On Numerical Predictive Accuracy for Electronic Component Heat Transfer in Forced Convection

Transcription

1 On Numerical Predictive Accuracy for Electronic Heat Transfer in Forced Convection Valérie Eveloy 1,2, Peter Rodgers 1, and John Lohan 2 1 Electronics Thermal Management, Upper Quay, Westport, Ireland. 2 Galway-Mayo Institute of Technology, Department of Mechanical & Industrial Engineering, Dublin Road, Galway, Ireland. rodgersp@eircom.net Abstract Numerical predictive accuracy is assessed for multi-component Printed Circuit Board (PCB) heat transfer in forced convection using a code dedicated to the thermal analysis of electronic systems. This is achieved by comparing numerical predictions with experimental component junction temperature measurements for an in-line array of fifteen, PCB-mounted 160-lead Plastic Quad Flat Pack (PQFP) components. Test case complexity is increased in controlled steps to enable the impact of both aerodynamic conditions and component thermal interaction on predictive accuracy to be assessed. Where highly disturbed flow conditions existed, component junction temperature predictive accuracy is overall within ± 10 C or ± 20%, and component thermal interaction is shown not to be fully captured. However, if the flow is steady with low upstream disturbance, predictive accuracy improves to ± 5 C or ± 10%. The flow field over the board is experimentally visualised to help explain prediction errors in aerodynamic sensitive regions. Keywords: numerical analysis, computational fluid dynamics, CFD, benchmark, calibration, validation, prediction accuracy, component modelling, component heat transfer, electronics cooling, thermal management, reliability, virtual prototyping, flow visualisation. Introduction In absence of time for comprehensive prototyping, diminishing product development cycles have placed greater reliance on the use of numerical analysis to predict electronic system thermal performance. However, thermal designs produced by Computational Fluid Dynamics (CFD) analysis still require experimental verification due to inherent limitations in the code numerics and low-order turbulence models used [1]. Furthermore, at the final design stage, it is generally accepted that the accuracy of component junction temperature predictions needs to be within ±3 C or ±5% of measurement for these predictions to be used as boundary conditions in subsequent product reliability and electrical performance analyses [2]. A stringent need therefore exists to comprehensively evaluate the predictive accuracy of CFD codes dedicated to the thermal analysis of electronic systems so as to; firstly, better anticipate the extent of concurrent experimentation required during the various design phases, and secondly, determine whether component junction temperature predictions may be used in subsequent reliability and performance analyses. To achieve these objectives, CFD predictions should be assessed against accurate benchmark data for a range of different thermofluid systems that span from component to system level. This study focuses on component-printed Circuit Board (PCB) heat transfer and a review of previous benchmark studies in this area is given in [1,3-6]. Recently, Eveloy et al. [7] reported component junction temperature prediction errors of up to ±11 C (25%) for the analysis of the multi-component PCB shown in Figure 1. This PCB, which is also the focus of this study, is populated with an inline array of fifteen, equally spaced 160-lead PQFPs, and was characterised in a wind tunnel environment at 2 and 5.5 m/s freestream air velocities. To generate a higher degree of component thermal interaction present in double-sided PCB applications, an adiabatic boundary condition was imposed along its non-component side using a 50 mm thick, thermally-insulating Styrofoam block. The geometry of this block, represented in Figure 3, was machined to a smooth, uniform, semi-elliptical profile so that leading edge flow separation over the PCB and its impact on component thermal resistance might be minimised. However, as subsequently shown by flow visualisation presented in a later section, this insulating block actually induced a complex, unsteady flow field over the PCB leading edge at 2 m/s. While significantly increasing the difficulty to numerically model this flow field and hence component-pcb heat transfer, the induced flow disturbance serves to mimic those encountered in electronic systems, but retains welldefined boundary conditions for numerical modelling. This flow disturbance therefore permitted predictive accuracy to be assessed for system level flow conditions that are more realistic than uniform free-stream conditions. To better understand sources of component junction temperature prediction errors, the analysis reported by Eveloy et al. [7] is extended in this study by assessing predictive accuracy for the noninsulated PCB, over which less complex, steady flow conditions exist. Based on both test PCB configurations a novel, methodical approach is employed, whereby test case complexity is incremented in controlled steps from; i) individually powered components on the non-insulated PCB, where only one component is powered at a time, ii) individually powered components on the insulated PCB, to iii) all fifteen components simultaneously powered on the insulated PCB. The step from cases (i) to (ii) 1

2 enables the impact of aerodynamic conditions on junction temperature prediction accuracy to be quantified, while stepping from (ii) to (iii) permits the ability of the code to predict component thermal interaction to be assessed. For all test cases measured component junction temperature is used as the primary benchmark criterion. In addition, the flow field over the board is experimentally visualised to help identify aerodynamic sensitive regions on the PCB, thereby permitting, in some instances, sources of prediction error in component junction temperature to be further investigated. Objectives of the work presented in this paper are therefore to assess the applicability of CFD analysis to predict multi-component PCB heat transfer in forced air flows approximating those encountered in electronic systems. This is achieved by investigating; (i) the ability of different flow models to predict component-pcb heat transfer in forced air flows that exhibit significant flow disturbance, and (ii) the significance of component junction temperature prediction errors associated with component thermal interaction, present on many densely populated PCBs. X Note: PCB size = 233 x 160 x 1.6 mm. The position of each component on the PCB is identified by the lettering, A to O. A, F and K are leading edge components. Axis X-X refers to the PCB central, stream-wise axis. Figure 1 Multi-component thermal test PCB [8]. Experimentation Air flow direction junction temperature measurements; The measurements used to assess predictive accuracy were reported by Lohan & Davies [8], who undertook an experimental investigation into component thermal interaction on the multi-component PCB shown in Figure 1. -PCB architectures and experimental details are described by Eveloy et al. [7]. The test PCB was a 1.6 mm thick FR-4 design of double Euro Card standard size 233mm x 160 mm, with one-ounce copper tracking on both sides. The components were thermally enhanced PQFP packages, Figure 2, having either an embedded 18 mm square or a 20 mm diameter, circular heat slug. location on the PCB was identified by the lettering A to O, Figure 1. Junction temperature was measured using thermal test dies calibrated to an accuracy of ±0.4 C. In all cases component power dissipation was 3 Watts. Uniform free-stream air velocities were generated using a wind tunnel [8]. Air flow visualisation; The air flow over both the insulated and non-insulated multi-component-pcbs was visualised using a smoke-wire method described by Rodgers et al. [1,5]. This method X X Y permitted the flow streamlines above the PCB surface to be visualised, thereby helping to identify aerodynamically sensitive regions, where large prediction errors are most likely to occur. In this instance, the 0.18 mm diameter nichrome wire was placed 4 mm upstream of the PCB leading edge and flush with its surface. The smoke was produced by burning Dantec s Safex standard fog fluid [9] from the heated wire. The visualised flow was recorded over a five-second period using a Sony Cybershot digital video camera. Though undertaken at both 2 and 5.5 m/s, air flow visualisation is only presented for the lower velocity to highlight unsteady flow features over the insulated PCB All dimensions in mm Thermal test chip x 40 Leads Plastic package Copper heat slu Figure lead PQFP component geometry with a square heat slug design [8]. Numerical analysis Numerical analysis was undertaken using Flotherm, Version 3.1, a code widely used within the industry for the analysis of electronics cooling. The computational method is given in [10]. Due to computational constraints, Eveloy et al [7] analysed the multi-component PCB using a half geometry model, Figure 3(a), by taking advantage of the symmetry of the heat transfer processes in the span-wise direction about the PCB central stream-wise axis, as experimentally characterised by Lohan & Davies [8]. This modelling approach was appropriate as instances of prediction error greatly exceeded the measured asymmetry in component junction temperature. However in this study, improved computational facilities permitted the complete board geometry to be modelled, in an attempt to more accurately quantify the prediction errors highlighted by Eveloy et al. [7]. The component and PCB modelling methodologies are based on Rosten s et al. [11] approach, with minor alterations described in Eveloy et al. [7]. All dimensions and constituent material thermal properties correspond to nominal vendor specifications listed in Table 1, with the exception of FR-4 thermal conductivity [12]. The modelling methodology was validated in Eveloy et al. [7] by analysing a less complex case, consisting of a single component mounted at position H on the non-insulated PCB. For both 2 and 5.5 m/s free-stream air velocities, component junction temperature prediction accuracy was found to be within ±5% of measurements. This accuracy permitted the analysis to be extended to the populated PCB with confidence, on the premise that any significant decay in predictive accuracy would probably be associated with potential limitations of the CFD code. The complete model is shown in Figure 3(b) for the insulated PCB case, with the detailed component model represented in Figure 3(c). As infra-red measurements of PCB surface temperature showed negligible thermal interaction between the fully powered PCB and the test assembly mechanical support, made of lowconductivity Tufnol material, this support was modelled as nonconducting. 2

3 Table 1 - Vendor specified thermal conductivity values for component and PCB constituent elements. Element Thermal conductivity (W/m.K) Density (kg/m 3 ) Specific heat capacity (J/kg.K) Encapsulant Die (T-100) Die attach Leadframe Heat slug Leadframe insulation PCB substrate k ip = 0.81, k tp = PCB copper tracking Styrofoam insulation Note: T = temperature in C. k ip and k tp are in-plane and through-plane thermal conductivities values respectively. The computational domain was confined to the fluid domain in the vicinity of the PCB to permit the computational grid to be effectively used to focus on the resolution of the component-pcb thermofluids. The following artificial boundaries were imposed at a sufficient distance from the PCB assembly so that no significant unintentional elliptical effects were introduced. A uniform mass flow inlet boundary was applied upstream of the PCB leading edge, and an outlet vent downstream of the PCB trailing edge. Freestream boundaries were applied parallel to the PCB componentand non-component sides, and its stream-wise edge. Their location relative to the PCB are defined in Table 2 for both the noninsulated and insulated models. A free-stream boundary was also applied coincident with the base of the mechanical support. These free-stream boundaries fixed the relative pressure to zero with any incoming air entering at both the prescribed ambient temperature of 20 C and inlet free-stream velocity. The overall computational domain size is given in Table 3 for both the non-insulated and insulated models. In the absence of a dominant length scale for describing the fluid flow regime in non-dimensional form, the fluid domain was solved using both the laminar, and two-equation (k-e) eddy-viscosity turbulence model with the revised wall function formulation. Based on flow visualisation, the flow field was solved as steady for all analyses, with the exception of the insulated model at 2 m/s. For this test case the flow field was modelled as unsteady. Variable fluid property treatment was applied. Radiative heat transfer was modelled from the component top and bottom surfaces, FR-4 substrate and copper tracking surfaces on the PCB component-side. Die Symmetry plane (a) Insulated PCB: half geometry [7]. Inlet (fixed mass flow) Mechanical support PCB connector Inlet (fixed mass flow) Insulation (b) Insulated PCB: complete geometry. Heat slug Note: Unmarked computational domain boundaries are free-stream boundaries. Gravity vector acts in (-y) direction. Track PCB Outlet (vent) Note: Unmarked computational domain boundaries are free-stream boundaries. Gravity vector acts in (-y) direction. Track Insulation Leadframe PCB Outlet (vent) External leads (c) 160-lead PQFP component with square heat slug design. Figure 3 Numerical model. 3

4 Table 2 - Location of applied artificial boundary conditions relative to the PCB for numerical analysis. PCB model Inlet Outlet side Free-stream boundaries Non-comp. / insulation side Stream-wise edge Non-insulated Insulated Note: All dimensions are in mm. Table 3 - Computational domain size and grid discretization detail for the multi-component PCB numerical model. Non-insulated model Insulated model x y z x y z Domain size (mm) Computational grid body grid Note: x, y and z co-ordinates refer to the stream-wise, span-wise and transverse directions respectively, Figure 3(b). A non-uniform grid was applied having highest density both within the component bodies to resolve conductive heat spread, and in the near-wall regions, to resolve the high velocity and temperature gradients within the hydrodynamic and temperature boundary layers respectively and thus their near-wall effects on both surface friction and heat transfer. This generic grid permitted the application of both laminar and turbulent flow models, for both free-stream air velocities analysed. The grid was refined until it captured the main flow features anticipated in bluff-body flows, namely: leading edge separation, downstream reattachment and flow recirculation, as these may impact on the prediction of both near-wall fluid temperature and heat transfer coefficient. The overall computational domain grid volumes are given in Table 3 for both the non-insulated and insulated models. Both models were verified to produce solutions that were independent to both grid and computational domain size, for either flow model applied and both free-stream air velocities analysed. Solution convergence was defined when the residual error sum for each variable was reduced to the termination error level, which was set to the software default setting of 0.5%. However using the laminar model for the insulated PCB case at 2 m/s, the residual error sums for pressure and x-velocity could only be reduced to within 6 % and 2% respectively of their termination error levels. The field errors for pressure were found to be located in the wake flow downstream of the board. As grid aspect ratios were maintained within satisfactory ranges to minimise convergence difficulties, it was suspected that the downstream domain outlet was not positioned at a sufficient distance from the PCB trailing edge to fully resolve the transient wake region. However, increasing the stream-wise length of the domain was computationally unfeasible as the grid volume exceeded 3.4 Million cells. Instead, to assess if convergence could be improved, a simplified PCB model was used, where both the component internal architecture and PCB copper tracking details were not modelled. As the software uses a structured grid, this permitted the superfluous grid detail that would otherwise be required to maintain low grid aspect ratios to be redeployed, thereby enabling the domain stream-wise length to be extended by 150 mm. However, convergence for pressure was only improved by 2%, while temperature predictions for the simplified components remained unchanged relative to corresponding predictions for the smaller domain. On this basis, it was concluded that the wake flow was sufficiently resolved so as not to adversely impact on component junction temperature predictive accuracy. Computation was performed using a DELL Precision 420 workstation with dual 1 GHz Pentium III processor and 1024 MB RAM, operating on Windows 2000 Professional. Results and discussion Before assessing component junction temperature prediction accuracy, the visualised air flows over the multi-component PCB are presented to identify aerodynamically sensitive regions on the board. This will aid, in some instances, sources of prediction errors to be identified. Air flow visualisation Flow patterns visualised using smoke streamlines from a heated nichrome wire at 2 m/s are presented for the non-insulated and insulated PCBs in Figures 4 and 5 respectively. Views of the flow fields are shown from both the front view in Figures 4 and 5(a), and in plan view in Figures 5(b) and 5(c). A comparison of the flow patterns over the PCB front surface in Figures 4 and 5(a) highlights the significant influence of the thermally insulating block on the flow field. The tightly packed streamlines over the non-insulated PCB, Figure 4, tend to flow closely to the component-pcb surfaces near the leading edge. This is particularly evident both in the regions between the components, and for the streamlines that impact close to the front face corners of the leading edge components and sweep inwards over their top surface as they flow downstream. However, for the same wire position in the insulated case, Figure 5(a), the leading edge streamlines no longer follow the component-pcb contours. Instead, the flow separates upstream of the PCB leading edge and re-attaches in a region just downstream of the leading row components A, F and K. This feature, which was not evident on the non-insulated PCB, can be clearly seen from the plan views of a single streamline over the central F component in Figures 5(b) and 5(c). As both these images were taken from the same image sequence, but 230 ms apart, it is also obvious that the location of the re-attachment point varies, indicating unsteady flow characteristics. Analysis of many such image sequences revealed that this flow phenomenon was non-periodic and fluctuated randomly at frequencies between 3 and 9 Hertz. While a similar, but thinner separation did occur at 5.5 m/s, the unsteady behaviour was not evident. It was concluded therefore that localised characteristics of the flow over the insulated PCB at 2 m/s were unsteady, and that the strong re-attachment zone immediately downstream of the leading row components could pose problems for modelling heat transfer in this region. Predictive accuracy While there is no consensus in the literature on maximum permissible prediction error for component junction temperature, the authors propose the following guidelines. During the early design phase, to select the thermal management strategy, a predictive accuracy within ±10 C or ±20% of measurement is typically acceptable. For the refinement of the thermal design by parametric analysis, a more stringent criterion of ±5 C or ±10% would apply. Ultimately, most thermal design processes in the final design phase would aim for component junction temperature predictions to be within ±2 C or ±5% of measurement as this variable forms a critical boundary condition for both life-cycle costing and reliability calculations, and concurrent numerical Electrical Design Analysis (EDA) of circuit performance, which 4

5 Note: Smoke introduced 4 mm upstream and flush with the PCB surface. Figure 4 - Experimentally visualised flow field on the non-insulated PCB at 2 m/s. (a) Front view. (b) Plan view still 1. (c) Plan view still 2. PCB insulation Note: Time lapse between stills 1 and 2 is approximately 230 ms. Smoke introduced 4 mm upstream and flush with the PCB surface, and in plan view, aligned with the central stream-wise axis of component F. Figure 5 - Experimentally visualised flow field on the insulated PCB at 2 m/s. may be highly temperature dependent. In agreement with the above criterion, Mack & Venus [13] note that temperature prediction accuracy needs to be within ±10% at an intermediate product design phase, and for reliability analysis, Lasance [2] recommends an accuracy of ±3 C. junction temperature prediction accuracy is categorised based on these guidelines, and is therefore presented both as an absolute temperature error ( C), and percentage value. To isolate the impact of aerodynamic conditions on predictive error, numerical accuracy is assessed for individually powered components on both the non-insulated and insulated PCBs in Tables 4 and 5 respectively. The fully powered, insulated PCB case serves to assess the ability of the code to predict component thermal interaction using the results in Tables 6 and 7. Individually powered components; Predictive accuracy is firstly assessed on the non-insulated PCB. In Table 4 at 5.5 m/s, using either flow model predictive accuracy (±3 C) would qualify for reliability predictions, when account is taken of experimental uncertainty. However at 2 m/s, predictive accuracy for the leading edge components (A, F, K), and G is outside this band for both flow models, indicating that the leading edge flow phenomena visualised in Figure 4 is inaccurately captured. Table 4 - Comparison of measured and predicted component junction temperatures for individually powered components on the non-insulated PCB. Prediction discrepancy ( C) A (14.6%) -7.9 (16.3%) (5.6%) +0.6 (2.0%) K (15.9%) -8.3 (17.3%) (5.4%) +0.4 (1.3%) F (13.2%) -7.0 (14.9%) (6.4%) +0.7 (2.3%) G (12.0%) +3.0 (6.8%) (11.2%) +2.7 (8.4%) H (0.4%) -1.4 (3.0%) (1.1%) -0.8 (2.3%) I (0.2%) +0.4 (0.9%) (0.3%) +0.6 (1.8%) J (6.2%) -2.6 (5.6%) (0.9%) -1.2 (3.6%) Note: A, F, K are leading edge components. Measurement accuracy, ±0.4 C. component junction temperature rise above ambient air temperature. power dissipation = 3W. Ambient air temperature = 20 C. For the insulated PCB, predictive accuracy at both air flow velocities decays relative to the non-insulated case, Table 5. This decay clearly indicates a weakness in the code to capture the more complex aerodynamic conditions induced by the elliptical block. At 5.5 m/s, the k-e flow model displays overall slightly better accuracy than the laminar model, suggesting its better applicability for this flow. It is acknowledged that the k-e model is not suited to the analysis of the unsteady flow over the insulated PCB at 2 m/s. This is reflected in the time-invariant flow field predictions above component G at 2 m/s in Figure 6. The laminar model, however, does predict flow unsteadiness, as shown in Figure 6. It was noted that this unsteadiness was only captured for a grid volume above 2 Million cells. The time invariance of the k-e predictions possibly results from an overprediction of the turbulent viscosity damping out any transient flow features. This hypothesis is in line with the fact that the k-e model yielded the same component junction temperature predictions when solving the flow field as steady. However, the k-e model was assessed for this case to reflect normal design scenarios, where no a priori knowledge of the flow regime exists, and whether it is steady or unsteady. Though the k-e model 5

6 predictions should therefore be considered with scepticism, it yields better accuracy than the laminar model at 2 m/s. of component thermal resistance with increasing free-stream velocity. Velocity (m/s) Laminar, 1.5 mm Laminar, 6 mm Laminar, 12 mm k-e, 1.5 mm k-e, 6 mm k-e, 12 mm Fully powered insulated PCB; When account is taken of measurement uncertainty, prediction accuracy in Table 6 ranges from ±1 C to ±12 C (up to 30%) depending on component location and flow model. Reflecting more complex, but realistic flow conditions, these errors exceed those reported by Rodgers et al. [1,5,14] and Eveloy et al. [6]. The results from these studies combined therefore clearly indicate that component junction temperature needs to be experimentally measured to be used as a boundary condition in reliability analyses. Neither flow model yields best predictive accuracy for all components. For example, the k-e and laminar models more accurately predict the junction temperatures of the first and second span-wise row components respectively. These deviations are attributed to significant aerodynamic disturbance in this region of the board, to which the flow models display different sensitivities Note: Distance given is the transverse location of flow field monitoring point above the centre of the package top surface. Figure 6 Numerically predicted flow unsteadiness in the stream-wise direction above component G. Table 5 - Comparison of measured and predicted component junction temperatures for individually powered components on the insulated PCB Time (s) Prediction discrepancy ( C) A (21.3%) -9.5 (17.6%) (12.7%) -3.7 (10.0%) K (18%) -8.1 (15.2%) (14.5%) -3.5 (9.4%) F (21.4%) -9.6 (17.7%) (10.8%) -0.8 (2.3%) G (14.3%) -1.1 (2.0%) (10.2%) +1.6 (4.1%) H (16.8%) -7.1 (12.4%) (12.3%) -4.8 (11.1%) I (17.5%) -8.3 (14.6%) (10.6%) -3.9 (9.6%) J (17.2%) (18.2%) (12.4%) -6.0 (14.9%) Note: A, F, K are leading edge components. Measurement accuracy, ±0.4 C. component junction temperature rise above ambient air temperature. power dissipation = 3W. Ambient air temperature = 20 C. The air flow at 2 m/s is solved as unsteady. Using the laminar model, the decay in predictive accuracy relative to the non-insulated case at 2 m/s is most pronounced for the downstream components (H, I, J). It is suspected that this decay may be coupled with inaccurate prediction of upstream flow conditions, as numerical sensitivity checks showed that it was not related to the modelling of the wake flow downstream of the PCB. The flow visualisation in Figure 5 shows that a complex, unsteady flow field exists over the insulated PCB s two leading edge component rows, which resulted in predictive accuracy also decaying in this region. Overall, predictive accuracy would only be sufficient in the early design phase at 2 m/s, but satisfactory in the intermediate design phase at 5.5 m/s. This difference reflects the decreasing sensitivity Table 6 - Comparison of measured and predicted component junction temperatures for the fully powered, insulated PCB. Prediction discrepancy ( C) A (21.4%) -9.0 (15%) (14.8%) -4.0 (10.1%) B (2.0%) +5.1 (9.3%) (13.4%) +7.9 (23.4%) C (4.4%) -0.3 (0.5%) (15.5%) +5.1 (13.6%) D (3.8%) -2.8 (4.2%) (26.5%) +4.9 (12.7%) E (4.0%) -4.5 (7.0%) (31.5%) +4.0 (11.0%) F (18.5%) -4.6 (7.5%) (9.1%) +2.0 (5.4%) G (3.1%) +8.4 (13.1%) (1.9%) +8.2 (19.6%) H (3.8%) -0.6 (0.8%) (2.5%) +2.0 (4.2%) I (1.4%) -3.2 (4.2%) (10.7%) +1.5 (3.1%) J (1.7%) -5.9 (7.9%) (12.5%) -0.6 (1.3%) K (14.2%) -3.1 (5.4%) L (10.4%) (24.2%) (13.7%) (30.6%) M N (10.1%) +5.2 (8.3%) (25.4%) +7.1 (18.6%) O (8.8%) +1.4 (2.2%) (23.2%) +3.3 (8.4%) Note: A, F and K are leading edge components. Measurement accuracy, ±0.4 C. component junction temperature rise above ambient air temperature. power dissipation = 3W. Ambient air temperature = 20 C. No junction temperature measurements for component M, and K at 5.5 m/s were available. The component junction temperature rise between the individually- and fully powered configurations, that is its temperature rise due solely to component thermal interaction, is significantly overpredicted in almost all cases, Table 7. This clearly shows a weakness of the code to predict downstream component thermal interaction. This is in line with Anderson s results [15], who found incorrect prediction of downstream fluid flow mixing using the same CFD code. Overprediction of the temperature rise therefore effectively compensates for the underprediction of the individually powered component junction temperatures, Table 5. Therefore, the predictive accuracy obtained for the fully powered case in Table 6 are only net values, and a function of component power dissipation. 6

7 Apart from highlighting predictive errors, it is worth noting that the overall computational grid volume required in this study would be considered impractical in a design environment. Table 7 - Comparison of measured and predicted component junction temperature rise between the individually- and fully powered configurations on the insulated PCB. Prediction discrepancy ( C) A (22%) +0.5 (8%) (41%) -0.3 (10.0%) K (40%) +5.0 (132%) F (2.7%) +5.0 (68%) (10.3%) +2.8 (97%) G (56%) +9.5 (93%) (178%) +6.6 (244%) H (43%) +6.5 (41%) (144%) +6.8 (151%) I (46%) +5.1 (27%) (134%) +5.4 (77%) J (48%) +4.5 (25%) (174%) +5.4 (87%) Note: A, F, K are leading edge components. Measurement accuracy, ±0.4 C. component junction temperature rise between the individually and the fully powered configurations. power dissipation = 3W. Ambient air temperature = 20 C. Summary and conclusions This study has given an insight into numerical heat transfer predictive accuracy for a practical, forced air-cooled multicomponent PCB application, where both significant flow disturbance and a high degree of component thermal interaction exist. Numerical analysis was undertaken using a widely-used CFD code specifically designed for the thermal analysis of electronics cooling, thereby permitting the findings of the research to have direct industrial applicability. Predictive accuracy for the fully powered multi-component PCB ranged from ±1 C to ±12 C (up to 30%) depending on component location and flow model. Neither the laminar or turbulent flow model accurately predicted the temperature of all components on the populated board, suggesting the need for a flow model capable of modelling transition. Both the impact of aerodynamic conditions and component thermal interaction on predictive error were assessed. Tentatively, the results show that in applications where the flow is highly disturbed, overall predictive accuracy (±10 C or ±20%) is only sufficient for the early design phase. However, if the flow is steady with low upstream disturbance, predictive accuracy (±5 C or ±10%) may be sufficient for the intermediate design phase. As component junction temperature would therefore need to be experimentally measured to be used as a boundary condition for reliability calculations, CFD analysis could not reliably be used as a virtual prototyping tool in this instance. References [1] Rodgers, P., 2000, An Experimental Assessment of Numerical Predictive Accuracy for Electronic Heat Transfer, Ph.D. Thesis, Department of Mechanical & Aeronautical Engineering, University of Limerick, Limerick, Ireland. [2] Lasance, J. M., 1995, The need for a change in Thermal Design Philosophy, Electronics Cooling, Vol. 1, No. 2, pp [3] Rodgers, P., Eveloy, V., Lohan, J., Fager, C.M., Tiilikka, P., and Rantala, J., 1999, "Experimental Validation of Numerical Heat Transfer Predictions for Singleand Multi- Printed Circuit Boards in Natural Convection Environments," Proceedings of the Fifteenth IEEE Semiconductor Thermal Measurement and Management Symposium (SEMI-THERM XV), pp [4] Rodgers, P., Eveloy, V., Lohan, J., Fager, C.M., and Rantala, J., 1999, "Experimental Validation of Numerical Heat Transfer Predictions for Single- and Multi- Printed Circuit Boards in a Forced Convection Environment: Part I Experimental and Numerical Modelling," 33rd ASME National Heat Transfer Conference, NHTD [5] Rodgers, P., Lohan, J., Eveloy, V., Fager, C.M., and Rantala, J., 1999, "Validating Numerical Predictions of Thermal Interaction on Electronic Printed Circuit Boards in Forced Convection Air Flows by Experimental Analysis," Advances in Electronic Packaging; Proceedings of The PACIFIC RIM/ASME International Intersociety Electronic & Photonic Packaging Conference (InterPACK 99), EEP-Vol. 26-1, pp [6] Eveloy, V., Lohan, J., and Rodgers, P., 2000, A Benchmark Study of Computational Fluid Dynamics Predictive Accuracy for -Printed Circuit Board Heat Transfer, IEEE Transactions on s and Packaging Technology (CPT), Vol. 23, No. 3, pp [7] Eveloy, V., Rodgers, P., and Lohan, J., 2001, Numerical Heat Transfer Predictive Accuracy for an In-line Array of Board-mounted PQFP s in Forced Convection, To be presented at the PACIFIC RIM/ASME International Intersociety Electronic & Photonic Packaging Conference (InterPACK 01), July 8-13, Kauai, Hawaii, USA. [8] Lohan, J., and Davies, M., 1996, Thermal Interaction Between Electronic s, 31st ASME National Heat Transfer Conference, HTD-329, Vol. 7, pp [9] Dantec, Safex Standard Fog Fluid, Dantec Measurement Technology A/S, Tonsbakken 16-18, P.O. Box 121, DK-2740 Skovlunde, Denmark. [10] Flotherm, 2000, Version 3.1 Reference and User Manuals, Flomerics Limited, Bridge Road, Hampton Court, Surrey, KT8 9HH, United Kingdom. [11] Rosten, H., Parry, J. Addison, R. Viswanath, M. Davies, and E. Fitzgerald, 1995, Development, Validation and Application of a Thermal Model of a PQFP, Proceedings of the 45th ECTC, pp [12] Azar, K., and Graebner, J. E., 1996, Experimental Determination of Thermal Conductivity of Printed Wiring Boards, Proceedings of the Twelfth IEEE SEMI-THERM Symposium (SEMI-THERM XII), pp [13] Mack, B., and Venus, T., 2000, Thermal Challenges in the Telecom and Networks Industry, Electronics Cooling, Vol 6, No. 2, pp [14] Rodgers, P., Eveloy, V., Lohan, J., Fager, C.M., and Rantala, J., 1999, "Experimental Validation of Numerical Heat Transfer Predictions for Single- and Multi- Printed Circuit Boards in a Forced Convection Environment: Part II Results and Discussion," 33rd ASME National Heat Transfer Conference, NHTD [15] Anderson, A., 1997, A Comparison of Computational and Experimental Results for Flow and Heat Transfer from an Array of Heated Blocks, Transactions of the ASME, Journal of Electronic Packaging, Vol. 119, March, pp Acknowledgments The authors gratefully acknowledge Flomerics, UK, for the use of Flotherm and their technical support. The test vehicles were built and thermally characterised at the Stokes Research Institute, University of Limerick, Ireland, while the flow visualisations were performed at the Galway-Mayo Institute of Technology, Ireland. The computational facilities of Electronics Thermal Management, Ireland were used to undertake the numerical analyses. 7