is a very physical constraint: heat.

If chips cannot be cooled efficiently and reliably, performance is limited. If cooling systems consume too much energy, operational efficiency suffers. If thermal management is not robust, hardware reliability becomes a risk.

Research into acoustically assisted phase-change cooling and nanoarray-coated surfaces addresses this problem at its source. It looks at the chip-level physics that ultimately influences system-level performance.

The Mansim study shows how advanced surface engineering, bubble control and machine-learning analysis can be combined to improve understanding of high-heat-flux cooling. This does not mean one technology will solve the entire data-centre cooling challenge. But it does show the direction thermal management needs to move: towards more intelligent, physics-informed and optimised cooling architectures.

As AI hardware becomes more powerful, cooling will need to become more sophisticated. The next generation of data centres will depend not only on better infrastructure, but on better control of heat transfer at the smallest scales.