Next-generation microfluidic cooling for AI hardware

Babak Baghaei
Nov 18
2 min read

Updated: 5 days ago

New Method for Data Centre Cooling using Microfluidic

The rapid global expansion of AI infrastructure and data-centre capacity is pushing thermal-management technologies toward their limits. As rack densities rise and next-generation AI accelerators draw unprecedented power, traditional air cooling, and even single-phase liquid cooling, is struggling to keep pace. Addressing these emerging constraints requires compact, high-performance, and intelligently controlled cooling architectures capable of operating at chip scale.

New research led by Dr Amir Keshmiri, Founder and Technical Director of Mansim, together with Seyed Hamed Godasiaei and Pouyan Sardari, proposes a breakthrough solution. Published in Scientific Reports (Springer Nature), the study demonstrates a two-phase microfluidic cooling architecture that uses acoustically controlled boiling, nanoarray-engineered micropin fins, and machine learning guided optimisation to stabilise two-phase flow and dramatically enhance chip-level heat transfer.

Why two-phase cooling?

Two-phase cooling, where liquid transitions to vapour directly on the heated surface, provides exceptional heat-removal capability. However, it is notoriously difficult to control. Instability, surface drying, and thermal hotspots have historically limited its practical deployment in data centre environments.

This research tackles those long-standing challenges head-on.

Acoustically driven bubble control

Instead of relying on passive boiling, the system introduces ultrasound-based bubble actuation. By actively influencing bubble formation, detachment and collapse, the architecture:

stabilises boiling behaviour
suppresses local dry out
smooths temperature distribution
delays critical heat flux (CHF) failures

This active control provides the stability that traditional two-phase cold plates lack.

Nanoarray-coated micropin fins

The microchannel cold plate includes nanoarray-coated micropin structures, engineered to:

enhance surface wettability
promote rapid rewetting after bubble departure
amplify capillary liquid refill
increase boiling surface area

Together, these effects create a highly resilient cooling surface capable of handling aggressive heat fluxes representative of next-generation AI chips.

Machine-learning-guided optimisation

To understand which design parameters have the strongest impact on cooling performance, the team applied a suite of machine-learning models:

Deep Neural Networks (DNN)
Long Short-Term Memory (LSTM) networks
SHAP (DeepSHAP) interpretability
Partial Dependence Plots (PDP)
Spearman and Kendall correlation analysis

Among these, the LSTM model outperformed all others, achieving exceptionally low prediction errors (MAE: 0.0055; SMAPE: 0.8; RMSE: 0.0072), making it the most accurate predictor of heat-transfer behaviour.

Machine-learning insights revealed that:

initial temperature is the single most influential factor
chipset material (S30-120, stainless steel) strongly affects heat-transfer coefficient
nanostructure configuration and nanoparticle type (SiO₂, ZnO) significantly modulate bubble dynamics
acoustofluidic excitation is the dominant positive contributor to thermal performance

These interpretability tools provide a data-driven roadmap for engineering future cold-plate designs.

Why this matters for AI, HPC and data-centre engineering

Cooling is becoming one of the defining bottlenecks of modern computing. With AI workloads requiring multi-kilowatt per-chip thermal envelopes, this research points toward a scalable path for the next generation of high-performance cooling: