Intermediate Layer Collaborative Heat Dissipation Structure and Design Method for Chiplet Heterogeneous Integration Package

By functional partitioning on the interposer layer and employing locally enhanced cooling zones, electrical interconnect avoidance zones, and thermal diffusion compensation structures, the conflict between cooling channels and high-density electrical interconnect structures in the Chiplet heterogeneous integrated package is resolved, achieving adaptive and efficient heat dissipation and improving the overall performance and reliability of the package.

CN123124803APending Publication Date: 2026-11-17WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611116054.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing Chiplet heterogeneous integrated packaging, the cooling channels and high-density electrical interconnect structures conflict with each other. Uniform microchannel cooling resources have low utilization efficiency, and areas with both high heat flux density and high interconnect density cannot simultaneously achieve heat dissipation and interconnect reliability.

Method used

By functionally partitioning the intermediate layer into high heat flux density-low interconnect density region, low heat flux density-high interconnect density region, and thermal-electrical conflict region, adaptive and efficient heat dissipation is achieved by adopting local enhanced cooling region, electrical interconnection avoidance region and heat diffusion compensation structure.

Benefits of technology

While ensuring the integrity of electrical interconnect signals, it achieves adaptive and efficient heat dissipation for non-uniform heat flux density, improving the overall performance and reliability of the Chiplet heterogeneous integrated package.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123124803A_ABST
    Figure CN123124803A_ABST
Patent Text Reader

Abstract

This invention provides a collaborative heat dissipation structure and design method for interposer layer in chiplet heterogeneous integrated packaging. The interposer layer includes a redistribution layer, a microfluidic network, and a thermal diffusion compensation structure. The planar region of the interposer layer is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermal-electrical conflict region, based on the heat flux density of the chip and the interconnect density of the redistribution layer. The microfluidic network flows through the high heat flux density-low interconnect density region to form a locally enhanced cooling region, where the flow density is greater than that of the rest of the microfluidic network. A thermal diffusion compensation structure is provided between the thermal-electrical conflict region and the adjacent high heat flux density-low interconnect density region to conduct heat from the thermal-electrical conflict region to the adjacent locally enhanced cooling region. The low heat flux density-high interconnect density region includes an electrical interconnect avoidance region where no microfluidic network is arranged.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of advanced electronic packaging, chiplet heterogeneous integration, 2.5D / 3D packaging interposer, package-level liquid cooling, multiphysics collaborative design, machine learning-assisted structural optimization, and package reliability monitoring. Specifically, it relates to an interposer collaborative heat dissipation structure and design method for chiplet heterogeneous integrated packaging. Background Technology

[0002] With the development of high-performance computing, artificial intelligence acceleration, automotive computing, edge servers, and high-bandwidth storage systems, single large-chip solutions are limited by manufacturing yield, area cost, and heterogeneous process compatibility. Chiplet heterogeneous integration combines CPUs, GPUs, AI accelerators, high-bandwidth memory (HBM), I / O chips, power management chips, and optoelectronic interconnect chips within the same package, enabling chips with different process nodes, functional units, and thermal characteristics to work collaboratively within a system-in-package. The interposer, as the high-density interconnect carrier between chips, typically simultaneously supports redistribution layers (RDL), through-silicon vias (TSVs), microbump arrays, power supply networks, inter-chip bridging interconnects, and package substrate connection structures, and its space resources are very limited.

[0003] In the aforementioned heterogeneous integrated packaging, the heat flux density, interconnect density, and thermal sensitivity of each chip are highly uneven. High-power computing chips and AI accelerator chips often form local hotspots, while memory chips such as HBM are sensitive to temperature gradients and thermal crosstalk. I / O chips and high-speed interface areas have high interconnect density and signal integrity requirements. Traditional top heat sinks, uniform heat diffusion layers, or uniform microchannel cooling structures usually only consider heat dissipation and cannot simultaneously address the space constraints of high-density redistribution layer routing, TSV placement, microbump arrays, power supply networks, and sealing structures.

[0004] Existing interposer microchannel cooling solutions often emphasize fluid heat transfer capacity or the ability to remove local hot spots. Simply placing microchannels below hot spots may cause them to traverse high-density interconnect regions, disrupting the continuity of redistribution layers, TSVs, microbump arrays, and power supply networks, increasing the risks of open / short circuits, impedance abrupt changes, stress concentration, leakage, and package warpage. If local channels are completely eliminated to avoid interconnects, that area loses its active cooling capability. Furthermore, uniformly spaced microchannels below the entire interposer can generate unnecessary ineffective pressure drops in low heat flux density regions, while the heat transfer capacity in high heat flux density regions is insufficient to eliminate local hot spots, making it impossible to allocate cooling resources on demand within a limited pressure drop budget. Moreover, existing technologies lack effective methods for handling regions with both high heat flux density and high interconnect density. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of conflict between cooling channels and high-density electrical interconnect structures in existing chiplet heterogeneous integrated packaging, low utilization efficiency of uniform microchannel cooling resources, and inability to balance heat dissipation and interconnect reliability in areas with both high heat flux density and interconnect density. This invention proposes a collaborative heat dissipation structure and design method for interposer layer in chiplet heterogeneous integrated packaging. By performing partitioned collaborative design on the interposer layer based on heat flux density and interconnect density, adaptive and efficient heat dissipation for non-uniform heat flux density is achieved while ensuring the integrity of electrical interconnect signals.

[0006] To achieve the above objectives, the specific technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a cooperative heat dissipation structure for intermediary layer in Chiplet heterogeneous integrated packaging, wherein the intermediary layer is provided with a redistribution layer, a microfluidic network and a thermal diffusion compensation structure.

[0008] The planar region of the interposer layer is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermo-electrical conflict region, based at least on the heat flux density of the core and the interconnect density of the redistribution layer. The high heat flux density-low interconnect density region is a region where the heat flux density is higher than a heat flux threshold and the interconnect density is lower than an interconnect density threshold. The low heat flux density-high interconnect density region is a region where the heat flux density is lower than a heat flux threshold and the interconnect density is higher than an interconnect density threshold. The thermo-electrical conflict region is a region where the heat flux density is higher than the heat flux threshold and the interconnect density is higher than the interconnect density threshold.

[0009] The microchannel network flows through the high heat flux density-low interconnect density region to form a locally enhanced cooling region, where the channel density of the locally enhanced cooling region is greater than the channel density of the rest of the microchannel network.

[0010] The thermal diffusion compensation structure is provided between the thermal-electric conflict zone and the adjacent high heat flux density-low interconnect density zone to conduct the heat from the thermal-electric conflict zone to the adjacent local enhanced cooling zone.

[0011] The low heat flux density-high interconnect density region is provided with an electrical interconnection avoidance zone where the microchannel network is not arranged.

[0012] Furthermore, the localized enhanced cooling zone includes at least one of the following: dense microchannels, parallel branched microchannels, microneedle rib array channels, turbulence rib channels, and locally high thermal conductivity filled channels.

[0013] Furthermore, the electrical interconnection avoidance area is provided with at least one electrical interconnection structure selected from high-speed signal lines, redistribution layer interconnects, power supply networks, microbump arrays, through-silicon vias, and inter-chip bridging interconnects; the electrical interconnection avoidance area and the adjacent microchannel network are provided with an insulation support spacing that meets a preset threshold, and the boundary of the electrical interconnection avoidance area is provided with a sealing ring, an insulating dielectric layer, or a mechanical support rib.

[0014] Furthermore, the thermal diffusion compensation structure includes at least one of a copper layer, a diamond composite material layer, a graphite layer, a graphene layer, a silicon carbide layer, an aluminum nitride layer, and a copper-molybdenum composite layer, with a thickness of 1 μm to 500 μm.

[0015] Furthermore, the interposer layer collaborative heat dissipation structure for heterogeneous chipplet packaging also includes a sensor array, which is disposed at at least one of the following locations: the projection area of ​​the high-power chipplet on the interposer layer, the location of the microfluidic network, the boundary of the electrical interconnection avoidance area, and the edge of the thermal diffusion compensation structure.

[0016] The sensor array includes at least one type of temperature sensing node, pressure sensing node, flow sensing node, leakage monitoring node, and interconnect impedance monitoring node.

[0017] Furthermore, the cooling medium is at least one of deionized water, dielectric coolant, ethylene glycol aqueous solution, fluorinated liquid, phase change cooling medium, or gas-liquid two-phase cooling medium.

[0018] In a second aspect, the present invention provides a Chiplet heterogeneous integrated package, comprising the interposer-assisted heat dissipation structure described in the first aspect, a core layer disposed above the interposer-assisted heat dissipation structure, and a package substrate layer disposed below the interposer-assisted heat dissipation structure.

[0019] Thirdly, the present invention provides a design method for a cooperative heat dissipation structure of an interposer layer for Chiplet heterogeneous integrated packaging, comprising:

[0020] Based on the obtained heat flux density of the core and the interconnect density of the redistribution layer, as well as the heat flux threshold and interconnect density threshold, the planar region of the interposer layer is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermal-electrical conflict region. The low heat flux density-high interconnect density region is provided with an electrical interconnection avoidance region.

[0021] Generate candidate intermediary layer structures, wherein the candidate intermediary layer structures include at least the design parameters of the partition boundary and the microchannel network;

[0022] The performance parameters of the candidate intermediate layer structure are predicted using the trained machine learning proxy model, and the design parameters of the candidate intermediate layer structure are adjusted under preset constraints to perform iterative optimization with the goal of minimizing the performance parameters, so as to obtain the target intermediate layer structure.

[0023] Based on the target intermediary layer structure, a thermal diffusion compensation structure is provided between the thermal-electric conflict region and the adjacent high heat flux density-low interconnect density region, and a locally enhanced cooling region is provided within the high heat flux density-low interconnect density region, wherein the flow channel density of the locally enhanced cooling region is greater than the flow channel density of the rest of the microchannel network.

[0024] Furthermore, the performance parameters include at least peak temperature, flow channel pressure drop, and interconnection congestion.

[0025] The preset constraints include at least the following: the peak temperature of the candidate interposer structure does not exceed the maximum junction temperature, the channel pressure drop does not exceed the upper limit of the allowable pressure drop, no microchannel network is set in the electrical interconnection avoidance area, and the channel size in the design parameters of the microchannel network is not less than the minimum manufacturing feature size.

[0026] Furthermore, the machine learning agent model is one of graph neural networks, multi-task neural networks, Gaussian process regression models, and physically constrained neural networks.

[0027] Compared with the prior art, the advantages of the present invention are:

[0028] This invention provides a collaborative heat dissipation structure and design method for interposer layers in chiplet heterogeneous integrated packaging. Based on the heat flux density of the chip and the interconnect density of the redistribution layer, the planar region of the interposer layer is functionally partitioned. A microchannel network flows through the high heat flux density-low interconnect density region, and a cooling medium is used to cool this region, thereby concentrating the cooling medium on the region that most needs heat dissipation and is most suitable for channeling. For the low heat flux-high interconnect density region, no microchannel network is arranged, which fundamentally avoids the damage to the continuity of the redistribution layer, TSV, microbump array and power supply network caused by channel etching, sealing and thermal cycling stress. This reduces the risk of open circuits, impedance changes and leakage, and ensures the electrical signal integrity and long-term reliability of the high-density interconnect region. For the thermal-electric conflict zone, where both heat flux density and interconnect density are high, a microchannel network is not deployed in this area to avoid damaging the interconnect structure. Instead, a heat diffusion compensation structure is set between the thermal-electric conflict zone and the adjacent high heat flux density-low interconnect density zone. This structure actively conducts heat laterally to a locally enhanced cooling zone with strong convective heat transfer capabilities, where it is ultimately carried away by the cooling medium. This achieves a three-in-one synergistic heat dissipation path that integrates interconnect protection in the avoidance zone, heat conduction by the compensation structure, and strong cooling in the adjacent zone. This solves the technical problem of lacking effective means to handle thermal-electric conflict zones in existing technologies, and achieves adaptive and efficient heat dissipation for non-uniform heat flux density while ensuring the integrity of electrical interconnect signals. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A schematic diagram of the functional partitioning of the intermediary layer planar region provided in an embodiment of the present invention;

[0031] Figure 2 This is a cross-sectional schematic diagram of the intermediary layer collaborative heat dissipation structure provided in an embodiment of the present invention;

[0032] Figure 3 A flowchart illustrating the design method of a collaborative heat dissipation structure for interposer layer in a Chiplet heterogeneous integrated package according to an embodiment of the present invention;

[0033] Figure 4 A flowchart illustrating a method for manufacturing a Chiplet heterogeneous integrated package provided in an embodiment of the present invention;

[0034] Figure 5The figure shows a comparison of simulation results between the Chiplet heterogeneous integrated package and the uniform microfluidic structure provided in the embodiments of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to clearly and completely understand the technical solution of the present invention, the present invention will be further described in detail below with reference to embodiments. Obviously, the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0036] To address the problems of conflict between interposer microchannel cooling and high-density electrical interconnect structures, low efficiency of uniform cooling resource utilization, and disconnect between machine learning optimization and actual package structure in existing chiplet heterogeneous integrated packaging, this invention provides an interposer collaborative heat dissipation structure and design method for chiplet heterogeneous integrated packaging. Instead of simply listing the interposer, microchannel network, and machine learning model side-by-side, this invention uses functional partitioning of the interposer as its core technical feature. The output of the machine learning proxy model is directly transformed into the physical partition boundaries of the interposer, the location of locally enhanced cooling zones, the microchannel topology, electrical interconnect avoidance boundaries, thermal diffusion compensation structural parameters, and sensor array arrangement. This allows the algorithm features and package structure features to functionally support each other and jointly solve the thermal-electrical-current coupling constraint problem at the package level.

[0037] For ease of understanding, the terminology used in this invention will be explained as follows:

[0038] Chiplets: These are bare chips with independent functions integrated into the same package using advanced packaging technology, such as CPU chipslets, GPU chipslets, AI accelerator chipslets, high-bandwidth memory (HBM) chipslets, I / O chipslets, control chipslets, power management chipslets, or optical interconnect chipslets.

[0039] Heat flux density: refers to the heat power per unit area projected onto the plane of the interlayer, with units of W / cm². 2 The heat flux density of the core particle can be determined by the ratio of the core particle power consumption to the projected area of ​​the core particle on the interlayer, or it can be obtained by combining the transient power consumption waveform of the core particle with thermal simulation mapping.

[0040] Interconnect density: refers to the amount of interconnect resources occupied by the redistribution layer per unit area in the intermediate layer plane, specifically the number of interconnects per unit area (unit: interconnects / mm). 2 The interconnection density can be measured by wiring occupancy or wiring congestion; the interconnection density can be extracted from the wiring results of electronic design automation (EDA), wiring congestion diagram or power supply network distribution diagram.

[0041] Channel density: refers to the number of channels arranged in the microchannel network per unit area of ​​the interlayer plane, specifically the total length of channels per unit area (unit: mm / mm). 2 The flow channel density in a localized enhanced cooling zone is greater than that in the rest of the microchannel network. For example, the flow channel spacing in a localized enhanced cooling zone is 30% to 70% of the flow channel spacing in the rest of the network.

[0042] Proximity: refers to two regions that are directly adjacent in the plane direction of the intermediary layer, or the distance between them is not greater than a preset proximity distance; the preset proximity distance is, for example, 500μm to 5000μm, so that the thermal resistance of the lateral heat conduction path formed by the heat diffusion compensation structure within this distance meets the heat dissipation requirements.

[0043] Performance parameters: These are indicators used to evaluate the thermal, fluid, and electrical performance of candidate interposer structures; peak temperature refers to the highest temperature of the core chip within the package; channel pressure drop refers to the total pressure loss of the cooling medium flowing through the microchannel network; interconnect congestion refers to the degree of occupancy of wiring resources in the redistribution layer, which can be characterized by the number of wiring violations or the wiring occupancy rate.

[0044] Example 1

[0045] like Figure 1 and Figure 2 As shown, this invention proposes a collaborative heat dissipation structure for interposer layers in heterogeneous chiplet integrated packaging. The interposer layer is no longer considered a single functional layer, but rather a collaborative heat dissipation functional layer integrating a redistribution layer (RDL), a microfluidic network, and a thermal diffusion compensation structure, achieving a deep integration of electrical interconnection and thermal management functions. The interposer layer of this embodiment can be used in existing packages, including 2.5D packaging, 3D stacked packaging, fan-out chiplet packaging, glass interposer packaging, or silicon interposer packaging. The package may include a chip layer, a packaging substrate layer, and the interposer layer of this embodiment. The chip layer is disposed above the interposer layer, and the packaging substrate layer is disposed below the interposer layer. The chip layer contains at least two of the following: CPU chips, GPU chips, AI accelerator chips, high-bandwidth memory (HBM) chips, I / O chips, control chips, power management chips, or optical interconnect chips. Each chip is electrically connected to the redistribution layer of the interposer layer via microbumps, and the interposer layer is electrically connected to the packaging substrate layer via through-silicon vias (TSVs). Different chips differ significantly in area, power density, operating duty cycle, thermal sensitivity, and interconnect requirements, making them unsuitable for a completely uniform cooling and wiring strategy.

[0046] The planar region of the interposer layer is functionally partitioned at least according to the heat flux density of the core and the interconnect density of the redistribution layer. Specifically, such as Figure 1As shown, the interposer plane region is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermal-electrical conflict region.

[0047] A high heat flux density-low interconnect density region refers to an area in the chiplet projection area or its vicinity where the heat flux density is significantly higher than the package average, while the density of the RDL, TSV, microbump array, or high-speed signal line is relatively low. Specifically, within this region, the heat flux density value is higher than a preset heat flux threshold, and the interconnect density value of the redistribution layer is lower than a preset interconnect density threshold. This type of region typically corresponds to the area below the projection region of a GPU chiplet, AI accelerator chiplet, or high-power CPU chiplet, where the RDL wiring density is relatively low, making it suitable for deploying a high-density microchannel network. The microchannel network (used to transport the cooling medium) flows through the high heat flux density-low interconnect density region to form a locally enhanced cooling zone within this area. The channel density of the locally enhanced cooling zone is greater than the channel density of the rest of the microchannel network, allowing the cooling medium to preferentially pass near the hotspot under limited pressure drop, achieving near-source heat dissipation.

[0048] The low heat flux density-high interconnect density region refers to an area with relatively low heat load but dense wiring, power supply, or high-speed interconnect structures. Specifically, in this region, the heat flux density value is lower than a preset heat flux threshold, and the interconnect density value of the redistribution layer is higher than a preset interconnect density threshold. This type of region typically corresponds to the area below I / O chips, high-speed SerDes interfaces, power supply network convergence areas, or external interface fan-out areas. Although the heat load in this region may not be the highest, the requirements for electrical interconnect continuity and signal integrity are high. To protect the integrity of electrical signals and without requiring strong cooling, an electrical interconnect avoidance zone is set up in this region, and microchannel networks are not arranged there. Open channel crossings are further restricted or prohibited to avoid damage to the electrical interconnect structures in this region during channel processing, sealing, or thermal cycling. Redistribution layers, through-silicon vias, microbump arrays, and power supply networks are preferentially preserved.

[0049] A thermal-electric conflict zone refers to a region that simultaneously possesses high heat flux density and high interconnect density. Specifically, within this region, the heat flux density exceeds a heat flux threshold, and the interconnect density of the redistribution layer exceeds an interconnect density threshold. Instead of forcibly creating flow channels in this region, priority is given to ensuring the continuity of the electrical interconnect structure and signal integrity. A thermal diffusion compensation structure is then placed between the thermal-electric conflict zone and the adjacent high heat flux density-low interconnect density region, forming a comprehensive thermal diffusion compensation zone. For example... Figure 2 As shown, the heat diffusion compensation structure is used to conduct the heat generated in the thermal-electric conflict zone to the adjacent local enhanced cooling zone through the lateral heat conduction path. With the help of the efficient heat exchange capacity of the local enhanced cooling zone, the heat is carried away, so that the area where the flow channel cannot be directly arranged can still release heat to the adjacent local enhanced cooling zone through the lateral heat conduction path, thereby achieving effective heat dissipation in the area.

[0050] The aforementioned heat flux threshold can be determined based on the statistical characteristics of the heat flux density distribution of each chip within the package. For example, the 70th percentile of the heat flux density distribution can be used as the heat flux threshold, and the 30th percentile as the heat flux threshold. An exemplary range for the heat flux threshold is 30 W / cm². 2 Up to 300W / cm 2 The interconnect density threshold can be determined based on the statistical characteristics of the interconnect density distribution of the redistribution layer. An exemplary range for the interconnect density threshold is 100 interconnects / mm. 2 Up to 5000 pieces / mm 2 The above thresholds are exemplary values, and those skilled in the art can adjust them according to package size, redistribution layer process node, chip power, and cooling medium type; the thresholds can also be adaptively corrected by subsequent machine learning proxy models based on maximum junction temperature, allowable voltage drop limit, and interconnect congestion constraints.

[0051] Compared to existing technologies that use uniform flow channels across the entire area or only flow channels based on hot spots, this invention divides regions based on heat flux density and redistribution layer interconnect density, and adopts differentiated designs for different regions. This allows the limited cooling medium to prioritize cooling high heat flux density regions, while implementing electrical interconnect avoidance in high interconnect density regions. This ensures the integrity of high-density electrical interconnect signals and achieves effective thermal management of non-uniform extreme heat flux densities, significantly improving the overall performance and reliability of the Chiplet heterogeneous integrated package.

[0052] In this embodiment, the interposer can be a silicon interposer, a glass interposer, an organic interposer, a ceramic interposer, or a composite interposer. When the interposer is a silicon interposer, the microchannel network is formed by deep reactive ion etching or wet etching; when the interposer is a glass interposer, the microchannel network is formed by laser-induced etching, wet etching, or precision grinding; when the interposer is an organic interposer, the microchannel network is formed by laser grooving, imprinting, lamination, or additive manufacturing.

[0053] The microfluidic network allows for the flow of cooling media, which can be at least one of deionized water, dielectric coolant, ethylene glycol aqueous solution, fluorinated liquid, phase change cooling media, or gas-liquid two-phase cooling media. When using a cooling media with a high risk of conductivity, an anti-corrosion layer, an insulating layer, a hydrophilicity / hydrophobicity control layer, or a particle filter structure is installed on the inner wall of the flow channel, and leakage monitoring nodes are installed at the boundary of the electrical interconnection avoidance zone.

[0054] Further, see Figure 1The microchannel network can include a main channel, branch channels, and an outlet channel (not shown in the figure). Both the branch channels and the outlet channel flow through the main channel, with the branch channels upstream and the outlet channels downstream. The main channel has inlets and outlets for the cooling medium in the cooling zone. The cooling medium flows into the main channel through the inlet, then through the branch channels to the locally enhanced cooling zone. After heat exchange with the high heat flux density-low interconnect density zone, the cooling medium is discharged back into the main channel through the outlet channel and then discharged from the outlet, thus forming a complete flow path for the cooling medium, effectively dissipating heat from the high heat flux density-low interconnect density zone.

[0055] In this embodiment, the boundaries of the functional regions, the location of the locally enhanced cooling zones, the microchannel topology, and the boundaries of the electrical interconnect avoidance zones are determined by a machine learning proxy model under the combined influence of thermal performance, electrical interconnect constraints, channel pressure drop, and manufacturing size constraints, thus forming a corresponding manufacturable intermediate layer physical structure. Specifically, the partitioning of the intermediate layer planar region can be determined as follows: First, the intermediate layer plane is discretized according to a regular grid or irregular cells consistent with the core particle projection boundary. Each cell records the heat flux density q, interconnect density d, micro-bump or TSV occupancy b, fluid path accessibility r from the inlet of the cooling medium, safe distance s from the sealing boundary, and minimum manufacturing feature size m. The heat flux density q can be obtained from the core particle power consumption diagram, thermal simulation results, or operating power consumption mapping; the interconnect density d can be obtained from EDA wiring results, wiring congestion diagram, or power supply network distribution diagram; the fluid path accessibility r can be determined by the location of the cooling medium inlet, the location of the cooling medium outlet, the minimum bending radius of the channel, and the location of the sealing ring (the sealing ring is set around the microchannel network, and the sealing ring can be a metal seal, a glass seal, a polymer seal, or a hybrid bonded seal).

[0056] In a preferred embodiment, when the heat flux density q is greater than 50 W / cm² 2 And the interconnect density d is less than 1000 wires / mm 2 At that time, the unit was initially divided into a high heat flux density - low interconnect density region; when the heat flux density q is less than 50 W / cm² 2 And the interconnect density d is greater than 1000 strands / mm 2 At that time, the unit was initially divided into a low heat flux density-high interconnect density region; when the heat flux density q is greater than 50 W / cm² 2 And the interconnect density d is greater than 1000 strands / mm 2In some cases, the unit is designated as a thermal-electrical conflict zone; in others, it can be incorporated into an adjacent functional area based on fluid path accessibility and thermal coupling with adjacent areas. The above thresholds are not the only limitations, and those skilled in the art can adjust them according to package size, RDL process node, chiplet power, and cooling medium. For example, in high-power AI packages, the heat flux threshold can be set to 80W / cm² to 150W / cm²; in ultra-high-density interconnect packages, the interconnect density threshold can be set to 2000 interconnects / mm² or higher.

[0057] After initial threshold partitioning, the partition boundaries can be further corrected using a trained machine learning proxy model (the model will be described in detail in Example 2). Using the intermediate layer functional partitioning as a core technical feature, the output of the machine learning proxy model is directly converted into physical partition boundaries, local enhanced cooling zone locations, microchannel topology, electrical interconnection avoidance zone boundaries, and thermal diffusion compensation structure parameters of the intermediate layer. This allows the algorithm features and packaging structure features to functionally support each other and jointly solve the thermal-electrical-fluid coupling constraint problem at the packaging level. In the final correction result, if a certain initial cooling region has a high heat flux density but is too far from the inlet of the cooling medium and has a large contribution from the channel pressure drop, part of it can be converted into a region covered by the thermal diffusion compensation structure. Heat transfer can be achieved by thickening the thermal diffusion compensation structure or increasing the heat transfer intensity of adjacent channels. If a certain region has a low actual interconnection congestion and is close to the main channel, a closed heat-conducting groove without cooling medium flow can be set up to enhance heat dissipation without damaging the electrical interconnection. Through the above processing, the functional zoning boundaries are not drawn manually based on experience, but rather are structural boundaries formed under the combined effects of thermal performance, electrical performance, fluid resistance, and manufacturability constraints.

[0058] In one specific embodiment, the locally enhanced cooling zone can be located below the high heat flux density-low interconnect density zone, meaning that some branch channels are arranged within the vertical projection range of this region, and the area covered by these branch channels constitutes the locally enhanced cooling zone. Furthermore, the locally enhanced cooling zone can employ at least one of the following structural forms: dense microchannels, parallel branch microchannels, micro-needle fin array channels, turbulence fin channels, or locally high thermal conductivity filled channels, to further improve the local heat transfer coefficient.

[0059] Specifically, densified microchannels: by reducing the hydraulic diameter of these branch channels, the heat exchange area is increased, thereby enhancing the convective heat transfer effect.

[0060] Parallel branched microchannels: This part of the branched channel is divided into multiple parallel short branch channels, which significantly reduces the channel pressure drop while ensuring the total flow rate.

[0061] Micro-needle rib array flow channel: Micron-sized columnar structures are arranged in the branch flow channel of this part, so that the cooling medium generates a tail vortex when flowing around it, which enhances the turbulence disturbance and improves the heat transfer efficiency.

[0062] Turbulence ribbed flow channel: Horizontal or oblique ribs are set on the inner wall of the branch flow channel in this part to force the change of flow direction and induce secondary flow, thereby enhancing the mixing and heat exchange of the cooling medium near the wall.

[0063] Localized high thermal conductivity filled flow channels: The bottom wall or side wall interior of this branch flow channel is made of a high thermal conductivity material (such as copper, diamond, graphene or aluminum nitride, etc.). By utilizing the thermal conductivity of the high thermal conductivity material, the heat exchange capacity between the cooling medium and the hot spot area is improved.

[0064] For example, below high-power CPU / GPU or AI accelerator chips, the locally enhanced cooling zone preferably employs a combination of dense microchannels and micro-needle rib array channels. For voltage drop-sensitive packages, parallel branched microchannels can be used to ensure that locally enhanced cooling does not significantly increase the total voltage drop. For areas with high turbulence enhancement requirements, turbulence-enhancing rib channels can be used. For areas where it is not suitable to create channels but where enhanced thermal conductivity is required, locally high thermal conductivity filling channels can be used. For example, the main channel width can be 100μm to 1000μm, the branch channel and outlet channel width can be 20μm to 300μm, the channel depth can be 10μm to 1000μm, the channel spacing can be 10μm to 2000μm, the equivalent diameter of the micro-needle ribs can be 5μm to 500μm, and the height of the micro-needle ribs can be the same as or slightly lower than the channel depth. Furthermore, all the dimensions of the aforementioned channel structures are not less than the minimum feature size required for the corresponding interposer processing technology.

[0065] In the structural design of the aforementioned locally enhanced cooling zone, densified microchannels, parallel branched microchannels, micro-needle rib array channels, turbulence rib channels, and locally high thermal conductivity filling channels can be selected individually or arbitrarily combined according to actual heat dissipation requirements, processing conditions, and channel pressure drop budget. This achieves synergistic optimization of heat dissipation performance and pressure drop constraints, thereby achieving optimal heat dissipation performance under different operating conditions.

[0066] In one specific embodiment, the electrical interconnect clearance area includes at least one electrical interconnect structure selected from high-speed signal lines, redistribution layer interconnects, power supply networks, microbump arrays, through-silicon vias (TSVs), and inter-chip bridging interconnects. The core function of the electrical interconnect clearance area is to protect the electrical interconnect structures. Therefore, microchannel networks are not arranged within the electrical interconnect clearance area to prevent channel etching, sealing, pressure cycling, or cooling medium leakage from affecting the electrical interconnect structures. An insulating support spacing is provided between the electrical interconnect clearance area and the adjacent microchannel network. This insulating support spacing meets a preset threshold to ensure that the clearance area boundary does not experience dielectric breakdown or mechanical failure when the channel is under pressure. For example, the preset threshold for the insulating support spacing is 20 μm to 200 μm, which can be determined based on the cooling medium pressure within the channel, the strength of the interlayer material, and the sealing process.

[0067] The boundary of the electrical interconnect avoidance zone is equipped with a sealing ring, an insulating dielectric layer, or mechanical support ribs. The sealing ring can be a metal seal, a glass seal, a polymer seal, or a hybrid bond seal. Different protection strategies are adopted for different types of electrical interconnect structures: for dense high-speed signal lines, a continuous reference ground plane and shielded traces can be set in the electrical interconnect avoidance zone to reduce impedance abrupt changes; for dense power supply network areas, the voltage drop can be reduced by widening the power / ground metal layer; for dense micro-bump array areas, thermal cycling warpage and mechanical stress can be reduced by locally reinforcing support ribs.

[0068] By setting an insulating support gap between the microfluidic network and the electrical interconnect avoidance area, the microfluidic network is prevented from covering the electrical interconnect avoidance area under pressure deformation or thermal expansion, thus avoiding impact on the electrical interconnect structure within the avoidance area. Simultaneously, adding a sealing ring, insulating dielectric layer, or mechanical support ribs at the boundary of the electrical interconnect avoidance area further blocks potential leakage paths of the cooling medium along the interface. Through the synergistic effect of the aforementioned insulating support gap and boundary reinforcement structure, safe isolation between the microfluidic network and the electrical interconnect avoidance area is achieved, ensuring the signal integrity and physical reliability of the electrical interconnect structure throughout the entire package lifecycle.

[0069] In one specific embodiment, a thermal diffusion compensation structure is provided between the thermal-electrical conflict region and the adjacent high heat flux density-low interconnect density region. Furthermore, the thermal diffusion compensation structure can also be disposed between the redistribution layer and the microfluidic network, between the microfluidic network and the packaging substrate layer, or in a locally embedded region of the interposer. The thermal diffusion compensation structure includes at least one of a copper layer, a diamond composite layer, a graphite layer, a graphene layer, a silicon carbide layer, an aluminum nitride layer, and a copper-molybdenum composite layer. These ultra-high thermal conductivity materials enable rapid lateral heat conduction to the adjacent locally enhanced cooling region, effectively shortening the heat transfer path from the heat source to the cooling medium. For high thermal conductivity materials with a significant difference in thermal expansion coefficient compared to the interposer substrate material, buffer media layers can be provided on their upper and lower surfaces to reduce thermal stress concentration. The thickness of the thermal diffusion compensation structure can range from 1 μm to 500 μm, preferably from 10 μm to 100 μm.

[0070] For regions with low heat flux density and high interconnect density, due to the extremely high interconnect density, signal integrity must be absolutely protected even if the heat generation is not high. Therefore, electrical interconnect avoidance zones are set up, and microfluidic network cooling methods are not used. Furthermore, for heat-electric conflict zones, which have both extremely high heat generation and extremely high interconnect density, electrical interconnect avoidance zones are also set up to protect the interconnects. To solve the heat dissipation problem, a heat diffusion compensation structure must be used to laterally conduct heat to adjacent locally enhanced cooling zones. Therefore, the heat diffusion compensation structure is not a single heat spreader; its placement and size are determined collaboratively with the electrical interconnect avoidance zones and locally enhanced cooling zones. The heat diffusion compensation structure extends along the plane of the interposer from the electrical interconnect avoidance zone to the adjacent locally enhanced cooling zone. Furthermore, when the area or span of the electrical interconnect avoidance zone is large, the thickness and equivalent thermal conductivity of the heat diffusion compensation structure are correspondingly increased to maintain the heat dissipation capacity of the lateral heat conduction path. When the heat transfer intensity of the adjacent locally enhanced cooling zone is high, the heat diffusion compensation structure can be appropriately thinned to reduce manufacturing difficulty and warpage risk.

[0071] In one specific embodiment, the interposer-based collaborative heat dissipation structure further includes a sensor array, which is disposed at at least one of the following locations: the projection area of ​​the high-power chipplet on the interposer, the location of the microfluidic network, the boundary of the electrical interconnection avoidance area, and the edge of the thermal diffusion compensation structure. The sensor array includes at least one type of temperature sensing node, pressure sensing node, flow sensing node, leakage monitoring node, and interconnect impedance monitoring node. The sensor array is used to collect operational data such as temperature, pressure, flow rate, cooling medium leakage, and interconnect impedance.

[0072] During packaging operation, the sensor array collects temperature, pressure, flow rate, humidity, leakage, and interconnect impedance data according to a preset sampling period. When the temperature of a high-power core region rises and the flow rate of the corresponding branch decreases, it can be determined that there is blockage, bubbles, or local valve malfunction in that branch channel. When the temperature rises but the pressure and flow rate do not change significantly, it can be determined that the core power consumption load has increased or the thermal diffusion compensation structure has locally degraded. When the interconnect impedance near the boundary of the electrical interconnect avoidance area increases, the data rate of the corresponding high-speed channel can be reduced, redundant channels can be switched, or packaging maintenance prompts can be triggered. The above operating data can form a real-world operating condition dataset, which is fed back to the machine learning proxy model to correct power density estimation, flow resistance model, equivalent thermal conductivity of the thermal diffusion layer, and sensor node placement strategy.

[0073] Example 2

[0074] Based on the same inventive concept, this embodiment proposes a design method for a cooperative heat dissipation structure of the interposer layer for heterogeneous chiplet integrated packaging, such as... Figure 3 As shown, the specific steps may include:

[0075] S1. Based on the obtained heat flux density of the core and the interconnect density of the redistribution layer, as well as the heat flux threshold and interconnect density threshold, the planar region of the interposer layer is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermal-electrical conflict region. The low heat flux density-high interconnect density region is provided with an electrical interconnect avoidance region. Specifically, the division of the above three regions is based on the comparison results of the heat flux density of the core and the interconnect density of the redistribution layer with the heat flux threshold and interconnect density threshold. For details, please refer to Embodiment 1, which will not be repeated here.

[0076] S2. Generate candidate intermediate layer structures. Each candidate intermediate layer structure includes at least the design parameters for the partition boundaries and the microfluidic network. The partition boundary parameters include the boundary location, shape, and size between high heat flux density-low interconnect density regions, low heat flux density-high interconnect density regions, and thermal-electrical conflict regions. The microfluidic network design parameters include the microfluidic layout topology (e.g., serpentine, branching, grid-like), channel width, channel depth, channel spacing, channel density distribution, inlet / outlet location, and number of channels. The initial generation of candidate intermediate layer structures can employ a regularized layout (e.g., based on a preset template) or a randomized layout (e.g., random sampling within the design space). This embodiment does not specifically limit this approach.

[0077] S3. Using the trained machine learning surrogate model, predict the performance parameters of the candidate intermediate layer structure, and iteratively optimize the design parameters of the candidate intermediate layer structure under preset constraints, with the goal of minimizing the performance parameters, to obtain the target intermediate layer structure. Specifically, this may include the following steps:

[0078] S31. Input the design parameters of the current candidate intermediate layer structure into the trained machine learning agent model, and the machine learning agent model outputs the predicted performance parameters of the candidate intermediate layer structure.

[0079] Machine learning surrogate models are trained on historical design data and are used to approximate high-fidelity multiphysics simulations (such as finite element analysis (FEA) and computational fluid dynamics (CFD)). The input features of the machine learning surrogate model include at least: core type, core heat flux density distribution, interconnect density distribution, partition boundary parameters, microchannel network design parameters, and electrical interconnect clearance zone parameters. The output performance parameters of the machine learning surrogate model include at least: peak temperature, channel pressure drop, and interconnect congestion.

[0080] S32. Determine whether the current candidate intermediate layer structure meets the preset constraints.

[0081] The preset constraints include at least the following: the peak temperature of the candidate interposer structure does not exceed the maximum junction temperature, which is determined based on the chip's material properties, reliability requirements, and process node; for example, it can be set to 85℃, 105℃, or 125℃. The channel voltage drop does not exceed the upper limit of the allowable voltage drop, which is determined based on the pumping capacity and energy consumption budget of the cooling system. No microchannel networks are set within the electrical interconnect avoidance area. The channel size in the microchannel network design parameters is not smaller than the minimum manufacturing feature size, which is determined based on the manufacturing process used (e.g., deep reactive ion etching (DRIE), photolithography, etc.). The preset constraints may also include: maintaining an insulating support spacing between the sealing ring and the redistribution layer that meets a preset threshold; package warpage not exceeding the warpage threshold; and maximum thermal stress not exceeding the reliability threshold.

[0082] When a candidate intermediate layer structure satisfies all the above preset constraints, the candidate intermediate layer structure is determined to be the target intermediate layer structure, and the process proceeds to step S4; otherwise, the process proceeds to sub-step S33.

[0083] S33. With the goal of minimizing performance parameters, use an optimization algorithm to adjust the design parameters of the candidate intermediate layer structure, generate a new candidate intermediate layer structure, and return to sub-step S31 for the next round of iterative optimization.

[0084] Performance parameters include at least peak temperature, channel pressure drop, and interconnect congestion. During optimization, the loss function of the machine learning surrogate model includes temperature prediction loss, pressure drop prediction loss, interconnect congestion prediction loss, warpage prediction loss, and physical constraint penalties. Physical constraint penalties include: penalties for channel dimensions smaller than the minimum processing size, penalties for the presence of microchannel networks within the electrical interconnect clearance zone, penalties for total pressure drop exceeding the allowable pressure drop, penalties for peak temperature exceeding the preset junction temperature, and penalties for insufficient safety clearance between the sealing ring and the redistribution layer. These penalties can be implemented using step functions, exponential functions, or sigmoid functions. For example, the channel size penalty can be set as follows:

[0085] (1)

[0086] In the above formula (1), Penalty term indicating flow channel size; Indicates the minimum machining dimension; Indicates the actual flow channel size; This represents the first penalty coefficient. When the actual flow channel size is greater than or equal to the minimum machining size, the penalty value is zero; when the actual flow channel size is less than the minimum machining size, the penalty value increases twice as the deviation increases.

[0087] Similarly, other penalty terms can also take the form of corresponding functions, and the penalty coefficient of each penalty term can be set according to the importance of each constraint and the severity of the consequences of violation.

[0088] The optimization algorithm can employ gradient descent, genetic algorithm (GA), particle swarm optimization (PSO), simulated annealing (SA), Bayesian optimization, or multi-objective optimization algorithms such as NSGA-II. This embodiment does not specifically limit the specific algorithm used. Each iteration generates candidate intermediate layer structures, whose performance parameters are quickly predicted by a machine learning surrogate model. Candidate intermediate layer structures that violate preset constraints are directly eliminated, while those that meet the preset constraints are sorted using Pareto. After multiple iterations, the target intermediate layer structure that achieves a balance between peak temperature, channel pressure drop, and interconnection congestion is output. In this embodiment, the termination condition for iterative optimization can be set as: satisfying all preset constraints, reaching the maximum number of iterations, or the total loss function value converging (the objective function change is less than a preset threshold for several consecutive iterations).

[0089] S4. Based on the target intermediary layer structure, a thermal diffusion compensation structure is set between the thermal-electric conflict zone and the adjacent high heat flux density-low interconnect density zone, and a local enhanced cooling zone is set in the high heat flux density-low interconnect density zone. The flow channel density of the local enhanced cooling zone is greater than the flow channel density of the rest of the microchannel network.

[0090] In this embodiment, the machine learning agent model can be obtained in the following way:

[0091] First, a sample database is established. Samples can come from thermal-fluid-solid multiphysics simulations, electrical interconnect congestion assessments, and a small amount of prototype test data. For example, for a packaging platform including a CPU, GPU, AI chip, two HBM chips, and an I / O chip, 3,000 to 20,000 candidate interposer structures can be randomly generated or generated based on design rules. Each candidate interposer structure includes at least the design parameters of the partition boundary and microfluidic network. The candidate interposer structure may also include the material and thickness of the thermal diffusion compensation structure, the boundary of the electrical interconnect avoidance zone, and candidate sensor node locations. Subsequently, finite element thermal simulation is used to obtain peak temperature and temperature gradient, computational fluid dynamics simulation is used to obtain branch flow rate and channel pressure drop, thermal stress simulation is used to obtain package warpage and maximum thermal stress, and wiring congestion assessment is used to obtain interconnect congestion degree and the number of interconnect avoidance violations.

[0092] The input features of the machine learning proxy model include at least the chip type, size, coordinates, heat flux density distribution, transient power consumption waveform, interconnect density distribution of the redistribution layer, via location, microbump array density, coolant inlet and outlet locations, coolant type, channel width, channel depth, channel spacing, material of the thermal diffusion compensation structure, thickness of the thermal diffusion compensation structure, upper limit of allowable voltage drop, maximum junction temperature, and minimum feature size. The performance parameters output by the machine learning proxy model include at least peak temperature, channel voltage drop, and interconnect congestion, and may also include average temperature, maximum temperature gradient, branch flow rate, maximum thermal stress, package warpage, and sensor node sensitivity score. For ease of implementation by those skilled in the art, the main input and output variables of the machine learning proxy model can be set as shown in the table below.

[0093] Table 1. Examples of main input and output variables for machine learning agent models.

[0094]

[0095] The machine learning proxy model is one of the following: graph neural network, multi-task neural network, Gaussian process regression model, and physically constrained neural network. When using a graph neural network, the core, high heat flux density-low interconnect density region, low heat flux density-high interconnect density region, thermal-electrical conflict region, cooling medium inlet, cooling medium outlet, and sensing node are represented as graph nodes. The thermal coupling relationship between cores, the electrical interconnect relationship of the redistribution layer, and the fluid connectivity relationship of the microchannel network are represented as graph edges. Node features include heat flux density, area, material thermal conductivity, interconnect density, minimum manufacturing feature size, and partition type. Edge features include distance, thermal resistance estimate, flow resistance estimate, electrical interconnect length, and whether it crosses an electrical interconnect avoidance zone. After extracting the thermal-electric-fluid coupling features, the graph neural network predicts peak temperature, channel pressure drop, interconnect congestion, package warpage, and maximum thermal stress through multiple output heads.

[0096] During training, sample data can be divided into a 70% training set, a 15% validation set, and a 15% test set. The loss function includes temperature prediction loss, pressure drop prediction loss, interconnection congestion prediction loss, and a physical constraint penalty term. In an exemplary training result, the mean absolute error of peak temperature on the test set can be controlled within 2°C, the relative error of channel pressure drop can be controlled within 8%, and the accuracy of interconnection congestion classification can reach over 90%. The above indicators are used to illustrate the implementation of the model and do not limit the present invention to using only this range of accuracy.

[0097] Example 3

[0098] Based on the same inventive concept, this embodiment proposes a Chiplet heterogeneous integrated package, including the interposer-coordinated heat dissipation structure in Embodiment 1, a core layer disposed above the interposer-coordinated heat dissipation structure, and a package substrate layer disposed below the interposer-coordinated heat dissipation structure.

[0099] The Chiplet heterogeneous integrated package provided in this embodiment of the invention has a similar implementation principle and technical effect to that of Embodiment 1, and will not be repeated here.

[0100] Example 4

[0101] Based on the same inventive concept, this embodiment proposes a method for manufacturing a Chiplet heterogeneous integrated package, such as... Figure 4 As shown, the specific steps may include:

[0102] T1. Obtain the design parameters of the target intermediary layer structure. These design parameters are obtained through iterative optimization based on the machine learning proxy model in Embodiment 2 above.

[0103] T2. Preparation of the interposer substrate. The interposer substrate can be silicon, glass, organic materials, ceramics, or composite materials. Depending on the interposer substrate material, corresponding processing techniques are used to form microchannels: for silicon interposers, deep reactive ion etching can be used to form microchannels; for glass interposers, laser-induced modification combined with wet etching can be used to form channels; for organic interposers, laser grooving, lamination, imprinting, or additive manufacturing can be used to form channels.

[0104] T3. Form the redistribution layer, through-silicon vias (TSVs), pads, and power supply network, and reserve a channelless protection area according to the location of the electrical interconnect avoidance area. The electrical interconnect avoidance area corresponds to the low heat flux density to high interconnect density area. In this area, the complete redistribution layer routing space is reserved, and the microchannel network is not fabricated.

[0105] T4. Based on the design parameters of the microchannel network and the localized enhanced cooling zone in the target interposer structure, fabricate the microchannel network and the localized enhanced cooling zone. The localized enhanced cooling zone corresponds to the high heat flux density-low interconnect density region, and its channel density is greater than that of the rest of the microchannel network.

[0106] T5. Form a thermal diffusion compensation structure and a sensor array. A thermal diffusion compensation structure is provided between the thermal-electrical conflict region and the adjacent high heat flux density-low interconnect density region. The thermal diffusion compensation structure can be formed by electroplating, sputtering, bonding, lamination or transfer printing; the sensor array can be formed by thin film resistors, diode temperature sensors, MEMS pressure sensors, thermal flow sensors or interconnect monitoring lines.

[0107] T6. A sealing ring and an insulating support structure are formed around the microchannel network. The sealing ring can be a metal seal, a glass seal, a polymer seal, or a hybrid bonded seal.

[0108] T7. The heterogeneous chip is mounted or bonded to the interposer, and the interposer is connected to the packaging substrate through a through-silicon via (TSV) to complete the assembly of the package.

[0109] T8. Perform package performance tests, including: pressure holding test (to check flow channel sealing), leakage rate test (to check cooling medium leakage), thermal resistance test (to check heat dissipation performance), RDL open and short circuit test (to check electrical interconnect integrity), high-speed signal eye diagram test (to check signal integrity), thermal cycling test (to check reliability), warpage test (to check structural stability), and sensor calibration test (to calibrate sensor nodes). Feed the test results back to the machine learning agent model for model iteration optimization or anomaly warning threshold correction.

[0110] Within a feasible process window, the number of cores is 2 to 30, and the area of ​​a single core is 1 mm. 2 Up to 2500mm 2The microchannel width ranges from 10 μm to 1000 μm, the microchannel depth ranges from 10 μm to 1000 μm, the angle between the main channel and the branch channels ranges from 30° to 150°, the thermal diffusion compensation structure thickness ranges from 1 μm to 500 μm, the electrical interconnection avoidance zone width ranges from 20 μm to 5000 μm, the sealing ring width ranges from 50 μm to 3000 μm, and the number of sensing nodes ranges from 4 to 500. The cooling medium can be deionized water, dielectric coolant, ethylene glycol aqueous solution, fluorinated liquid, or a two-phase cooling medium. When using a cooling medium with a high risk of conductivity, an insulating protective layer and leakage monitoring nodes should be preferentially installed on the inner wall of the channel.

[0111] During mass production, key process parameters are recorded for each batch of interposer samples, including interposer material batch, flow channel width, flow channel depth, thermal diffusion layer thickness, redistribution layer (RDL) number, sealing ring width, alignment error, bonding temperature, bonding pressure, holding time, and leakage rate. Each sample undergoes visual inspection, dimensional measurement, flow channel connectivity testing, holding pressure testing, thermal resistance testing, open / short circuit testing, impedance testing, thermal cycling testing, and warpage testing. If a batch of samples exhibits a high pressure drop but no significant temperature improvement, it indicates excessive flow resistance in the localized enhanced cooling zone or uneven flow distribution; the next design round can add parallel branches or adjust the inlet and outlet positions. If a batch of samples shows abnormal electrical interconnection tests, it indicates that the clearance zone width, sealing boundary, or support spacing needs to be increased.

[0112] Exemplarily, the technical effects of the present invention are illustrated using a 2.5D Chiplet package as an example. This package arranges a CPU chip, a GPU chip, an AI accelerator chip, two HBM chips, and an I / O chip on an interposer. The CPU, GPU, and AI chips have high power densities and are initially identified as high heat flux density-low interconnect density regions; the I / O chip and the redistribution layer (RDL) of the high-speed interconnect region at the edge of the HBM have high wiring densities and are identified as low heat flux density-high interconnect density regions or thermal-electrical conflict regions. Traditional uniform microchannel schemes, which set equally spaced channels under the entire interposer, are simple to manufacture but traverse parts of the dense RDL region and generate unnecessary voltage drops in low-heat areas.

[0113] With this invention, a locally enhanced cooling zone is set below the CPU, GPU, and AI chips. The HBM region adopts a temperature homogenization flow channel and a thermal diffusion compensation structure. Electrical interconnection avoidance zones are set in the I / O chips and high-speed interface fan-out areas. The machine learning proxy model eliminates schemes where the flow channel crosses the RDL dense region from the candidate structures, while enhancing the branch flow in the vicinity of the enhanced cooling zone. Simulation results are as follows. Figure 5As shown, under the same inlet temperature, total flow rate, and total power consumption, the peak temperature of the uniform microchannel scheme is approximately 96°C, the total pressure drop is approximately 42 kPa, and the interconnection congestion is approximately 31%; the peak temperature of the partitioned collaborative structure of this invention is approximately 78°C, the total pressure drop is approximately 35 kPa, and the interconnection congestion is approximately 18%. This indicates that this invention does not simply increase the channel density, but rather achieves a synergistic balance between cooling resource allocation, interconnection avoidance, and thermal diffusion compensation.

[0114] In terms of reliability, with the electrical interconnect avoidance zone in place, high-speed signal lines are no longer cut by open channels or sealed boundaries, reducing local impedance abrupt changes in the RDL. The heat diffusion compensation structure allows heat below the avoidance zone to still be conducted to the adjacent enhanced cooling zone, preventing the formation of new hot spots due to the protection of electrical interconnects. The sensor array is arranged at hot spots, inlets, outlets, branch intersections, and avoidance zone boundaries, enabling the identification of problems such as channel blockage, bubbles, insufficient cooling flow, abnormal chiplet load, and heat diffusion layer degradation during operation.

[0115] In the exemplary comparative evaluation, the uniform microchannel structure can be used as the control group, and the partitioned collaborative structure of the present invention can be used as the experimental group. The evaluation indicators include peak temperature, temperature gradient, total pressure drop, interconnection congestion, and thermal stress. Exemplary results are shown in the table below, which is used to illustrate the disclosure method of the test or simulation data in the specification.

[0116] Table 2 Comparison of evaluation indicators between the control group and the inventor

[0117]

[0118] The beneficial effects of this invention are as follows:

[0119] This invention abandons the traditional single-dimensional design approach of uniform microchannels or simply arranging channels according to hot spots. Instead, it uses heat flux density and redistribution layer interconnect density as the basis for parallel partitioning, dividing the interposer plane into high heat flux density-low interconnect density regions, low heat flux density-high interconnect density regions, and thermal-electrical conflict regions. By concentrating the locally enhanced cooling zone in the high heat flux density-low interconnect density region, the cooling medium flows through this region for heat dissipation. An electrical interconnect avoidance zone is set in the low heat flux density-high interconnect density region, thus concentrating the cooling medium on the areas most in need of heat dissipation and most suitable for channel opening, avoiding ineffective pressure drops caused by uniform microchannels in low-heat areas. Furthermore, the electrical interconnect avoidance area does not have a microchannel network, which fundamentally avoids the damage to the continuity of the redistribution layer, TSV, microbump array and power supply network caused by channel etching, sealing and thermal cycling stress, and reduces the risk of open circuits, impedance changes and leakage. At the same time, the heat is conducted to the adjacent local enhanced cooling area through the heat diffusion compensation structure, so that the high-density interconnect area can still release heat through the lateral heat conduction path without being crossed by the channel.

[0120] More importantly, this invention directly maps the design parameter predictions of the machine learning proxy model to manufacturable physical structures such as partition boundaries, flow channel topology, avoidance zone locations, and thermal diffusion compensation structure layouts of the interposer layer. This enhances the interaction between AI-related features and engineering features, improving the feasibility, verifiability, and licensing stability of the solution. Simultaneously, temperature, pressure, flow, and interconnect impedance sensing nodes are deployed in key areas of the interposer layer to collect multiphysics data of the package under real-world operating conditions. This data is then fed back to the machine learning proxy model, enabling retraining and subsequent design corrections based on the measured results. This significantly improves the long-term adaptability and reliability of the interposer layer's collaborative heat dissipation structure under complex dynamic load conditions.

[0121] This invention does not limit the use of a specific machine learning model or microfluidic fabrication method. As long as the functional partitioning of the interposer layer is determined based on the chip power density, redistribution layer wiring density, cooling medium accessibility, and manufacturing size constraints, and the locally enhanced cooling zone, electrical interconnect avoidance zone, thermal diffusion compensation structure, and sensor array are collaboratively formed into a manufacturable physical structure, it falls within the scope of this invention. The above alternative embodiments ensure the adequacy of the disclosure in the specification while avoiding limiting the scope of protection to a single material, a single cooling medium, or a single algorithm architecture.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A collaborative heat dissipation structure for interposer layer in a chiplet heterogeneous integrated package, characterized in that, The interposer layer includes a redistribution layer, a microfluidic network, and a thermal diffusion compensation structure; The planar region of the interposer layer is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermo-electrical conflict region, based at least on the heat flux density of the core and the interconnect density of the redistribution layer. The high heat flux density-low interconnect density region is a region where the heat flux density is higher than a heat flux threshold and the interconnect density is lower than an interconnect density threshold. The low heat flux density-high interconnect density region is a region where the heat flux density is lower than a heat flux threshold and the interconnect density is higher than an interconnect density threshold. The thermo-electrical conflict region is a region where the heat flux density is higher than the heat flux threshold and the interconnect density is higher than the interconnect density threshold. The microchannel network flows through the high heat flux density-low interconnect density region to form a locally enhanced cooling region, wherein the channel density of the locally enhanced cooling region is greater than the channel density of the rest of the microchannel network, and wherein the microchannel network is used to transport the cooling medium. The thermal diffusion compensation structure is provided between the thermal-electric conflict zone and the adjacent high heat flux density-low interconnect density zone to conduct the heat from the thermal-electric conflict zone to the adjacent local enhanced cooling zone. The low heat flux density-high interconnect density region is provided with an electrical interconnection avoidance zone where the microchannel network is not arranged.

2. The interposer-based collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 1, characterized in that, The localized enhanced cooling zone includes at least one of the following: dense microchannels, parallel branched microchannels, microneedle rib array channels, turbulence rib channels, and locally high thermal conductivity filled channels.

3. The interposer layer collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 1, characterized in that, The electrical interconnect avoidance area is provided with at least one electrical interconnect structure selected from high-speed signal lines, rewiring layer interconnects, power supply networks, microbump arrays, through-silicon vias, and inter-chip bridging interconnects; the electrical interconnect avoidance area and the adjacent microchannel network are provided with an insulation support spacing that meets a preset threshold, and the boundary of the electrical interconnect avoidance area is provided with a sealing ring, an insulating dielectric layer, or a mechanical support rib.

4. The interposer layer collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 1, characterized in that, The thermal diffusion compensation structure includes at least one of a copper layer, a diamond composite material layer, a graphite layer, a graphene layer, a silicon carbide layer, an aluminum nitride layer, and a copper-molybdenum composite layer, with a thickness of 1 μm to 500 μm.

5. The interposer layer collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 1, characterized in that, It also includes a sensor array, which is disposed at at least one of the following locations: the projection area of ​​the high-power Chiplet on the interposer layer, the location of the microfluidic network, the boundary of the electrical interconnection avoidance area, and the edge of the thermal diffusion compensation structure; The sensor array includes at least one type of temperature sensing node, pressure sensing node, flow sensing node, leakage monitoring node, and interconnect impedance monitoring node.

6. The interposer layer collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 1, characterized in that, The cooling medium is at least one of deionized water, dielectric coolant, ethylene glycol aqueous solution, fluorinated liquid, phase change cooling medium, or gas-liquid two-phase cooling medium.

7. A Chiplet heterogeneous integrated package, characterized in that, It includes the interposer-assisted heat dissipation structure as described in any one of claims 1 to 6, a core layer disposed above the interposer-assisted heat dissipation structure, and a packaging substrate layer disposed below the interposer-assisted heat dissipation structure.

8. A design method for a collaborative heat dissipation structure of the interposer layer in a chiplet heterogeneous integrated package, characterized in that, include: Based on the obtained heat flux density of the core and the interconnect density of the redistribution layer, as well as the heat flux threshold and interconnect density threshold, the planar region of the interposer layer is divided into a high heat flux density-low interconnect density region, a low heat flux density-high interconnect density region, and a thermal-electrical conflict region. The low heat flux density-high interconnect density region is provided with an electrical interconnection avoidance region. Generate candidate intermediary layer structures, wherein the candidate intermediary layer structures include at least the design parameters of the partition boundary and the microchannel network; The performance parameters of the candidate intermediate layer structure are predicted using the trained machine learning proxy model, and the design parameters of the candidate intermediate layer structure are adjusted under preset constraints to perform iterative optimization with the goal of minimizing the performance parameters, so as to obtain the target intermediate layer structure. Based on the target intermediary layer structure, a thermal diffusion compensation structure is provided between the thermal-electric conflict region and the adjacent high heat flux density-low interconnect density region, and a locally enhanced cooling region is provided within the high heat flux density-low interconnect density region, wherein the flow channel density of the locally enhanced cooling region is greater than the flow channel density of the rest of the microchannel network.

9. The design method of the interposer layer collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 8, characterized in that, The performance parameters include at least peak temperature, flow channel pressure drop, and interconnection congestion. The preset constraints include at least the following: the peak temperature of the candidate interposer structure does not exceed the maximum junction temperature, the channel pressure drop does not exceed the upper limit of the allowable pressure drop, no microchannel network is set in the electrical interconnection avoidance area, and the channel size in the design parameters of the microchannel network is not less than the minimum manufacturing feature size.

10. The design method of the interposer layer collaborative heat dissipation structure for Chiplet heterogeneous integrated packaging according to claim 8, characterized in that, The machine learning agent model is one of the following: graph neural network, multi-task neural network, Gaussian process regression model, and physical constraint neural network.