The rapid development of high-density power electronics has led to remarkably challenging thermal issues. Over the past several decades, transistor development has followed Moore's Law, which states that device sizes decrease exponentially over time. Although advanced power semiconductor technologies have delayed the need for aggressive cooling by several years, the heat flux from the power device has risen significantly, approaching 500 W/cm2. This level is beyond the capability of conventional heatsinks used for silicon-based devices, which can achieve only about 20 W/cm2 when maintaining junction temperatures below 150°C. Therefore, novel technologies must be developed for thermal management. One such technology is microchannel cooling.
The plot in Fig. 1 compares commercially available water-cooled heatsinks, including two microchannel designs, which exhibit dramatically better performance than conventional centimeter-scale channel designs. However, these microchannel heatsinks are limited to small footprints in terms of manufacturability and scalability, thus they have yet to see widespread use in power electronics.
Fig. 2 displays the fractions of the thermal resistivity attributed to different individual layers for a typical module. In the case of microchannel cooling, the largest contributor to the resistivity is the thermal grease layer between the baseplate and the heatsink.
Integrated Microchannel Heatsink
Because of the increased importance of the conductive resistance of the stack, a better solution would be to eliminate some of these layers from the structure. To that end, a team from GE Global Research designed an integrated heatsink with a series of microchannels fabricated directly into the bottom copper layer of the active metal braze (AMB) substrate. This concept, shown schematically in Fig. 3, has advantages in reducing both the convective and conductive resistances of the module.
By using microchannels, the convective thermal resistivity is reduced dramatically. In addition, this stack removes the baseplate solder, the copper baseplate and, most importantly, the thermal grease from the conductive path. As seen in Fig. 2, this results in the elimination of the two largest resistances in the structure. The overall result is a reduction of the total stack resistivity by a factor of two when compared to the best microchannel heatsink in Fig. 1.
An integral microchannel heatsink should be systematically designed for optimum thermal response. This process consists of first-order analysis, detailed 3-D computational fluid dynamics (CFD) simulations and experimental validation.
Step One: Analytical Optimization
First-order heat transfer is governed by the thermal resistivity (R"), which is defined as the temperature rise divided by the heat flux. Note that this is different from the common metric of thermal resistance, which also divides by the device area. This alternate resistivity metric is more appropriate for this situation due to the challenges of high heat flux. For convective heat transfer in channels having hydraulic diameter (DH), the thermal resistivity is calculated as:
In this equation, K is the fluid thermal conductivity, and NuDH is the Nusselt number for the appropriate flow condition. For example, for laminar, fully developed flow in a circular passage with constant heat flux, NuDH = 4.36. In addition, the pressure loss (Δp) is calculated using the friction factor (f) as:
For this expression, is the fluid density and L is the passage length. In laminar, fully developed flow in a circular passage, the friction factor is f = 64/ReDH. Here, ReDH is the Reynolds number (UDH/ν), where U is the fluid velocity and is the fluid kinematic viscosity.
These basic expressions for R" and p can be used to select the optimal channel sizing. Similar to the method described by Knight et al., the passage dimensions are chosen to minimize the thermal resistivity R" subject to pumping constraints on the maximum pressure loss and flow rate. These expressions can be further modified for more complex situations by using the appropriate equations for NuDH and f, such as those for laminar developing and turbulent flows.
The primary variables to be optimized are the channel width and pitch. In GE's integrated microchannel heatsink, these values were altered by changing the number of cooling passages (from 10 to 200) and the ratio of wall thickness to channel width (from 0.1 to 2) beneath a heat source of 2-cm × 2-cm size. Note that these dimensions result in a range of flow conditions, including laminar, turbulent, developing and fully developed regimes. In addition, the channel height in the AMB substrate was varied from 0.05 mm to 0.3 mm, which was the maximum depth allowed due to the thickness of the bottom copper layer. The coolant was water at room temperature, which is the typical coolant used for comparison by commercial heatsink vendors. The pump constraints were specified as 1-gallon per minute (GPM) maximum flow rate and 25-psi maximum pressure loss, which are representative values for power electronics cooling applications.
An additional constraint is to keep the channel width at or above 100 m, given the difficulty of manufacturing passages below that width in copper. This constraint was applied to GE's heatsink.
After analyzing this array of potential channel dimensions, the optimum shape is then selected based on these pump and manufacturing restrictions. For the integrated microchannel heatsink, the preferred shape has a width of 100 µm, a depth of 300 µm and a wall thickness between channels of 100 µm, a configuration consistent with earlier microchannel studies[2,6]. The narrow channel width and small pitch results in a high surface area-to-volume ratio, while the tall channel height tempers the pressure loss through the passage. The calculated thermal resistivity for this design was 0.042 (K)(cm2)/W, comparable to the best published results for existing microchannel heatsinks.
Step Two: Computational Simulation
Although traditional first-order correlations provide a preliminary estimate of the heatsink performance, a more detailed 3-D analysis is necessary to understand the potential response. This is because the previous optimization method assumes 1-D flow through the microchannels, which does not consider entrance, exit and turning effects in the actual heatsink structure. A complete model for a device can be developed using a CFD tool, Icepak, which is a commercial package based on Fluent software. This program allows the study of a variety of shapes and materials, and it includes the effects of conduction, free and forced convection, and radiation. The software can examine both the laminar and the turbulent flow regimes, and it permits steady-state and transient solutions. The relative accuracy of this tool depends on the problem, but it can be accurate to within a few degrees Celsius. To determine absolute temperatures, it is best to anchor the results to experimental data, but the relative values will be fairly reliable.
As an example, consider an integrated microchannel heatsink on which a substrate having four 6-mm × 8-mm silicon rectifier diodes is mounted. In this case, the substrate consists of two 300-µm-thick layers of copper on either side of a 625-µm-thick insulating layer of aluminum nitride (AlN) ceramic. Beneath the four power devices, the bottom copper layer features a series of 65 parallel microchannels of 100-µm width, 300-µm height and 200-µm pitch. This substrate is then bonded to a copper baseplate of 6.35-mm thickness, the interior of the baseplate having a manifold topology that distributes the coolant from an inlet port through the microchannels and out an exhaust port. The baseplate has a footprint of 31 mm × 25 mm. If each silicon diode is assumed to dissipate 150 W, corresponding to 1.5 V and 100 A, the resulting heat flux is 313 W/cm2. This energy is generated at the top of each diode, as the power dissipation occurs at the p-n junction, which is located near the top surface of the device.
This design was simulated with water coolant entering at 20°C at various flow rates. Fig. 4 shows the diode temperature contours at an inlet flow rate of 0.95 LPM. The average diode temperature at this condition was 63.9°C. Based on a grid convergence study, the uncertainty of this model was found to be ±5% in temperature and ±10% in pressure.
At this condition, the diode-averaged thermal resistivity is about 0.14 (K)(cm2)/W from the semiconductor junction to the coolant. After subtracting the conductive thermal resistivity of each layer from the ceramic to the diode, the actual heatsink resistivity is approximately 0.06 (K)(cm2)/W, which is reasonably close to the results obtained from the 1-D analysis.
The next step in the design process is to experimentally validate the simulation results. When the microchannel passages are machined directly into the bottom copper layer of the AMB substrate via laser ablation of the copper, this results in a uniform series of regular trapezoidal passages, which have acute interior angles above 75 degrees, as seen in Fig. 5. While these are not rectangular in profile, the trapezoidal channels should still have heat-transfer performance within 10% of the base profile. This substrate is then bonded to the heatsink base with the 65 microchannels centered beneath the four silicon diodes, resulting in an active cooling area of 2.2 cm2.
Step Three: Experimental Validation
Finally, the thermal performance of the fabricated microchannel heatsink is measured to validate the simulations. For this example, the four diodes were operated in the forward-voltage drop mode, producing constant heat fluxes on the order of 100 W/cm2. The diode top-surface temperatures can be measured using infrared thermography, and these values can be converted directly to thermal-resistivity values. The heatsink should be connected to a chiller that controls the flow rate and temperatures of a water-cooled loop. The pressure losses and coolant temperatures are then measured across the device. Based on the instrument capabilities, the uncertainties in thermal resistivity and pressure loss measurements were ±5% and ±3%, respectively, for GE's integrated microchannel heatsink.
The thermal performance for the GE heatsink was studied for a range of diode power dissipations (0 W to 200 W) and coolant flow rates (0 LPM to 2 LPM). Fig. 6 displays an infrared image of the diode temperature contours for a dissipation of 199 W (104 W/cm2) when cooled by 20°C water at 0.24 LPM. Note that the higher temperatures on the right side are due to power dissipation in the wirebonds and not due to diode variation.
The measured thermal results are compared to several existing commercial heatsinks in Fig. 7. This comparison assumes an identical electrical stack to the GE microchannel heatsink, using silicon diodes attached to an AMB substrate. Because these other heatsinks require a thermal grease or epoxy layer for attachment, a thin layer (75 mm) with relatively high thermal conductivity (9 W/m/K) was assumed. Note that the GE heatsink did not include this layer, as there was no such interface required due to the heatsink design. Because of this benefit, the performance is superior to any of these existing heatsinks. The experimental results demonstrated about 15% lower thermal resistivity than the predicted values, validating the simulations. This discrepancy is partly explained by power-dissipation losses in the experimental connector and wirebonds, as well as by the uncertainty of the Icepak simulation absolute temperatures. Overall, the comparison is quite accurate, clearly demonstrating that the heatsink is very effective. In fact, the overall thermal resistivity of a power module equipped with this heatsink would be less than 0.15 (K)(cm2)/W, resulting in less than 75°C junction-to-coolant temperature rise for the heat flux of 500 W/cm2. This thermal performance is better than any existing heatsink using a comparable material stack.
Moore, G.E., “Cramming More Components onto Integrated Circuits,” Electronics, Vol. 38, No. 8, April 19, 1965.
Tuckerman, D.B. and Pease, R.F.W., “High-Performance Heatsinking for VLSI,” IEEE Electronic Device Letters, Vol. EDL-2, 1981, pp. 126-129.
Wang, B.X. and Peng, X.F., “Experimental Investigation on Liquid Forced Convection Heat Transfer Through Microchannels,” International Journal of Heat and Mass Transfer, Vol. 37, 1994, pp.73-82.
Adams, T. M.; Abdel-Khalik, S.I.; Jeter, S.M.; and Qureshi, Z.H., “An Experimental Investigation of Single-Phase Forced Convection in Microchannels,” International Journal of Heat and Mass Transfer, Vol. 41, 1998, pp. 851-857.
Hetsroni, G.; Mosyak, A.; Pogrebnyak, E.; and Yarin, L.P., “Fluid Flow in Microchannels,” International Journal of Heat and Mass Transfer, Vol. 48, 2005, pp. 1982-1998.
Knight, R.W.; Hall, D.J.; Goodling, J.S.; and Jaeger, R.C., “Heatsink Optimization with Application to Microchannels,” IEEE Transactions on Components, Hybrids and Manufacturing Technology, Vol. 15, No. 5, October 1992, pp. 832-842.
Shah, R.K. and London, A.L., “Laminar Flow Forced Convection in Ducts,” Advances in Heat Transfer, Supplement 1, Academic Press, 1978.
Gnielinski, V., “New Equations for Heat and Mass Transfer in Turbulent Pipe and Channel Flow,” International Chemical Engineering, Vol. 16, 1976, pp. 359-368.
Kakaç, S., Shah, R. K., and Aung, W., Handbook of Single-phase Convective Heat Transfer, New York: John Wiley & Sons, 1987, pp. 68-70.
The Circular Path of the Microchannel Cooling
Twenty-five years ago, Tuckerman and Pease introduced the concept of microchannels to the electronics cooling industry. Their idea was fairly simple. Because heat-transfer coefficients generally increase with decreasing size, the passage size should be made as small as possible. This results in a dense package with higher heat transfer and a larger surface area-to-volume ratio than a conventional cooling device.
However, the benefits are tempered by increased pressure losses with minute passages, in addition to manufacturing challenges. Using traditional fluids analysis, Tuckerman and Pease determined that there was an optimum passage size for realistic pressure differences, selecting a 50-m-wide, 300-m-deep, 1-cm-long passage, which experienced a 30-psi drop with a 0.66 LPM water flow. Through the use of this microchannel, a heated device could dissipate 790 W/cm2 while only experiencing a 71°C temperature rise, as verified by subsequent experimentation. With this remarkable result, it appeared that this would provide a new solution to thermal management of high heat-flux power electronics.
Surprisingly, this improved cooling technology coincided with another important milestone in the history of electronic devices, as Intel released the 80286 microprocessor in 1982. This chip marked the first widely used commercial application of CMOS processing technology in the semiconductor industry, which was thermally significant because these devices only consume power in their switching state. Hence, the power dissipation of typical electronics chips decreased dramatically, and the need for aggressive cooling technology was reduced. As a result, microchannels were largely ignored commercially soon after their invention.
In spite of the lack of interest from industry, microchannels still garnered much academic study in the subsequent decades. The fascination grew particularly strong after a series of papers in the early 1990s raised questions about fundamental fluid dynamics in channels of this scale. These results ranged from unusual laminar-to-turbulent transition to remarkably higher heat-transfer coefficients, which implied that this technology might be even more promising than previously suggested.
A flurry of research was seen around the world, with hundreds of studies each year of these phenomena — even leading to the organization of conferences solely devoted to the subject such as the ASME International Conference on Microchannels and Minichannels. Ironically enough, most of the surprising results were later attributed to experimental errors or faulty assumptions, but this fact did not quell the renewed focus on microchannels. Regardless of these academic pursuits, microchannels were bound to become more prominent once again due to Moore's Law.