A Generative-AI Cooling Designer Beat the Trade-Off That Haunts Every Chip

The hottest spots inside a modern AI accelerator run so close to their thermal ceiling that a handful of degrees can be the difference between peak performance and a throttled, stuttering chip. Now imagine a computer that doesn't just cool those chips with hand-tuned channels but designs the cooling network itself — generating thousands of bizarre, branching, organic-looking flow paths, keeping the good ones, and throwing the rest away. That's what Michael Acquah and Zheng Liu at the University of Michigan-Dearborn have built, and in a simulated 2.7 kilowatt package holding two GPUs and a CPU, their best machine-generated coolant layout ran the hottest chip 33.6% cooler than a conventional design, with 72.8% less pressure drop, meaning dramatically less pumping power.

The result, posted on arXiv in August 2026, isn't just a clever bit of engineering optimization. It's a demonstration that generative AI — the same family of models that makes images and text — can be pointed at physics and discover cooling architectures a human engineer would be unlikely to ever sketch by handAlexander. And it hints at how the next generation of increasingly hot, tightly packed chips will avoid frying themselves.

The Science

The problem starts with a paradox of modern computing. Chips keep getting more powerful, but they're also being packed ever closer together in multi-chip packages — modules holding two GPUs and a CPU that share a single substrate. More transistors switching in a smaller area means more heat generated per square millimeter. The researchers' target package, loosely modeled on architectures like NVIDIA's GB200 module, dissipates 2,700 watts: each GPU burns 1,200 W and the CPU another 300 W, crammed into a cold plate measuring roughly 190 by 270 millimeters (Acquah & Liu, 2026).

Air cooling simply can't carry that heat away. So engineers turn to liquid cooling — channels routed directly above the chips through which water flows. But here's the catch: the shape of those channels determines everything. And there are effectively infinite shapes.

Classical approaches to designing cooling channels fall into predictable patterns — straight channels, wavy channels, branching manifolds. Human designers start from a known template, then tweak it. But the space of possible channel topologies is enormous lasughan an engineer can reason about, and the objectives conflict: you want to remove heat evenly (low maximum temperature, low temperature spread), but you also want low pressure drop, because pressure drop is pumping power and energy. Maximizing cooling often means squeezing coolant through narrow, tortuous paths that cost enormous hydraulic pressure.

The radical move here is to let a machine invent the topology. The researchers used a conditional diffusion model — the same deep-learning architecture behind modern image generators — trained to produce binary "flow masks" (each pixel either fluid or solid). The model generates the left half of the channel layout, then mirrors it across the package centerline to enforce symmetry and guarantee equal cooling for the two mirrored GPUs (Acquah & Liu, 2026). As it generates, the model is conditioned on target performance: the user can steer it toward layouts predicted to achieve a certain maximum GPU temperature, a certain temperature spread, a certain pressure drop (

).

But generating a plausible-looking channel is only the first step. A pretty picture isn't a working cooler. So every generated layout had to pass a brutal physics gauntlet.

The researchers generated 5,000 candidate layouts. Of these, only 2,220 — about 44% — contained a continuous path from the inlet to the outlet that water could actually flow through. Many diffusion-generated channels had dead ends, disconnected islands, or no through-flow at all. Of those 2,220 connected designs, only 229 survived the final topology-screening criteria, which rejected layouts with excessive dead-end branching and other features that would be physically or manufacturably problematic (Acquah & Liu, 2026). Nearly 96% of the machine's first drafts were unusable. That's expected — generative models are broadly proposing, and the physics filter is doing the hard selecting.

The 229 survivors then had to be evaluated. Running full three-dimensional computational fluid dynamics on every candidate would be prohibitively expensive — each full simulation is a heavy lift. So the researchers built a reduced-order model (ROM), a fast approximate physics solver that captures the dominant mechanisms: heat spreading through the copper and silicon layers, depth-averaged coolant flow, local convective heat transfer, progressive warming of the water as it travels downstream, and hydraulic resistance (Acquah & Liu, 2026). This 2.5D model runs on a 200×200 grid and solves the coupled conduction-flow-energy problem in seconds rather than hours.

The ROM isn't perfect — it's calibrated with two global correction coefficients. The thermal coefficient fitted to match reference cases produced a mean temperature-rise error of about 7.7% and a maximum error around 12.6%. But the key insight is that the ROM doesn't need to be exact; it needs to be consistently good at ranking designs relative to each other, so the best candidates float to the top. Only the final champion is subjected to the expensive, accurate full-order CFD.

What They Found

After screening and evaluation, the researchers had 229 feasible designs, each scored on three competing objectives: maximum GPU temperature $T_{\max ,\mathrm{GPU}}$, GPU temperature spread $\Delta T_{\mathrm{GPU}}$, and pressure drop $\Delta P$. These objectives genuinely fight each other — a layout that cools hot spots brilliantly might demand huge pressure, while a gentle low-pressure design might leave hot spots burning.

Multi-objective analysis revealed a trade-off frontier, the set of designs where improving one metric necessarily worsens another. From this Pareto front (

), the researchers selected a design designated G1016, which balanced all three objectives best.

G1016's predicted performance: a maximum GPU temperature of 70.30 °C, a GPU temperature spread of 24.90 °C, and a pressure drop of 89.72 kPa (Acquah & Liu, 2026).

Those numbers become meaningful only in comparison with the conventional reference design — a simple, straight-channel topology that a reasonable human engineer might well build. Against that baseline, G1016 cut the maximum GPU temperature by 33.6%, slashed the temperature spread by 52.5%, and reduced pressure drop by an astonishing 72.8% (Acquah & Liu, 2026). The generated channel simultaneously cooled the hottest chip far better, made the temperature across the GPUs much more uniform, and demanded far less pumping power. Improvements in all three objectives at once — something conventional wisdom often says is impossible because of the thermal-hydraulic trade-off.

Machine-generated cooling cuts hot-spot temperature by a third

Label	Value
Conventional reference (S1)	105.9
Generated G1016	70.3

Source: Acquah & Liu (2026), arXiv. ROM-predicted maximum GPU junction temperature.

But a fast approximate model predicting great numbers isn't proof. So the researchers took G1016, reconstructed it as a full three-dimensional geometry, and ran high-fidelity conjugate heat-transfer simulations in OpenFOAM — the expensive, "ground truth" CFD that resolves the real flow, turbulence, and heat exchange in detail (

). The independent simulation predicted a maximum GPU temperature of 66.70 °C and a pressure drop of 92.1 kPa. Compared to the ROM's predictions, that's a difference of roughly 8.6% in temperature rise and 2.6% in pressure drop (Acquah & Liu, 2026). The fast model was not only fast — it was faithful, at least for the design that mattered most.

A 72.8% cut in pressure drop — and pumping power

Label	Value
Pressure drop (kPa)	330
Pressure drop (kPa)	89.72

Source: Acquah & Liu (2026), arXiv. ROM-predicted inlet-to-outlet pressure drop.

Why This Changes Things

The most striking thing about this work isn't the specific numbers — impressive though they are — but the workflow it validates. The researchers found a machine architecture that generates cooling-channel designs, screens out the unusable ones, ranks the rest with a cheap physics surrogate, and reserves the expensive high-fidelity simulation for just the final winner. That's a fundamentally different economic model for engineering design.

For context, consider how the cost of evaluating designs scales. Full 3D CFD of a single multi-chip package is computationally demanding enough that you'd only run it on a handful of hand-picked candidates. The traditional approach — human invents a few layouts, CFD-validates them, picks the best — explores a tiny slice of design space. This generative framework, by contrast, can cheaply explore 5,000 designs and pick from 229 feasible ones summary - it samples design space millions of times more broadly for roughly the same compute budget (Acquah & Liu, 2026).

There's a deeper conceptual lesson too. The trade-off between thermal performance and hydraulic resistance has long been treated as a law of nature in single-phase cooling: to cool better, you pay more pressure. This framework found layouts that broke that assumed trade-off in a multi-objective sense — improving cooling and drastically reducing pressure drop. The reason is that the machine discovered channel topologies with genuinely different geometric logic than human-designed straight or wavy channels — organizing flow so it spreads efficiently across both GPU regions while avoiding tortuous, high-resistance paths. The dead-end-branch screening and connectivity checks show that the machine's best work isn't just aesthetically organic; it's physically and hydraulically sensible.

The pressure-drop result deserves particular attention. A 72.8% reduction in pressure drop at constant flow rate translates directly into a 72.8% reduction in required pumping power for those channels. In data centers, where cooling can account for a substantial share of total energy consumption, that kind of hydraulic efficiency is not a footnote — it's a sustainability win. Lower pumping power means less energy, less heat dumped into the facility, and lower operating costs.

This also matters for the relentless march of chip packaging. As advanced packages push toward and beyond the kilowatt scale — and as 3D-stacked chiplets concentrate heat in ever-smaller footprints — thermal management is becoming the binding constraint on performance. Cooling-to-air is no longer enough; liquid cooling is becoming mandatory. But adopting liquid cooling is one thing; routing it well is another. This framework attacks exactly that problem: how to put coolant exactly where the heat is, in complex multi-source packages where hot spots are distributed and asymmetric.

What's Next

The work is, by the authors' own acknowledgment, a demonstration of a framework on a simulated package — not a proven, manufactured, tested-in-hardware cooling system. Several questions remain open.

First, the caveat of calibration. The ROM is fitted to reference cases with global coefficients, and while it predicted G1016 well within about 8.6% on temperature, a 12.6% maximum error on calibration cases means the ranking of designs near the Pareto front could be slightly imperfect. A design that the ROM ranks third might actually outperform the selected one in full CFD. The framework is robust enough to find a good champion, but not yet proven to find the optimal one.

Second, manufacturability. The generated topologies the machine calls "feasible" are feasible in the physics sense — connected, no dead ends, minimum channel widths. But translating an organic, branching channel mask into a manufacturable cold plate — machined, etched, or 3D-printed — is a nontrivial engineering step that this paper doesn't fully address. Real manufacturing imposes constraints (tool sizes, minimum feature widths, material anisotropy) that the current screening only partially captures.

Third, the ground truth gap. Only one design was validated with full 3D CFD. A stronger claim would come from validating several designs across the Pareto front, or from physical benchtop experiments with a fabricated prototype. The paper is a strong computational proof of concept; the hardware test is the natural next step.

Perhaps most excitingly, though, the framework itself is extensible. The performance conditions the diffusion model is conditioned on — maximum temperature, temperature spread, pressure drop — could be swapped for other objectives: pumping power, coolant outlet temperature, thermal cycling reliability, even carbon footprint of the cooling solution. The diffusion model could be trained on more diverse packages, or conditioned directly on geometry of a specific chip layout. And the whole pipeline — generate, screen, rank with cheap physics, validate the winner with expensive physics — is a template that generalizes well beyond cooling, to any engineering design problem where high-fidelity evaluation is costly and the design space is vast.

The deeper significance is a shift in how engineers will think about designing physical systems. For decades, the bottleneck in engineering design has been human imagination plus high-fidelity simulation. This work shows a future where a generative model proposes, a fast physics surrogate ranks, and only the finalists are subjected to the full cost of truth. The computer isn't replacing the engineer — it's vastly expanding what the engineer gets to choose from. And in a world where chips are getting hotter and energy is getting more precious, letting machines dream up better ways to cool them isn't just clever. It may be necessary.