GPU chip forming equipment and chip packaging method

Abstract

The invention discloses GPU chip forming equipment and a chip packaging method, and the method comprises the steps: injecting bottom filling resin containing an ultraviolet light initiator, a latent thermal initiator, up-conversion luminescence nanoparticles and magnetic induction heating nanoparticles into a gap between a bare chip and a substrate, and forming a functional resin layer; performing ultraviolet irradiation on the peripheral area, so that the photoinitiator is polymerized to release heat and activate the latent thermal initiator to generate a thermal polymerization reaction frontal surface which is propelled from outside to inside; near-infrared light is transmitted through a bare chip to excite up-conversion luminescent nanoparticles to release ultraviolet light in a shading area, and the ultraviolet light and a thermal polymerization reaction frontal surface cooperate to initiate resin polymerization in the shading area; applying an alternating magnetic field to the initially cured resin layer, exciting magnetic induction to heat the nanoparticles to generate induction heat to trigger the residual latent thermal initiator to complete deep curing and release residual stress; and gradient cooling and constant-temperature aging treatment are executed. According to the invention, the uniform deep curing of the filling resin at the bottom of the GPU chip can be realized in an all-black workshop environment.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of semiconductor packaging, and in particular to a GPU chip forming device and a chip packaging method. BACKGROUND

[0002] A fully black workshop refers to an intelligent manufacturing environment that relies entirely on automated equipment and sensor systems for production operations without the need for artificial lighting. In a fully black workshop, all material transportation, processing operations, quality detection, and equipment maintenance are completed through pre-programmed mechanical systems, machine vision, sensor networks, and artificial intelligence algorithms, completely eliminating the dependence on visible light illumination. This production mode not only significantly reduces energy consumption and operating costs, but also eliminates the interference of ambient light on high-precision optical detection and laser processing, while avoiding the maintenance costs and failure risks of the lighting system.

[0003] The GPU chip packaging process in the prior art includes key steps such as die mounting, wire bonding or flip-chip bonding, underfilling, and plastic encapsulation molding, wherein the underfilling process has a decisive impact on packaging reliability. In the underfilling process, photosensitive epoxy resin is filled into the tiny gap between the chip and the substrate through capillary action, and then cured and polymerized by ultraviolet light irradiation. For modern large-size GPU chips (usually with an area of 800-1000 square millimeters) with a three-dimensional stacked structure of HBM high-bandwidth memory, a complex three-dimensional light-blocking geometry is formed. The chip body and the HBM stack form a complete closure in three-dimensional space, not just a two-dimensional plane blockage, making it impossible for external ultraviolet light to directly irradiate the central area at the bottom of the chip from any angle. This physical shielding effect causes a significant and non-eliminable light dose gradient between the central area and the edge area of the chip during resin curing.

[0004] Under the constraints of a fully black workshop, the auxiliary means such as ambient light diffuse reflection and multi-angle supplementary illumination relied on in the traditional process are completely ineffective. Even if the number and intensity of ultraviolet light sources are significantly increased around the shielding area, due to the straight-line propagation characteristics of light and the physical limitations of complex geometric shielding, ultraviolet light still cannot effectively reach the key curing area at the bottom of the chip. In addition, excessive enhancement of peripheral ultraviolet light irradiation will cause rapid depletion of photoinitiators, resin surface carbonization, and rapid rise of thermal stress in the edge area, making the curing non-uniformity problem worse. Therefore, when the GPU packaging process is implemented in a fully black workshop environment, the resin in the central area of the chip will inevitably be undercured, causing stress concentration at the packaging interface, micro-crack initiation, and long-term reliability degradation. This technical bottleneck cannot be solved by traditional lighting optimization methods and requires a completely new technical path. SUMMARY

[0005] The main purpose of the present application is to solve the technical problems of poor packaging reliability and low yield caused by the non-uniform UV curing of the underfill resin at the bottom due to the large-area chip shading in the existing GPU chip packaging process in a full-black workshop environment, resulting in insufficient curing in the central area and the existence of interface stress concentration.

[0006] The first aspect of the present application provides a GPU chip packaging method, which comprises: injecting an underfill resin containing ultraviolet light initiators, latent heat initiators, up-conversion luminescent nanoparticles and magnetic induction heating nanoparticles into the gap between the die and the substrate to form a fully filled functional resin layer; performing ultraviolet light irradiation on the peripheral area of the functional resin layer to make the ultraviolet light initiators polymerize and release heat and activate the latent heat initiators, generate a thermal polymerization reaction front that advances from the outside to the inside of the resin layer, and obtain a peripheral cured resin area; using near-infrared light to transmit through the die to excite the up-conversion luminescent nanoparticles to release ultraviolet light in the shaded area, and cooperatively initiate resin polymerization in the shaded area with the thermal polymerization reaction front to obtain a continuous initial cured resin layer; applying an alternating magnetic field to the continuous initial cured resin layer to excite the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent heat initiators to complete deep curing and simultaneously release internal residual stress to obtain a fully cured resin layer; performing gradient cooling from the substrate side to the die side on the fully cured resin layer and performing isothermal aging to redistribute and stabilize the curing shrinkage stress to obtain a GPU chip packaging structure.

[0007] Preferably, the injecting an underfill resin containing ultraviolet light initiators, latent heat initiators, up-conversion luminescent nanoparticles and magnetic induction heating nanoparticles into the gap between the die and the substrate to form a fully filled functional resin layer comprises: calculating the shading coefficient according to the die area and the HBM stacking height, setting the molar ratio of the ultraviolet light initiators and the latent heat initiators according to the shading coefficient, preparing the epoxy resin matrix to obtain a resin premix liquid; performing hydroxylation surface modification on the up-conversion luminescent nanoparticles to form hydroxyl sites on the particle surface that can condense with epoxy groups, adding the modified particles to the resin premix liquid and dispersing under constant shear conditions to obtain an up-conversion dispersion liquid; adding magnetic induction heating nanoparticles to the up-conversion dispersion liquid, performing shear homogenization and vacuum degassing combined treatment to remove suspended bubbles and make the two types of functional particles uniformly distributed in the system to obtain a functional underfill resin; maintaining the gap between the die and the substrate under negative pressure in a constant temperature environment, and applying a local infrared radiation heat field to the outer edge of the gap to establish a temperature gradient and a viscosity gradient that decreases from the outer edge to the center, and according to the viscosity gradient, the functional underfill resin is injected into the outer edge area and the central area of the gap in two stages to form an initial filling resin layer; maintaining the negative pressure and supplementing the functional underfill resin in a stepwise manner, while applying micro-vibration mechanical waves to the two sides of the gap to make the residual bubbles migrate to the outer edge and be discharged to form a fully filled functional resin layer.

[0008] Preferably, the upconversion luminescent nanoparticles are subjected to hydroxylated surface modification to form hydroxyl sites on the particle surface that can condense with epoxy groups, the modified particles are added to the resin premix and dispersed under constant shear to obtain an upconversion dispersion, including: controlling the coordination environment of rare earth ions on the surface of upconversion luminescent nanoparticles by silane coupling agent treatment, building a directional arrangement of hydroxyl functional layer on the particle surface, and making the hydroxyl density reach the preset standard to meet the near-infrared light to ultraviolet light conversion efficiency requirement under full black environment; determining the concentration gradient distribution of upconversion luminescent nanoparticles in the resin according to the spatial distribution characteristics of the light shielding area of the GPU chip, and forming a concentration distribution matching the light shielding intensity of the modified particles in the resin premix by controlling the spatial non-uniformity of the shear field; synchronously applying an ultrasonic field during the constant shear dispersion process, using cavitation to break up particle agglomeration and activate the hydroxylated surface, forming a pre-crosslinking bond with the epoxy group, and enhancing the particle-resin interface bonding strength; controlling the pre-crosslinking density on the particle surface by adjusting the corresponding relationship between the shear rate and the ultrasonic power, so that the upconversion luminescent nanoparticles maintain stable light conversion performance during subsequent excitation, and an upconversion dispersion is obtained.

[0009] Preferably, the peripheral area of the functionalized resin layer is subjected to ultraviolet irradiation, the ultraviolet light initiates polymerization exotherm and activates latent heat initiators, a thermal polymerization reaction front is generated along the resin layer from the outside to the inside, and a peripheral cured resin area is obtained, including: modulating the ultraviolet light in a circumferential direction according to the profile of the die outer edge and the thickness of the functionalized resin layer, to obtain a circumferentially uniform circumferential ultraviolet light field; pulse timing control is performed on the circumferential ultraviolet light field, the pulse width and interval are set so that the polymerization exotherm peak temperature is in the activation temperature range of the latent heat initiator, and a stable exothermic pulse sequence is obtained; the ultraviolet light intensity is adjusted in a radial gradient according to the stable exothermic pulse sequence, to obtain a balanced energy distribution along the thickness direction of the resin layer, forming a thermal polymerization reaction front advancing from the outside to the inside; the thermal polymerization reaction front is tracked in real time, and when the front reaches a set distance from the light shielding boundary of the die, the irradiation parameters are kept constant until the exotherm is completed, forming a peripheral cured resin area.

[0010] Preferably, the use of near-infrared light transmits excitation of the up-conversion luminescence nanoparticles in the light-shielded area to release ultraviolet light, and cooperates with the thermal polymerization reaction front to initiate resin polymerization in the light-shielded area, to obtain a continuous initial curing resin layer, comprising: performing optical simulation on the silicon-based wiring layer, through hole and metal interconnection structure according to the GPU die layout structure data, obtaining a wavelength transmittance matrix corresponding to the coordinates, and obtaining a near-infrared transmittance spectrum parameter matrix; constructing a multi-wavelength irradiation parameter set according to the near-infrared transmittance spectrum parameter matrix, and registering the multi-wavelength irradiation parameter set with the space-time coordinates of the thermal polymerization reaction front to obtain a scanning control matrix; sequentially projecting near-infrared light beams of different wavelengths and powers on the front surface of the die according to the scanning control matrix, exciting the up-conversion luminescence nanoparticles at the corresponding coordinates in the light-shielded area to release ultraviolet light, and forming a spatially uniform endogenous ultraviolet light field; maintaining the synchronous overlap of the endogenous ultraviolet light field and the thermal polymerization reaction front until the resin conversion rate in the light-shielded area reaches the preset conversion rate standard, to form a continuous initial curing resin layer.

[0011] Preferably, the use of near-infrared light transmits excitation of the up-conversion luminescence nanoparticles in the light-shielded area to release ultraviolet light, and cooperates with the thermal polymerization reaction front to initiate resin polymerization in the light-shielded area, to obtain a continuous initial curing resin layer, comprising: performing optical simulation on the silicon-based wiring layer, through hole and metal interconnection structure according to the GPU die layout structure data, obtaining a wavelength transmittance matrix corresponding to the coordinates, and obtaining a near-infrared transmittance spectrum parameter matrix; constructing a multi-wavelength irradiation parameter set according to the near-infrared transmittance spectrum parameter matrix, and registering the multi-wavelength irradiation parameter set with the space-time coordinates of the thermal polymerization reaction front to obtain a scanning control matrix; sequentially projecting near-infrared light beams of different wavelengths and powers on the front surface of the die according to the scanning control matrix, exciting the up-conversion luminescence nanoparticles at the corresponding coordinates in the light-shielded area to release ultraviolet light, and forming a spatially uniform endogenous ultraviolet light field; maintaining the synchronous overlap of the endogenous ultraviolet light field and the thermal polymerization reaction front until the resin conversion rate in the light-shielded area reaches the preset conversion rate standard, to form a continuous initial curing resin layer.

[0012] Preferably, the method of applying an alternating magnetic field to the continuous primary cured resin layer to excite the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent heat initiators to complete deep curing and simultaneously release internal residual stress to obtain a fully cured resin layer, comprises: calculating a magnetic field parameter set based on the volume distribution data of the magnetic induction heating nanoparticles in the continuous primary cured resin layer and the thermal conductivity data of the resin, the magnetic field parameter set including the magnetic field frequency and the field strength, and setting an alternating magnetic field based on the magnetic field parameter set; applying a first stage alternating magnetic field to the peripheral area of the continuous primary cured resin layer according to the magnetic field parameter set, so that the peripheral magnetic induction heating nanoparticles generate induction heat and activate the remaining latent heat initiators in the peripheral area to obtain a peripheral deep curing area; applying a second stage alternating magnetic field to the central area of the continuous primary cured resin layer using the heat diffusion gradient generated by the peripheral deep curing area, adjusting the magnetic field duty cycle to maintain the resin temperature in the central area at the latent heat initiator activation temperature zone, and forming a central deep curing area; continuing to apply a third stage alternating magnetic field to the peripheral deep curing area and the central deep curing area, gradually reducing the magnetic field amplitude and introducing phase modulation to promote the completion of relaxation of the resin molecular chain segments and redistribute the internal residual stress; and stopping the alternating magnetic field when the heat flux density reaches a steady state and there is no further exothermic change, thereby forming a fully cured resin layer.

[0013] Preferably, the method of applying a third stage alternating magnetic field to the peripheral deep curing area and the central deep curing area to gradually reduce the magnetic field amplitude and introduce phase modulation to promote the completion of relaxation of the resin molecular chain segments and redistribute the internal residual stress, comprises: generating a differential magnetic field strength distribution based on the difference in curing degree between the peripheral deep curing area and the central deep curing area, applying a strong magnetic field mode to the area whose curing degree does not reach a preset curing standard, and applying a weak magnetic field mode to the area whose curing degree has reached the preset curing standard, the magnetic field strength of the weak magnetic field mode being smaller than that of the strong magnetic field mode; generating pulsed heating of the magnetic induction heating nanoparticles through magnetic field phase periodic modulation, and using periodic temperature fluctuations to induce cooperative thermal motion of the resin molecular chain segments to release the curing shrinkage stress; setting a magnetic field amplitude decay curve according to the thermal expansion matching requirements of the multilayer heterogeneous interface, so that the thermal shrinkage processes of the resin layer, the die, and the substrate are kept synchronous, and the interface shear stress concentration is inhibited; and switching to a constant low amplitude maintenance mode when the magnetic field amplitude drops to a lower limit value, maintaining weak heating of the particles until the temperature change rate and the stress change rate are both lower than the steady state determination standard, thereby completing the redistribution of residual stress.

[0014] Preferably, the gradient cooling and isothermal aging are performed on the fully cured resin layer from the substrate side to the die side to redistribute and stabilize the curing shrinkage stress, and a GPU chip packaging structure is obtained, including: establishing a temperature gradient curve according to the difference between the thickness of the fully cured resin layer and the linear expansion coefficient of the die and the substrate, setting the starting temperature of the substrate side and the target temperature of the die side, and generating a gradient cooling parameter set; performing segmented cooling from the substrate side to the die side according to the gradient cooling parameter set, maintaining a constant temperature platform at the end of each temperature segment by using the residual heat of the resin curing exothermic, forming a multi-stage gradient cooling process; maintaining the constant temperature section after the gradient cooling is completed until the internal heat flux density of the resin is lower than the preset threshold, and applying a micro-vibration mechanical wave to the packaging structure in the constant temperature stage to promote chain segment relaxation and stress redistribution; confirming that the temperature gradient in the resin layer is less than the temperature uniformity standard and the stress change rate tends to zero at the end of the constant temperature stage, completing the isothermal aging, and forming a GPU chip packaging structure.

[0015] The second aspect of the application provides a GPU chip forming device, including: a functional resin injection module for injecting a bottom filling resin containing ultraviolet light initiators, latent thermal initiators, up-conversion luminescent nanoparticles and magnetic induction heating nanoparticles into the gap between the die and the substrate to form a fully filled functional resin layer; a circumferential ultraviolet curing module for implementing ultraviolet light irradiation on the peripheral area of the functional resin layer to make the ultraviolet light initiators polymerize and release heat and activate the latent thermal initiators, generate a thermal polymerization reaction front that advances from the outside to the inside of the resin layer, and obtain a peripheral cured resin area; a near-infrared transmission excitation module for transmitting near-infrared light through the die to excite the up-conversion luminescent nanoparticles to release ultraviolet light in the shaded area, and cooperatively induce resin polymerization in the shaded area with the thermal polymerization reaction front to obtain a continuous initial cured resin layer; an alternating magnetic field deep curing module for applying an alternating magnetic field to the continuous initial cured resin layer to excite the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent thermal initiators to complete deep curing and simultaneously release internal residual stress, and obtain a fully cured resin layer; a gradient cooling and aging module for performing gradient cooling and isothermal aging on the fully cured resin layer from the substrate side to the die side to redistribute and stabilize the curing shrinkage stress, and obtain a GPU chip packaging structure.

[0016] The technical scheme provided by the embodiment of the application, at the material level, establishes a light-heat cascade reaction mechanism by introducing a synergistic system of ultraviolet initiators and latent thermal initiators into the bottom filling resin. When the peripheral area is irradiated by ultraviolet light, the polymerization heat generated by the polymerization reaction excited by the light initiator can reach the activation temperature threshold of the latent thermal initiator, thereby triggering the secondary thermal initiation polymerization. This thermal polymerization reaction has a self-sustaining characteristic, and the reaction front formed can advance to the center area of the chip at a stable rate, breaking through the limitation of the physical boundary of light irradiation. The key of this mechanism lies in the matching of the exothermic characteristic of the polymerization reaction and the temperature response characteristic of the thermal initiator, which enables the reaction to continue without external light. At the same time, the introduction of upconversion luminescent nanoparticles realizes the conversion of light shielding to light transmission. Since the silicon-based chip has good transmission characteristics for specific near-infrared light, near-infrared light can penetrate the chip body to reach the bottom shielding area. The upconversion luminescent nanoparticles convert low-energy near-infrared photons into high-energy ultraviolet photons through the anti-Stokes process, forming a distributed light source inside the shielding area. This endogenous light field and the advancing thermal polymerization reaction front form a spatial-temporal synergistic coupling, ensuring that the resin in the shielding area can be simultaneously excited by photochemical excitation and thermal chemical excitation. The dispersion system of the magnetic induction heating nanoparticles further perfects the deep curing mechanism. After the light-heat synergistic preliminary curing is completed, the nanoparticles are excited to generate internal induction heat by applying an alternating magnetic field. This endogenous heating method has spatial selectivity and temperature controllability. Unlike the traditional bulk heating method, magnetic induction heating only acts on the resin area containing nanoparticles, avoiding the thermal stress impact on the chip and substrate. More importantly, this internal uniform heating can activate the residual latent thermal initiator, complete the deep crosslinking reaction, and at the same time realize the stress release and rearrangement of molecular chains by controlling the time sequence of the heating process. Gradient cooling and constant temperature aging treatment solve the problem of thermal stress matching of multi-material interfaces. Since the GPU chip, the bottom filling resin and the substrate have different thermal expansion coefficients, the traditional uniform cooling method is easy to cause stress concentration at the interface. By establishing a temperature gradient from the substrate side to the chip side, the thermal contraction process of different materials shows an ordered characteristic, and the stress is preferentially released in the low-sensitive interface, reducing the stress accumulation in the high-stress-sensitive area. The constant temperature aging stage further promotes the relaxation and rearrangement of polymer molecular chains, so that the entire packaging system reaches a stress equilibrium state. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 An embodiment schematic diagram of the GPU chip packaging method in the embodiment of the application; Figure 2 An embodiment schematic diagram of the GPU chip forming device in the embodiment of the application. DETAILED DESCRIPTION

[0018] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0019] An embodiment of the present application provides a GPU chip packaging method. Figure 1 An embodiment of the present application provides a GPU chip packaging method. Please refer to Figure 1 The gap between the die and the substrate is injected with a bottom filling resin containing an ultraviolet initiator, a latent thermal initiator, up-conversion luminescent nanoparticles and magnetic induction heating nanoparticles to form a fully filled functional resin layer. In an embodiment of the present application, the gap between the die and the substrate is injected with a bottom filling resin containing an ultraviolet initiator, a latent thermal initiator, up-conversion luminescent nanoparticles and magnetic induction heating nanoparticles to form a fully filled functional resin layer, including: calculating the light shielding coefficient according to the die area and the HBM stacking height, setting the molar ratio of the ultraviolet initiator and the latent thermal initiator according to the light shielding coefficient, preparing the epoxy resin matrix and obtaining the resin premix liquid; performing hydroxyl surface modification on the up-conversion luminescent nanoparticles to form hydroxyl sites on the particle surface which can condense with epoxy groups, adding the modified particles to the resin premix liquid and dispersing under constant shear conditions to obtain an up-conversion dispersion liquid; adding magnetic induction heating nanoparticles to the up-conversion dispersion liquid, performing shear homogenization and vacuum exhaust combined treatment to remove suspended bubbles and make the two types of functional particles uniformly distributed in the system to obtain a functional bottom filling resin; maintaining the gap between the die and the substrate under negative pressure in a constant temperature environment, and applying a local infrared radiation heat field to the outer edge of the gap to establish a temperature gradient and a viscosity gradient decreasing from the outer edge to the center, according to the viscosity gradient, the functional bottom filling resin is injected into the outer edge region and the center region of the gap in two stages to form a preliminary filling resin layer; maintaining the negative pressure and supplementing the functional bottom filling resin in a stepwise feeding mode, while applying micro-vibration mechanical waves to the two sides of the gap to make the residual bubbles migrate to the outer edge and be discharged to form a fully filled functional resin layer.

[0020] The following is a specific description of the steps involved in the above embodiments: the calculation of the light shielding coefficient is based on the geometric parameters and optical shielding characteristics of the GPU chip. HBM is the abbreviation of High Bandwidth Memory, which is a three-dimensional stacked DRAM memory, and is integrated in the form of multi-layer stacking on the periphery or top of the chip through the through silicon via technology. The specific calculation process is as follows: use a caliper or laser measuring instrument to measure the length, width and thickness of the GPU die, and calculate the die projection area; measure the height and lateral size of the HBM stack to determine its shielding range for ultraviolet light; through the TracePro optical simulation software or using an ultraviolet radiometer to determine the light intensity attenuation data at different positions, the relative light intensity of each region is calculated. The light shielding coefficient is defined as the ratio of the light intensity at the center of the chip to the light intensity at the edge, and the smaller the value, the more serious the light shielding. Taking a typical 28x28mm GPU chip as an example, when equipped with a double-layer HBM stack, the ultraviolet light intensity received at the center is only 15-25% of that at the edge, so the light shielding coefficient is 0.15-0.25. Based on the light shielding coefficient, the molar ratio of the ultraviolet initiator and the latent thermal initiator is determined: when the light shielding coefficient is 0.2, the light initiator Irgacure 184 and the thermal initiator azobisisobutyronitrile are selected, and the molar ratio is set to 1:2.5, to ensure that the heat generated by thermal polymerization can compensate for the deficiency of photopolymerization in the light shielding area. The epoxy resin matrix selected is a bisphenol A type epoxy resin with a molecular weight of 380-400, and the double-initiator system is uniformly dispersed in the matrix by a planetary mixer, with a stirring speed of 300-500 rpm and a stirring time of 15-20 minutes. During this period, stop the machine every 5 minutes to scrape the wall once, to obtain a resin premix liquid with uniform color and no agglomerates. This quantitative ratio design based on the light shielding coefficient directly addresses the uneven light problem of large-area GPU chips in a completely dark workshop environment, and through the pre-compensation thermal curing mechanism, even in areas where ultraviolet light cannot reach, complete curing can still be achieved through thermal polymerization, fundamentally solving the technical bottleneck of insufficient curing in the center area in traditional processes.

[0021] The up-conversion luminescent nanoparticles are subjected to hydroxyl modification by a silane coupling agent treatment process. The up-conversion luminescent nanoparticles refer to rare earth doped nanomaterials that can convert low-energy near-infrared light into high-energy ultraviolet light. In this embodiment, NaYF4:Yb,Er nanocrystals are selected, in which Yb 3+ ions act as sensitizers to absorb 980nm near-infrared light, and Er 3+The ion emits 365 nm ultraviolet light as an activator. 3-aminopropyltriethoxysilane (APTES) was chosen as the coupling agent. The upconversion luminescent nanoparticles with a particle size of 20-50 nm were dispersed in anhydrous ethanol by ultrasonic dispersion. An ultrasonic cleaner with a power of 300 W was used to treat for 30 minutes to ensure that the particles were completely dispersed without agglomeration. A mass fraction of 2% APTES coupling agent was added, and the reaction was carried out in a three-necked flask equipped with a reflux condenser at 60°C for 4 hours under nitrogen protection to prevent oxidation. The hydroxyl site refers to the Si-OH groups formed on the surface of the nanoparticles through the hydrolysis-condensation reaction of APTES. These groups have active hydrogen atoms that can undergo ring-opening addition reaction with the epoxy groups of the epoxy resin to form chemical bonding. After the reaction was completed, the nanoparticles were separated by centrifugation at 8000 rpm for 10 minutes, washed with anhydrous ethanol 3 times to remove unreacted coupling agent, and dried at 80°C under vacuum for 2 hours to obtain the surface-hydroxylated nanoparticles. The modified particles were added to the resin premix at a ratio of 3-5% of the total mass of the resin. An IKA high-shear disperser was used to disperse the particles at a constant shear rate of 2000 s -1 The dispersion was carried out for 60 minutes. The constant shear condition means that the shear rate value is kept stable during the dispersion process, and the speed fluctuation is controlled within ±2% to avoid particle re-agglomeration or surface modification layer damage caused by uneven shear force. During the dispersion process, samples were taken every 15 minutes, and the particle size distribution was detected by Malvern laser particle size analyzer. The D90 value was controlled below 80 nm to indicate good dispersion.

[0022] FeCo alloy magnetic induction heating nanoparticles with high saturation magnetization and excellent high-frequency induction heating performance are added to the up-conversion dispersion. Magnetic induction heating nanoparticles refer to magnetic nanomaterials that can generate heat through hysteresis loss and eddy current loss under the action of an alternating magnetic field. The particle size is controlled within the range of 80-120 nm to balance the dispersion stability and heating efficiency. The nanoparticles are added at a proportion of 1-2% of the total mass of the resin, and a three-roll mill is used for shear homogenization. Shear homogenization refers to the destruction of van der Waals forces and electrostatic attraction between particles through mechanical shear force, allowing the two types of functional nanoparticles to achieve uniform distribution in three-dimensional space without phase separation. The front, middle, and rear roller gaps of the three-roll mill are set to 10 μm, 8 μm, and 5 μm, respectively, and the rotation speed ratio is set to 1:3:9. The material is continuously passed through 3-5 times until no visible agglomerates are observed under a microscope. Vacuum exhaust is performed simultaneously during the grinding process, and a rotary vane vacuum pump is used to control the system vacuum to -0.08 to -0.09 MPa, with an exhaust time of 30-45 minutes. Suspended bubbles refer to small bubbles introduced into the high-viscosity resin system due to stirring, feeding, and other operations. The diameter of these bubbles is usually within the range of 10-100 μm, and these bubbles can form void defects during the curing process, severely affecting the mechanical strength and electrical insulation performance of the package. Vacuum exhaust increases the pressure difference between the inside and outside of the bubbles by reducing the system pressure, driving the bubbles to migrate to the surface and break and escape. At the same time, the negative pressure environment reduces the solubility of gases in the resin, promoting the precipitation of dissolved gases. This dual-functional nanoparticle composite system provides unique multi-physical field response capabilities for GPU packaging in a full-black workshop environment. The up-conversion particles solve the problem of light blind area, and the magnetic induction particles provide internal heating capacity. The two work together to ensure uniform deep curing under large-area light-shielding conditions.

[0023] A stable negative pressure environment of -0.05 to -0.07 MPa is established in the gap between the die and the substrate using a vacuum pump through a special joint in a constant temperature environment at a temperature of 25±1°C. The local infrared radiation heat field at the outer edge of the gap is generated by a ceramic infrared heater installed in a ring-shaped position 2-3 mm away from the edge of the chip, with a power density of 0.8-1.2 W / cm 2, the irradiation angle is 45° to avoid direct irradiation of the chip surface. The temperature gradient refers to the temperature distribution field formed by infrared radiation, which decreases from the outer edge to the center. The outer edge temperature is controlled at 40-45°C by an infrared thermometer, and the center area is maintained at 25-30°C, with a radial temperature gradient of about 1.5-2.0°C / mm. The viscosity gradient is a direct result of the temperature gradient. According to the Arrhenius equation, the viscosity of the resin decreases exponentially with temperature. The viscosity at 40°C is about 60-70% of that at 25°C. This viscosity difference provides the basis for flow control in the staged filling process. In the two-stage injection process, the first stage uses a precision metering pump to inject functionalized underfill resin to the outer edge area of the gap. The injection rate is controlled at 0.2-0.3 ml / min. The low viscosity characteristics of the outer edge area and the negative pressure driving force are used to achieve rapid wetting and preliminary filling. The second stage changes the injection position to above the center area of the chip, and the injection rate is reduced to 0.1-0.15 ml / min. Capillary force and negative pressure gradient are used to achieve complete filling. During the injection process, the filling process is monitored in real time by an ultrasonic liquid level sensor. The sensor is installed on both sides of the gap, with a monitoring frequency of 5 MHz. It can detect the position and speed of the resin front. The pressure sensor is used to monitor the pressure change in the gap. When the pressure fluctuation exceeds ±0.005 MPa, it indicates the presence of bubbles or uneven filling. When abnormal signals are detected, the control system automatically adjusts the negative pressure or temperature parameters. This filling process based on temperature-viscosity gradient control is specifically designed for the large area and thin gap characteristics of GPU chips. Through precise fluid mechanics control, it avoids the common problems of bubble entrapment and incomplete filling in traditional uniform filling methods, creating ideal initial conditions for subsequent uniform solidification.

[0024] On the basis of maintaining -0.05 to -0.07 MPa negative pressure, the functionalized underfill resin is continuously supplemented by using a step feeding method. The step feeding refers to a process method of supplementing resin in multiple small doses. The amount of each feeding is accurately controlled to be 5-8% of the remaining space volume, the interval time between feedings is 3-5 minutes, and a total of 4-6 feeding cycles are performed. The physical basis of this feeding strategy is to compensate for the volume shrinkage of the resin under the action of negative pressure environment and temperature gradient, ensuring that there is no hollow area inside the gap. At the same time, piezoelectric ceramic vibrators are installed on the top surface of the bare chip and the bottom surface of the substrate to generate micro-vibration mechanical waves to assist bubble removal. The vibrator selects PZT-5 type piezoelectric ceramic, and the excitation voltage is controlled at 50-80V to generate micro-mechanical vibration with a frequency of 20-40 kHz and an amplitude of 2-5μm. The mechanical wave propagates into the resin system through the solid, and its mechanism is to break the surface tension balance of the bubble and resin interface through periodic compression-tension cycle, and at the same time to provide driving force to make the bubble migrate to the area with lower pressure. Residual bubbles refer to small bubbles that are still dispersed in the resin after vacuum degassing and preliminary filling. The diameter of these bubbles is usually in the range of 1-20μm, and these bubbles are difficult to completely remove by conventional methods due to their small size and scattered distribution. Under the action of 20-40 kHz mechanical waves, the bubbles undergo forced vibration, and when the vibration frequency approaches the resonance frequency of the bubbles, the amplitude increases sharply, and the bubble wall is subjected to periodic stress and gradually migrates and coalesces. The migration process follows Stokes' law, and the drag force on the bubble is proportional to its radius, so larger bubbles prefer to move to the outer low-pressure area, and finally they are removed from the system through the negative pressure exhaust channel. The entire feeding and mechanical wave assisted bubble removal process lasts for 15-20 minutes, during which a high-frequency ultrasonic detector is used to monitor the change of bubble content in the gap. When the detection signal shows that the volume fraction of bubbles is reduced to less than 0.1%, the process is stopped. The functionalized resin layer obtained after the process forms a continuous, homogeneous, and defect-free three-dimensional distribution state in the gap between the chip and the substrate, which lays the foundation for the successful implementation of subsequent multi-physical field collaborative curing process in a full-black workshop environment, and ensures the high reliability and process consistency of large-area GPU packaging.

[0025] In one embodiment of the present application, the up-conversion luminescent nanoparticles are subjected to hydroxylated surface modification to form hydroxyl sites on the particle surface that can condense with epoxy groups, the modified particles are added to the resin premix and dispersed under constant shear to obtain an up-conversion dispersion, comprising: regulating the coordination environment of rare earth ions on the surface of up-conversion luminescent nanoparticles by silane coupling agent treatment, building a directional hydroxyl functional layer on the particle surface, and making the hydroxyl density reach a preset standard to meet the conversion efficiency requirement of near-infrared light and ultraviolet light under full black environment; determining the concentration gradient distribution of up-conversion luminescent nanoparticles in the resin according to the spatial distribution characteristics of the light-shielding area of the GPU chip, and forming a concentration distribution matching the light-shielding intensity of the modified particles in the resin premix by controlling the spatial non-uniformity of the shear field; synchronously applying an ultrasonic field during the constant shear dispersion process, using cavitation to break up particle agglomeration and activate the hydroxylated surface, and forming a pre-crosslinking bond with the epoxy group to enhance the particle-resin interface bonding strength; controlling the pre-crosslinking density of the particle surface by adjusting the corresponding relationship between the shear rate and the ultrasonic power, so that the up-conversion luminescent nanoparticles maintain stable light conversion performance in the subsequent excitation process, and an up-conversion dispersion is obtained.

[0026] The following describes the steps involved in the above embodiment: The silane coupling agent treatment process regulates the coordination environment of rare earth ions on the surface of up-conversion luminescent nanoparticles by chemical modification. The coordination environment of rare earth ions refers to the spatial arrangement and electron cloud distribution state of the coordination atoms or molecules around the rare earth ions, which directly affects the energy level structure and optical transition efficiency of the rare earth ions. 3-aminopropyltriethoxysilane (APTES) is selected as the coupling agent to undergo hydrolysis-condensation reaction on the surface of NaYF4:Yb,Er up-conversion luminescent nanoparticles. The specific operation process is as follows: the up-conversion luminescent nanoparticles with an average particle size of 30 nm are ultrasonically dispersed in anhydrous toluene for 30 minutes, 3% by mass fraction of APTES coupling agent is added, and the reaction is carried out at 80°C for 6 hours in a three-necked flask equipped with a reflux condenser and a nitrogen protection device. During the reaction, the ethoxy groups of the APTES molecules are hydrolyzed to form silicon hydroxyl groups, which then condense with the hydroxyl groups on the particle surface to form Si-O-Si covalent bonds, while the amino end is exposed on the particle surface. The directional hydroxyl functional layer refers to the formation of an array of hydroxyl groups with a specific orientation on the particle surface through ring-opening reaction of the amino group with the epoxy group. The density and arrangement of these hydroxyl groups directly determine the interfacial bonding strength with the resin matrix. The hydroxyl density is determined by X-ray photoelectron spectroscopy (XPS), and the ratio of the intensity of the O 1s peak at 532.8 eV to the total surface area of the particles is used to represent the hydroxyl density. The value is controlled within 2.5-3.2 hydroxyl groups / nm 2 2 ​At this time, the quantum efficiency of the modified upconversion luminescent nanoparticles emitting 365 nm ultraviolet light reaches 2-3%, and this efficiency level can provide sufficient intensity of endogenous ultraviolet light to excite the polymerization reaction of the photosensitive resin in the light-shielded area of the GPU chip. For example, for a GPU chip with an area of 28x28 mm, the minimum ultraviolet light power density required by the central light-shielded area is 15-20 mW / cm 2 , and the modified nanoparticles can provide corresponding intensity of endogenous ultraviolet light under the excitation of near-infrared light at this power density. The surface modification process significantly improves the light conversion stability of the upconversion luminescent nanoparticles in a full-black workshop environment, avoids non-radiative transition loss caused by surface defects, and ensures stable ultraviolet light output intensity during long-term near-infrared light excitation.

[0027] According to the spatial distribution characteristics of the light-shielded area of the GPU chip, the concentration gradient distribution of the upconversion luminescent nanoparticles is designed. The spatial distribution characteristics refer to the light intensity distribution law of the three-dimensional light-shielded geometry formed by the GPU die and the HBM stack on the two-dimensional projection plane. The three-dimensional geometry of the GPU chip is established by the optical simulation software TracePro, and the 365 nm ultraviolet light is set to be incident from different angles to calculate the light intensity distribution of each point in the bottom filling area. The light-shielded intensity is defined as the ratio of the light intensity received by the point to the light intensity without shielding, and the smaller the value, the more serious the light shielding. The concentration gradient distribution refers to the non-uniform distribution state of the upconversion luminescent nanoparticles in the resin, and the particle concentration changes with the position in a gradient distribution, with high particle concentration in the high light-shielded area and low particle concentration in the low light-shielded area. In specific implementation, the resin premix liquid is divided into multiple containers, and different concentrations of upconversion luminescent nanoparticle dispersions are adjusted according to the light-shielded intensity data. The area with a light-shielded intensity of 0.1-0.2 corresponds to a particle mass fraction of 4-5%, the area with a light-shielded intensity of 0.2-0.4 corresponds to a particle mass fraction of 3-4%, and the area with a light-shielded intensity greater than 0.4 corresponds to a particle mass fraction of 2-3%. The spatial non-uniformity of the shear field is realized by a special gradient mixer, which has adjustable shear gap and rotational speed distribution and can generate a spatially varying shear intensity field during mixing. For example, when processing the bottom filling resin of a 28x28 mm GPU chip, the shear rate in the center area of the mixer is set to 3000 s -1 , the shear rate in the edge area is set to 1500 s -1 , and the shear rate gradient is about 75 s -1 / mm, so that high-concentration particles are preferentially dispersed in the center high-shear area, and low-concentration particles are mainly distributed in the edge low-shear area. This concentration gradient distribution design directly compensates for the non-uniformity of light caused by the complex geometry of the GPU chip, provides more endogenous light conversion units in the light-shielded area, realizes the spatial uniformization of the endogenous ultraviolet light field in the light-shielded area, and solves the physical shielding problem that cannot be overcome by traditional external illumination.

[0028] In the constant shear dispersion process, the ultrasonic field is applied synchronously to achieve the dual effect of breaking the particle agglomeration and activating the surface. The constant shear dispersion refers to keeping the shear rate constant at 2500±50 s -1 -1 throughout the dispersion process. A stable shear force field is generated by the rotor-stator system of a high-shear disperser. The ultrasonic field is generated by an ultrasonic processor with a power of 500 W and a frequency of 25 kHz. The ultrasonic probe is directly inserted into the dispersion container and is synchronized with the shear dispersion. Cavitation refers to the physical phenomenon of micro-bubble formation, growth, and violent collapse when ultrasonic waves propagate in a liquid. The micro-jet and shock wave generated in the instant of collapse can effectively destroy the van der Waals force and electrostatic attraction between nanoparticles, breaking the agglomerates with a particle size greater than 100 nm into a single dispersion state. Hydroxyl group activation refers to the high-energy shock wave generated by ultrasonic cavitation, which provides additional activation energy to the hydroxyl groups on the particle surface, enhancing the reactivity with the epoxy groups. Epoxy groups refer to the three-membered ring ether structure in epoxy resin molecules, which have high reactivity and can undergo ring-opening reaction with hydroxyl groups to form ether bonds and covalent bonding of hydroxyl groups. Pre-crosslinking bonding refers to the preliminary chemical bonding between the hydroxyl groups on the surface of the modified particles and some epoxy groups before the formal curing reaction, forming chemical bonding rather than physical adsorption. The interfacial bonding strength is determined by the single fiber pull-out test, and the interfacial shear strength of the modified particles-resin reaches 25-30 MPa, which is 60-80% higher than that of unmodified particles. For example, in the treatment of a resin system containing 4% up-conversion luminescent nanoparticles, after 60 minutes of shear-ultrasonic combined treatment, the average agglomeration degree of the particles is reduced from the initial 2.8 to 1.2, and the pre-crosslinking degree of the surface hydroxyl groups with the epoxy groups reaches 15-20%. This pre-crosslinking structure ensures the stability of the interface during subsequent near-infrared light excitation, avoids the separation of the particle-resin interface due to thermal stress or mechanical stress, and ensures the continuous and stable output of the endogenous ultraviolet light field.

[0029] The corresponding relationship between the shear rate and the ultrasonic power is adjusted to accurately control the pre-crosslinking density of the particle surface. The corresponding relationship refers to the quantitative matching rule between the shear rate and the ultrasonic power, and the optimal parameter combination is obtained by fitting a large amount of experimental data. When the shear rate is 2000 s -1 -1, the ultrasonic power is set to 400 W; when the shear rate is 2500 s -1 -1, the ultrasonic power is 500 W; when the shear rate is 3000 s -1 -1, the ultrasonic power is 600 W, and the power to shear rate ratio is maintained within the range of 0.2 W·s. The pre-crosslinking density refers to the number of hydroxyl groups forming pre-crosslinking bonds per unit surface area, which is monitored by an infrared spectrometer at 1650 cm -1The change of the intensity of the hydroxyl stretching vibration peak is quantitatively characterized. The pre-crosslinking density is controlled in the range of 1.2-1.8 bonds / nm 2 The range, too low will result in insufficient interface bonding force, too high will consume too much epoxy group affect the subsequent formal curing reaction. The stability of light conversion performance is verified by 12 hours of near infrared light excitation test, monitoring the decay of 365 nm ultraviolet light output intensity. Stable light conversion performance requires that the decay rate of ultraviolet light output intensity be less than 5% during continuous excitation, and the retention rate of light conversion efficiency be greater than 90% after 500 times of temperature cycle-65°C to +150°C. For example, for the resin system loaded with 3.5% modified up-conversion luminescent nanoparticles, after the optimized shear-ultrasonic parameter treatment, at 980 nm, 250 mW / cm 2 After 8 hours of continuous near-infrared light excitation, the 365 nm ultraviolet light output intensity only decays by 2.8%, and the light conversion efficiency retention rate reaches 94% after 300 times of thermal cycling. This high stability of light conversion performance ensures that during the GPU chip packaging process in a completely dark workshop environment, even after a long time of process cycle and temperature change, the up-conversion luminescent nanoparticles can still provide a stable internal ultraviolet light field, ensuring the continuity and consistency of the resin curing reaction in the light-shielded area.

[0030] Please continue to refer to Figure 1 The peripheral area of the functional resin layer is irradiated with ultraviolet light, and the ultraviolet light initiator is polymerized and exothermed, and the latent thermal initiator is activated, a thermal polymerization reaction front is generated along the resin layer from the outside to the inside, and a peripheral cured resin area is obtained. In an embodiment of the present application, the peripheral area of the functional resin layer is irradiated with ultraviolet light, and the ultraviolet light initiator is polymerized and exothermed, and the latent thermal initiator is activated, a thermal polymerization reaction front is generated along the resin layer from the outside to the inside, and a peripheral cured resin area is obtained, including: according to the profile of the die edge and the thickness of the functional resin layer, the ultraviolet light is circularly partitioned and modulated to obtain a circumferential ultraviolet light field uniform in circumference; the circumferential ultraviolet light field is subjected to pulse timing control, and the pulse width and interval are set so that the polymerization exothermic peak temperature is in the activation temperature zone of the latent thermal initiator, and a stable exothermic pulse sequence is obtained; the ultraviolet light intensity is radially gradient-adjusted according to the stable exothermic pulse sequence, to obtain an energy distribution balanced along the thickness direction of the resin layer, forming a thermal polymerization reaction front advancing from the outside to the inside; the thermal polymerization reaction front is tracked in real time, and when the front reaches a set distance from the light-shielded boundary of the die, the irradiation parameters are kept constant until the exothermic is completed, forming a peripheral cured resin area.

[0031] The following is a specific description of the steps involved in the above embodiments: The annular partition modulation process spatially partitions the ultraviolet light irradiation according to the geometric characteristics of the GPU die outer edge profile and the thickness parameters of the functional resin layer. The die outer edge profile refers to the geometric shape of the GPU chip edge, including straight edge length, fillet radius, chamfer size, and other geometric parameters, which directly affect the incident angle and reflection path of the ultraviolet light. Taking a 28x28mm square GPU chip as an example, each of the four edges of the chip is 28mm long, and the four corners are fillet transitions with a radius of 2mm. This geometric configuration requires a corresponding spatial layout for the ultraviolet light irradiation system. The thickness of the functional resin layer is measured by an ultrasonic thickness gauge and controlled within the range of 50-80μm, with a uniformity deviation of less than ±5μm. The annular partition modulation refers to the division of the ultraviolet light irradiation area around the chip edge into multiple independently controlled sector segments, each corresponding to a specific light intensity and irradiation angle. In specific implementation, a ring array composed of 16 ultraviolet LED units is used, each LED unit corresponding to a 22.5° sector area, with a wavelength of 365nm and a maximum power of 5W per LED. The power output of each LED is independently adjusted by a programmable controller, and the light intensity distribution is adjusted according to the resin layer thickness and chip edge geometry of the corresponding segment. For example, the LED power in the straight edge area of the chip is set to 3.5-4.0W, and the LED power in the fillet transition area is set to 4.5-5.0W to compensate for the larger light scattering loss in the fillet area. The circumferentially uniform annular ultraviolet light field refers to the spatial light intensity distribution formed around the chip edge, with a light intensity deviation of less than ±8% in each azimuthal angle. The light intensity uniformity is verified by measuring the light power density at 16 equally spaced positions on the chip edge using an ultraviolet power meter, with each measurement point maintaining a light power density of 80-95mW / cm 2 This annular partition modulation technique is specifically designed for the large-area square geometry of GPU chips, overcoming the problem of insufficient light intensity in the corner areas of square chips in traditional uniform irradiation methods through precise light intensity control in spatial partitioning, ensuring the high uniformity of ultraviolet light irradiation in the peripheral area and creating ideal light field conditions for the stable formation of the thermal polymerization reaction front.

[0032] The pulse timing control system implements precise temporal modulation on the annular UV light field, and controls the exothermic characteristics of the polymerization reaction by setting specific pulse width and interval parameters. The pulse width refers to the duration of each UV light pulse, which is set to 150-200 ms. Within this time window, the UV photoinitiator fully absorbs photon energy and decomposes to generate free radicals. The interval time refers to the pause time between adjacent two pulses, which is set to 80-120 ms. This interval time allows the exothermic heat of the preliminary polymerization reaction to be fully conducted and the latent thermal initiator to be activated. The polymerization exothermic peak temperature refers to the highest value of the system temperature during the photoinitiated polymerization reaction, which is monitored in real time by a K-type thermocouple embedded in the functional resin layer, and the temperature rise rate is controlled at 15-25 °C / s. The latent thermal initiator activation temperature range refers to the temperature range at which azobisisobutyronitrile starts to decompose and generate free radicals, which is determined by differential scanning calorimetry (DSC) to be 65-85 °C. The key of the timing control is to make the polymerization exothermic peak temperature generated by pulse irradiation exactly fall within this activation temperature range, triggering the thermal initiated polymerization reaction. The stable exothermic pulse sequence refers to the periodic temperature pulsation pattern formed after timing optimization, and the temperature peak value of each pulse cycle is stable in the range of 70-80 °C, with a temperature fluctuation amplitude of less than ±3 °C. For example, for a functional resin system containing 2.5% azobisisobutyronitrile, when the UV light pulse power density is 90 mW / cm 2 , the pulse width is 180 ms, and the interval is 100 ms, the polymerization exothermic heat makes the system temperature rise from 25 °C to 75 °C, completely activating the latent thermal initiator and initiating the stable thermal polymerization reaction. In the specific implementation process, the feedback control system automatically adjusts the pulse parameters according to the real-time data of the temperature sensor. When the temperature peak value deviates from the target range, the controller automatically corrects the width or interval time of the next pulse to ensure the stability of the exothermic pulse sequence. This pulse timing control technology realizes the precise coupling of photoinitiated reaction and thermal initiated reaction, avoids the local overheating or insufficient temperature problem easily occurred in the continuous irradiation mode, and ensures the synchronous initiation and stable progress of the thermal polymerization reaction in the entire peripheral area.

[0033] The radial gradient adjustment technology modulates the intensity of ultraviolet light according to the temperature distribution characteristics of the stable exothermic pulse sequence, realizes the energy equalization distribution in the thickness direction of the resin layer. Radial refers to the spatial direction from the edge of the chip to the center, and radial gradient adjustment refers to the gradual adjustment of the intensity of ultraviolet light in this direction. In specific implementation, an adjustable optical lens is installed on each ring-shaped LED unit, and the focal length of the lens can be adjusted in the range of 50-200mm, and the divergence angle and irradiation depth of the light beam are controlled by changing the focal length. The energy distribution along the thickness direction of the resin layer is obtained by Monte Carlo optical simulation, and the simulation results show that when the light intensity on the surface of the resin layer is set to a reference value of 100%, the light intensity at a depth of 20μm from the surface should be adjusted to 85-90%, at a depth of 40μm to 70-75%, and at a depth of 60μm to 55-60%. Energy equalization distribution means that the photon energy density in unit volume remains relatively stable at different depths of the resin layer, avoiding the problem of excessive concentration of surface light energy and insufficient deep light energy. The thermal polymerization reaction front refers to the boundary of the active area of the thermal polymerization reaction, which advances at a certain speed from the periphery to the center of the chip. The front advance speed is determined by monitoring the movement of the temperature field by an infrared thermal imager, and the typical advance speed is 2-4mm / min. The self-outer-to-inner advancing feature is reflected by the change of the spatial distribution of the temperature field. In the initial stage, the high-temperature area is concentrated on the outer edge of the chip, and as the thermal polymerization reaction proceeds, the high-temperature area gradually expands inward. For example, for a 28x28mm GPU chip, after the thermal polymerization reaction front is started from the edge, it reaches a position 10mm away from the edge after 8-12 minutes, and reaches a position 14mm away from the edge after 15-20 minutes. The temperature gradient during the advance of the front is maintained at 3-5°C / mm. During the radial gradient adjustment process, the focal length of the optical lens is automatically adjusted according to the preset program. When the front advances to a position 5mm away from the edge, the lens focal length is adjusted to 150mm; when it advances to a position 10mm away from the edge, the focal length is adjusted to 120mm; and when it advances to a position 12mm away from the edge, the focal length is adjusted to 100mm, ensuring that there is enough light energy density in front of the front to maintain the reaction activity. This radial gradient adjustment technology solves the problem of uneven light energy distribution in the thick layer resin curing process, dynamically adjusts the light field distribution to ensure that the thermal polymerization reaction front maintains stable reaction intensity during the advancing process, and avoids the common phenomena of front decay and reaction interruption in the traditional fixed light intensity irradiation method.

[0034] The real-time tracking system continuously monitors the position and state of the thermal polymerization reaction front, enabling precise process control and parameter adjustment. Real-time tracking uses an infrared thermal imager for temperature field monitoring, with a model number of FLIRA6751sc, a temperature resolution of 0.02°C, and a frame rate of 1000Hz, capable of capturing rapidly changing temperature field distributions. Auxiliary monitoring uses a high-frequency impedance spectrum analyzer to detect changes in the dielectric constant during resin curing, with a frequency range of 1kHz-1MHz, capable of reflecting the progress of the polymerization reaction and changes in the curing degree. The identification of the thermal polymerization reaction front is determined by a temperature gradient threshold, and when the temperature difference between adjacent regions exceeds 10°C, it is identified as the front boundary. The light-shielded boundary refers to the spatial boundary line formed by the GPU die shielding ultraviolet light, and the boundary position coordinates are determined by optical simulation. The set distance refers to the predetermined position where the thermal polymerization reaction front stops advancing, which is 2-3mm away from the light-shielded boundary. This distance ensures that the front has enough reaction reserves to reach the edge of the shielding area, while avoiding excessive advancement that would conflict with the subsequent near-infrared excitation process. The constant illumination parameters refer to maintaining the current light field distribution state by keeping the power output, pulse timing, and optical lens focal length of all LED units unchanged when the front reaches the set distance. The criterion for determining the completion of heat release is that the temperature change rate detected by the infrared thermal imager decreases to below 1°C / min and lasts for more than 2 minutes. At the same time, impedance spectrum analysis confirms that the dielectric constant change rate is less than 0.1% / min, with double confirmation of reaction completion. For example, when processing a 28x28mm GPU chip, the thermal polymerization reaction front advances to a set position 2.5mm away from the light-shielded boundary after 18 minutes from the edge start. At this point, the control system locks all illumination parameters and continues to irradiate for an additional 8 minutes until the heat release is complete. The peripheral cured resin area refers to the cured area formed after light-thermal co-polymerization. The curing degree is determined by measuring the change in carbonyl peak intensity using an infrared spectrometer, with a curing degree of 75-85%. The formation of the peripheral cured resin area provides a stable reaction basis and heat source for the subsequent near-infrared excitation stage, and its geometry and curing degree directly affect the stress distribution and interfacial bonding strength of the entire packaging structure. Real-time tracking technology ensures the high controllability of the thermal polymerization reaction front advancement process, avoiding the risk of loss of control during the reaction process through precise position monitoring and parameter adjustment, ensuring the geometric accuracy and curing quality of the peripheral cured resin area, and providing a reliable guarantee for the stability and reproducibility of the GPU chip packaging process in a full-black workshop environment.

[0035] Please continue to refer to Figure 1 , using near-infrared light to transmit through the die to excite the upconversion luminescent nanoparticles in the light-shielded area to release ultraviolet light, and to cooperatively initiate resin polymerization in the light-shielded area with the thermal polymerization reaction front, obtaining a continuous initial cured resin layer; In an embodiment of the present application, the use of near-infrared light transmits excitation of the up-conversion luminescence nanoparticles through the die to release ultraviolet light in the light-shielded area, and cooperates with the thermal polymerization reaction front to initiate resin polymerization in the light-shielded area, obtaining a continuous initial curing resin layer, comprising: performing optical simulation on the silicon-based wiring layer, through hole and metal interconnection structure according to the GPU die layout structure data, obtaining a wavelength transmittance matrix corresponding to the coordinates, and obtaining a near-infrared transmittance spectrum parameter matrix; constructing a multi-wavelength irradiation parameter set according to the near-infrared transmittance spectrum parameter matrix, and registering the multi-wavelength irradiation parameter set with the space-time coordinates of the thermal polymerization reaction front to obtain a scanning control matrix; projecting near-infrared light beams of different wavelengths and powers in sequence on the front surface of the die according to the scanning control matrix, exciting the up-conversion luminescence nanoparticles at the corresponding coordinates to release ultraviolet light in the light-shielded area, and forming a spatially uniform endogenous ultraviolet light field; maintaining the endogenous ultraviolet light field and the thermal polymerization reaction front to be synchronous and overlapped until the resin conversion rate in the light-shielded area reaches a preset conversion rate standard, and forming a continuous initial curing resin layer.

[0036] The following steps involved in the above embodiments are specifically described: The optical simulation process is based on GPU die layout structure data to analyze the near-infrared light transmission characteristics of the multi-layer composite structure inside the chip. The layout structure data refers to the geometric layout file generated during the design of the GPU chip, which contains the accurate position, size and material information of each functional layer. The data format is GDSII or OASIS, which is converted into three-dimensional geometric coordinate information by layout extraction software. The silicon-based wiring layer refers to the metal wire layer made on the silicon substrate. A typical GPU chip contains 8-12 layers of metal wiring, each layer is 0.2-0.8 μm thick, the material is mainly copper, and the wiring width ranges from 0.1 to 10 μm. The via is a vertical conductive structure that connects different metal layers, with a diameter of 0.05-0.2 μm and a length of 0.3-1.5 μm, filled with tungsten or copper material inside. The metal interconnection structure refers to the three-dimensional conductive network composed of wiring layers and vias, which presents a complex geometric distribution on the chip cross-section. The optical simulation uses the finite difference time domain (FDTD) software Lumerical to calculate, imports the layout structure data into the software and sets the material optical parameters: the refractive index of the silicon substrate at 980 nm wavelength is 3.42, and the extinction coefficient is 0.001; the refractive index of copper is 0.94, and the extinction coefficient is 6.8; the refractive index of tungsten is 3.5, and the extinction coefficient is 2.8. The simulation region is set to the full size of the chip 28x28mm, the grid accuracy is set to 50nm, and the light source is set to a plane wave vertically incident from the front of the chip. The coordinate corresponding to the wavelength transmittance matrix refers to the light transmittance value matrix of each coordinate point on the bottom surface of the chip at different near-infrared wavelengths, with a matrix dimension of 560x560x41, corresponding to a 28x28mm chip divided into a coordinate grid with a spacing of 50 μm and 41 wavelength points with an interval of 1 nm in the wavelength range of 940-980 nm. The transmittance value is calculated by simulation, which represents the ratio of transmitted light intensity to incident light intensity. The near-infrared transmittance spectrum parameter matrix is a further processing of the coordinate corresponding to the wavelength transmittance matrix, which extracts the best transmission wavelength, peak transmittance, half peak width and other key parameters of each coordinate point, forming a 560x560x8 parameter matrix. For example, for the coordinate point (14mm, 14mm) located at the center of the chip, due to the high density of metal wiring in this area, the best transmission wavelength is 975nm, the peak transmittance is 15%, and the half peak width is 8nm; while the coordinate point (2mm, 2mm) located at the edge of the chip has a low density of metal wiring, the best transmission wavelength is 950nm, the peak transmittance can reach 65%, and the half peak width is 12nm. This optical simulation based on real layout structure avoids the error of traditional uniform material assumption, accurately reflects the influence of complex internal structure of GPU chip on near-infrared light transmission, and provides reliable physical basis data for subsequent accurate beam control.

[0037] The multi-wavelength irradiation parameter set construction process determines the optimal irradiation parameters for each coordinate point according to the data distribution characteristics of the near-infrared transmission spectrum parameter matrix. The multi-wavelength irradiation parameter set refers to the combination of near-infrared light irradiation parameters set for different coordinate positions, including wavelength, power density, irradiation time, beam diameter, and other parameters. The parameter set construction algorithm determines the irradiation parameters according to the peak transmittance and the best transmission wavelength of each coordinate point: when the peak transmittance is greater than 50%, the irradiation power density is set to 200-250 mW / cm 2 ; when the peak transmittance is in the range of 20-50%, the power density is set to 300-400 mW / cm 2 ; when the peak transmittance is less than 20%, the power density is set to 450-500 mW / cm 2 . The irradiation wavelength directly uses the best transmission wavelength corresponding to the coordinate point, and the irradiation time is determined according to the density of the up-conversion luminescent nanoparticles to be excited and the expected ultraviolet light output intensity, ranging from 2 to 8 seconds. The space-time coordinate registration refers to the process of corresponding matching the spatial coordinates of the multi-wavelength irradiation parameter set with the space-time evolution coordinates of the thermal polymerization reaction front. The space-time coordinates of the thermal polymerization reaction front are obtained by real-time tracking with the aforementioned infrared thermal imager, and the data format is a three-dimensional array of time-X coordinate-Y coordinate. The registration algorithm uses the nearest neighbor interpolation method to match the time of the front reaching each coordinate point with the start time of near-infrared light irradiation at that point, ensuring the time synchronization of near-infrared light excitation and thermal polymerization reaction. The scanning control matrix is the final output of the registration process, with a dimension of 560x560x12, containing 12 control parameters such as irradiation wavelength, power density, irradiation time, start time, beam diameter, and scanning speed for each coordinate point. For example, for the coordinate point (10mm, 15mm), the parameters in the scanning control matrix are: wavelength 965nm, power density 320mW / cm 2 , irradiation time 4.5 seconds, start time 12.3 minutes, beam diameter 200μm, scanning speed 50mm / s. The data processing process of the scanning control matrix also includes a path optimization algorithm, which uses a genetic algorithm to optimize the movement path of the near-infrared laser head to minimize the total scanning time and avoid repeated scanning areas. The path optimization result shows that for a 28x28mm GPU chip, the optimized scanning path has a total length of 1.8m and a total scanning time of 18 minutes, saving 35% of time compared to random scanning. This parameter set construction method based on physical simulation data ensures that each coordinate point can obtain the most suitable near-infrared light irradiation for its local structural characteristics, maximizing the light energy utilization efficiency and the excitation effect of the up-conversion nanoparticles.

[0038] The near-infrared beam projection process performs precise spatial-temporal-spectral three-dimensional scanning of the front of the GPU die according to the parameters of the scan control matrix. The near-infrared laser system uses a tunable diode laser with a wavelength tuning range of 940-980nm, a tuning accuracy of ±0.1nm, a maximum output power of 2W, and a beam quality of M 2 <1.2, two-dimensional scanning is achieved by an optical scanning head. The optical scanning head includes an XY galvanometer system with a scanning range of 30×30 mm, positioning accuracy of ±5 μm, and a maximum scanning speed of 200 mm / s. Beam projection of different wavelengths and powers is achieved through laser current modulation and an optical attenuator. Wavelength switching time is less than 50 ms, and the power adjustment range is 10-2000 mW / cm². The corresponding coordinate refers to the spatial position specified in the scan control matrix. The laser beam is precisely positioned to this coordinate by the galvanometer system and maintained for the set exposure time. The light shielding area refers to the area formed by the GPU die and HBM stack that is shielded from direct UV light, primarily within a 14×14 mm area at the center of the chip. The intrinsic UV field refers to the UV light distribution generated within the resin by upconversion luminescent nanoparticles under near-infrared light excitation. The spatial uniformity of the light field distribution is measured using a UV power meter array. The optical power meter array consists of 36 miniature UV detectors, spaced 2 mm apart, arranged in the light shielding area at the bottom of the chip to monitor the intensity distribution of the intrinsic UV light in real time. Spatial uniformity is characterized by the ratio of the standard deviation of light intensity to the average light intensity. A value less than 0.15 indicates uniform distribution. For example, during the scanning process of a 28×28mm GPU chip, when the near-infrared laser is projected to the coordinates (12mm, 16mm) according to the scanning control matrix parameters, the upconversion luminescent nanoparticles at this point are at 975nm and 350mW / cm 2 After 4 seconds of near-infrared light excitation, the generated 365nm ultraviolet light power density reaches 18mW / cm 2 When scanning to the coordinates (8mm, 12mm), at 960nm, 280mW / cm 2 15mW / cm after 3 seconds of excitation 2 After scanning the entire shaded area, the spatial uniformity of the internal UV light field reached 0.12, with an average light intensity of 16.5mW / cm². This precise scanning technology, based on the chip's structural features, enables efficient light energy transfer under physical shielding conditions, overcoming the physical limitations of external illumination through internal light field generation.

[0039] The synchronization overlap control system maintains the spatiotemporal coordination between the endogenous ultraviolet light field and the thermal polymerization reaction front until the light-shielded area reaches the preset polymerization conversion level. Synchronization overlap refers to the spatial and temporal overlap state of the excitation area of the endogenous ultraviolet light field and the active area of the thermal polymerization reaction front, which is achieved through real-time scheduling of the control system. The control system receives front position data from the infrared thermal imager and light field intensity data from the ultraviolet light power meter array, and adjusts the scanning parameters of the near-infrared laser using a predictive control algorithm. When the thermal polymerization reaction front advances to within 1 mm of a certain coordinate point, the control system starts near-infrared light irradiation at that point in advance to ensure that the endogenous ultraviolet light generated by the upconversion luminescent nanoparticles is synchronized with the front. The preset conversion rate standard refers to the target conversion degree of the resin polymerization reaction in the light-shielded area, which is determined by monitoring the intensity changes of the characteristic peaks of the epoxy group and the acrylate group using a Raman spectrometer. A 50% decrease in the characteristic peak intensity of the epoxy group at 915 cm -1 indicates that the conversion rate has reached 50%, and a 60% decrease in the double bond stretching vibration peak intensity of the acrylate group at 1635 cm -1 indicates that the conversion rate has reached 60%. The preset conversion rate standard is set to 65-75%, which ensures that the resin has sufficient mechanical strength and chemical stability while retaining some active groups for subsequent deep curing reactions. The continuous primary cured resin layer refers to the spatially continuous and uniformly performant polymer network structure formed after light-thermal cooperative polymerization. Continuity is confirmed by ultrasonic C-scan detection with a scanning frequency of 10 MHz, which can detect internal defects such as cavities and delamination in the resin layer with a detection accuracy of 5 μm. The detection results show that there are no obvious phase separation or cavity defects in the resin layer. For example, during the processing of a 28x28 mm GPU chip, the synchronization overlap between the endogenous ultraviolet light field and the thermal polymerization reaction front was maintained for 22 minutes, during which the average conversion rate of the light-shielded area gradually increased from 5% to 70%. Raman spectroscopy showed that the conversion rate of the epoxy group was 68% and the conversion rate of the acrylate group was 72%, meeting the preset conversion rate standard. The formed continuous primary cured resin layer has a uniform thickness of ±3 μm and a smooth surface roughness of Ra<0.5 μm, creating an ideal reaction basis for subsequent magnetic field-induced deep curing. This light field-thermal field cooperative control technology realizes uniform polymerization in a large-area light-shielded area in a completely black workshop environment, solving the fundamental problem of insufficient curing in the light-shielded area in traditional processes and ensuring the integrity and reliability of the GPU packaging structure.

[0040] In an embodiment of the present application, the near-infrared light beams of different wavelengths and powers are sequentially projected on the front surface of the die according to the scanning control matrix to excite the up-conversion luminescent nanoparticles at the corresponding coordinates to release ultraviolet light in the light-shielded area, forming a spatially uniform endogenous ultraviolet light field, comprising: establishing a dynamic focusing processing strategy for the near-infrared light beams based on the scanning control matrix, so that the focal point of the light beam is positioned at different depth key structure layers inside the die to compensate for the attenuation difference of the near-infrared light by each metal layer; adjusting the light beam residence time and repetition frequency according to the distribution density of the heat generating units inside the chip to suppress local temperature rise and ensure that the particles obtain sufficient excitation energy; using multi-beam parallel scanning to synchronously irradiate the light-shielded area in different zones, balancing the endogenous ultraviolet light intensity in each zone by phase modulation between the light beams, and eliminating the curing asynchronization phenomenon; real-time acquisition of the endogenous ultraviolet light field distribution in each scanning zone, automatic correction of the corresponding light beam parameters when the light intensity deviation exceeds the light intensity uniformity standard, and maintenance of the spatiotemporal stability of the endogenous ultraviolet light field in the light-shielded area.

[0041] The following specifically describes the steps involved in the above embodiment: the dynamic focusing processing strategy adjusts the focal point of the near-infrared light beam in real time based on the depth information of the scanning control matrix, achieving accurate positioning of the different depth key structure layers inside the GPU die. The dynamic focusing processing strategy refers to a control method that adjusts the focal point depth of the laser beam in real time according to the position information of the target coordinate point in the thickness direction of the chip. The GPU die contains multiple key structure layers: a surface passivation layer (thickness 1-2 μm), a top metal layer M8-M12 (depth 5-15 μm), an intermediate metal layer M4-M7 (depth 20-40 μm), a bottom metal layer M1-M3 (depth 50-80 μm), and a silicon substrate. The key structure layer refers to a functional layer that significantly affects the transmission of near-infrared light, mainly including high-density metal wiring layers and large-area metal filling areas. In specific implementation, an electric zoom lens system is used, with a focal length that can be continuously adjusted in the range of 80-300 mm, an adjustment accuracy of ±0.5 mm, and a response time of less than 100 ms. Each coordinate point in the scanning control matrix contains target depth information, and when the laser beam moves to a new scanning position, the control system automatically adjusts the focal length of the lens according to the depth data of the point. Attenuation difference compensation is achieved through layered adjustment of power density: for the surface passivation layer, the near-infrared light power density is set to 200 mW / cm 2 ; for the top metal layer area, the power density is increased to 350 mW / cm 2 ; for the intermediate metal layer area, the power density is set to 450 mW / cm 2 ; for the bottom metal layer area, the power density reaches 600 mW / cm 2 . For example, when scanning to the coordinate (8 mm, 12 mm) of the intermediate metal layer with a target depth of 35 μm, the control system adjusts the focal length of the lens to 150 mm, and at the same time sets the laser power density to 450 mW / cm2 , ensuring that the focus is accurately positioned at that depth and providing sufficient photon flux to excite the upconversion luminescent nanoparticles. This layered dynamic focusing technology addresses the differentiated impact of the complex three-dimensional structure of the GPU chip on near-infrared light transmission. Through precise depth control and power compensation, it ensures that the upconversion luminescent nanoparticles in each structural layer can obtain sufficient and uniform excitation energy, avoiding the problems of insufficient deep excitation and excessive shallow excitation in the traditional fixed focal length method.

[0042] The beam dwell time and repetition frequency adjustment strategies are differentiated based on the spatial distribution density of heat generating units within the GPU chip. Heat generating unit distribution density refers to the number of functional units that generate heat per unit area, including computing cores, cache units, power management units, etc., and is determined through chip thermal design power consumption data and layout analysis. High-density heat generating areas are mainly concentrated in the GPU core computing array, with a heat density of 150-200W / cm 2 The medium-density area includes the secondary cache and control unit, with a heat density of 80-120W / cm 2 The low-density area is mainly composed of I / O interfaces and auxiliary circuits, with a heat density of 30-60W / cm 2 . The beam dwell time refers to the time the laser beam stays at each scanning point, and the repetition frequency refers to the frequency of multiple irradiations of the same scanning point. The specific parameters are set as follows: in high-density heating areas, the dwell time is set to 2-3 seconds and the repetition frequency is 0.5Hz to avoid local overheating caused by the superposition of laser heating and chip self-heating; in medium-density heating areas, the dwell time is 3-4 seconds and the repetition frequency is 0.8Hz to balance the excitation efficiency and thermal control; in low-density heating areas, the dwell time is 4-6 seconds and the repetition frequency is 1.2Hz to fully excite the upconversion luminescent nanoparticles without worrying about overheating. Local temperature rise suppression is achieved through real-time monitoring by an infrared thermal imager. When it is detected that the temperature of a certain area exceeds the preset threshold, the control system automatically reduces the laser power in that area or extends the irradiation interval. For example, when scanning the coordinates of the GPU core computing area (14mm, 16mm), since the density of the heating units in this area is as high as 180W / cm 2 The system sets a dwell time of 2.5 seconds and a repetition rate of 0.5 Hz. It also monitors temperature changes using a thermal imager to ensure that the local temperature rise does not exceed 15°C. This parameter optimization strategy based on thermal characteristics ensures that the upconversion luminescent nanoparticles receive sufficient excitation energy while preventing local overheating from damaging the GPU chip's functions, achieving a balanced optimization of optical excitation and thermal management.

[0043] The multi-beam parallel scanning system realizes the partition synchronous irradiation and the light intensity equalization control of the light shielding area through the phase modulation technology. The multi-beam parallel scanning refers to the technology of using multiple independent near-infrared lasers to irradiate different areas at the same time, and the embodiment adopts 4 lasers to work in parallel, and each laser is responsible for a 7*7mm scanning area, covering the entire 28*28mm chip area. The phase modulation between the beams refers to adjusting the light field interference effect by controlling the phase relationship of each laser to realize the precise control of the light intensity distribution. The phase modulation is realized by an electro-optical modulator, the modulation frequency is 1-10kHz, and the phase adjustment accuracy is ±1°. The partition synchronous irradiation is realized through the timing coordination of the main controller, and the starting time difference of the 4 lasers is controlled within ±10ms, ensuring the high synchronization of the excitation process of each area. The core of the light intensity equalization control is to compensate for the excitation efficiency difference caused by the concentration difference of the up-conversion luminescent nanoparticles and the chip structure difference in different areas. Through the interference effect generated by the phase modulation, constructive interference can be formed in the phase synchronous area to enhance the light intensity, and destructive interference can be formed in the phase reverse area to reduce the light intensity. For example, when it is detected that the intensity of the endogenous ultraviolet light in the upper left area is 15% lower than that in the lower right area, the control system adjusts the phase relationship of the corresponding lasers, so that the upper left area forms constructive interference, the light intensity increases by 20%, and the lower right area forms slight destructive interference, the light intensity decreases by 8%, and finally the light intensity of the two areas is balanced. The curing asynchronization phenomenon refers to the inconsistent progress of the resin polymerization reaction in different areas, resulting in spatial differences in the final curing degree. Through the coordinated control of multi-beam parallel scanning, the up-conversion luminescent nanoparticles in each area are excited at the same time, and the endogenous ultraviolet light field generated maintains consistency in time and space, ensuring that the polymerization reaction in the entire light shielding area proceeds synchronously. This parallel scanning technology shortens the processing time of 28 minutes of single-beam sequential scanning to 7 minutes, greatly improving the process efficiency, and at the same time, the light intensity equalization control realized through the phase modulation ensures the high consistency of the endogenous ultraviolet light field in the large-area light shielding area.

[0044] The real-time light field acquisition system continuously monitors and automatically corrects the distribution of the internal source ultraviolet light field in each scanning area through an ultraviolet power meter array. The internal source ultraviolet light field distribution acquisition uses a detection array composed of 64 miniature ultraviolet power meters with a spacing of 3.5 mm uniformly distributed in the light-shielded area at the bottom of the 28x28 mm chip. The detection wavelength range of each power meter is 350-380 nm, the response time is less than 1 ms, and the detection accuracy is ±2%. The light intensity uniformity standard is defined as the ratio of the standard deviation of the light intensity of each detection point to the average light intensity. A value less than 0.12 indicates uniform distribution, and a value exceeding 0.15 requires correction. The data processing frequency of the real-time acquisition system is 100 Hz, which can timely capture the dynamic changes of the light field distribution. The automatic correction algorithm is based on the feedback control principle. When the light intensity deviation of a certain area is detected to exceed the standard, the control system analyzes the spatial distribution pattern of the deviation, identifies the corresponding laser and scanning parameters. The correction measures include: laser power adjustment (±20% range), scanning speed adjustment (±30% range), dwell time fine tuning (±0.5 seconds), phase relationship optimization (±5° range). For example, when the detection array shows that the average light intensity of the left area is 14.2 mW / cm 2 and the right area is 17.8 mW / cm 2 , the light intensity uniformity is 0.18, which exceeds the standard, the control system automatically increases the left laser power by 12% and reduces the right laser power by 8%. After 30 seconds of adjustment, the light intensity of the two areas is 16.1 mW / cm² and 16.4 mW / cm 2 , respectively, and the light intensity uniformity improves to 0.09. The temporal and spatial stability is maintained through continuous monitoring-correction cycles. The light intensity deviation is maintained within ±8% during the entire scanning process, ensuring the long-term stability of the internal source ultraviolet light field. This real-time monitoring and automatic correction technology solves the technical problem of difficult accurate control of light field distribution in large-area complex structure chips. Through closed-loop feedback control, high-precision temporal and spatial stability of the internal source ultraviolet light field is achieved, ensuring the uniformity and consistency of the light-shielded area polymerization reaction.

[0045] Please continue to see Figure 1 , an alternating magnetic field is applied to the continuous primary cured resin layer to excite the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent heat initiator to complete deep curing and simultaneously release internal residual stress, obtaining a fully cured resin layer. In one embodiment of the present application, the method for applying an alternating magnetic field to the continuous primary cured resin layer to excite the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent heat initiators to complete deep curing and simultaneously release internal residual stress to obtain a fully cured resin layer comprises: calculating a magnetic field parameter set based on the volume distribution data of the magnetic induction heating nanoparticles in the continuous primary cured resin layer and the thermal conductivity data of the resin, the magnetic field parameter set including the magnetic field frequency and the field strength, and setting an alternating magnetic field based on the magnetic field parameter set; applying a first-stage alternating magnetic field to the peripheral region of the continuous primary cured resin layer according to the magnetic field parameter set to make the peripheral magnetic induction heating nanoparticles generate induction heat and activate the remaining latent heat initiators in the peripheral region to obtain a peripheral deep curing zone; applying a second-stage alternating magnetic field to the central region of the continuous primary cured resin layer using the heat diffusion gradient generated by the peripheral deep curing zone, adjusting the magnetic field duty cycle to maintain the resin temperature in the central region at the latent heat initiator activation temperature zone to form a central deep curing zone; continuing to apply a third-stage alternating magnetic field to the peripheral deep curing zone and the central deep curing zone, gradually reducing the magnetic field amplitude and introducing phase modulation to promote the relaxation of the resin molecular chain segments and redistribute the internal residual stress; and stopping the alternating magnetic field when the heat flux density reaches a steady state and there is no further exothermic change to form a fully cured resin layer.

[0046] The following describes the steps involved in the above embodiment: The calculation of the magnetic field parameter set is based on the spatial distribution characteristics of the magnetic induction heating nanoparticles in the continuous primary cured resin layer and the thermal conductivity of the resin material. The volume distribution data is obtained by high-frequency induction current scanning, with a scanning frequency of 5 MHz. The concentration distribution of the magnetic induction heating nanoparticles is determined by measuring the induction response intensity at different positions. The scanning results show that the particle volume fraction in the peripheral region is 1.8-2.2%, and the particle volume fraction in the central region is 1.2-1.6%, showing a gradient distribution of high outside and low inside. The thermal conductivity data of the resin is determined by the transient plane heat source method, and the thermal conductivity of the continuous primary cured resin is 0.25-0.32 W / (m·K), and the thermal diffusivity is 1.2×10 -7 m 2 The magnetic field parameter set refers to a data set including key parameters such as magnetic field frequency, field strength, and waveform type, which is calculated by a special algorithm. In the calculation process, the optimal excitation frequency is determined according to the magnetic hysteresis loss characteristics and distribution density of the nanoparticles: the magnetic hysteresis loss peak of the FeCo alloy nanoparticles appears in the frequency range of 3.2-4.8 MHz, so the magnetic field frequency is set to 4.0 MHz. The magnetic field strength is calculated according to the desired heating power and particle distribution density: the peripheral region requires a heating power density of 150-200 W / cm 3 , corresponding to a magnetic field strength of 18-22 kA / m; the central region requires a heating power density of 100-130 W / cm 3The corresponding magnetic field strength is 12-16 kA / m. The alternating magnetic field setting uses a programmable magnetic field generator, which can generate a uniform alternating magnetic field with a frequency of 0.1-10 MHz and a magnetic field strength of 0-30 kA / m, and the magnetic field uniformity is better than ±3%. For example, for a GPU chip package of 28x28 mm, the magnetic field parameters of the peripheral 14 mm wide annular region are set to a frequency of 4.0 MHz and a strength of 20 kA / m, and the magnetic field parameters of the central 14x14 mm region are set to a frequency of 4.0 MHz and a strength of 14 kA / m. This quantitative parameter calculation based on material distribution and thermal performance ensures the spatial accuracy and temperature controllability of magnetic induction heating, avoiding the problems of local overheating or insufficient heating caused by empirical parameter setting.

[0047] The first-stage alternating magnetic field treatment is specifically aimed at precise heating control of the peripheral region of the continuous primary cured resin layer. The peripheral region refers to the annular region 0-7 mm away from the chip edge, which has a curing degree of 70-75% due to sufficient ultraviolet light irradiation in the previous photo-thermal co-curing process, but still contains 15-20% of unreacted latent thermal initiator. The magnetic field application uses an annular induction coil with an inner diameter of 30 mm, an outer diameter of 45 mm, and 36 turns, driven by a high-frequency power amplifier. The induction heat generation mechanism is based on eddy current loss and magnetic hysteresis loss of magnetic induction heating nanoparticles. When the alternating magnetic field acts on the FeCo alloy nanoparticles, eddy currents and magnetic domain flips are generated inside the particles, converting magnetic energy into heat energy. The temperature control of the heating process is monitored in real time by an infrared thermal imager, and the target temperature is set to 85-95°C, with a heating rate controlled at 8-12°C / min. Latent thermal initiator activation refers to the generation of free radicals by azobisisobutyronitrile after reaching the decomposition temperature, which initiates the polymerization reaction of the remaining epoxy groups and acrylate groups. The formation of the peripheral deep curing zone is monitored by impedance spectrum analysis, and when the dielectric constant at 1 kHz frequency rises from the initial 4.2 to 6.8, it indicates that the curing degree reaches 90-95%. For example, when processing a 28x28 mm GPU chip, the peripheral annular region is heated under a 20 kA / m, 4.0 MHz alternating magnetic field for 8 minutes, and the temperature rises from room temperature to 90°C. The latent thermal initiator is completely activated, and the curing degree is improved from 75% to 92%, forming a peripheral deep curing zone with high mechanical strength and good chemical stability. This regional precise magnetic field heating technology realizes the deep curing perfection of the peripheral region, while providing a stable heat source and structural support for the central region.

[0048] The second stage of alternating magnetic field treatment utilizes the heat conduction effect of the outer deep curing zone to synergistically heat the central region. The thermal diffusion gradient refers to the driving force for heat transfer from the outer high-temperature region to the inner low-temperature region. According to Fourier's law of heat transfer, the heat flux is proportional to the temperature gradient. After the first stage of treatment, the outer deep curing zone maintains a temperature of 85-90°C, while the initial temperature of the central region is approximately 35-40°C, resulting in a temperature difference of 50-55°C and a temperature gradient of approximately 4-6°C / mm. This thermal diffusion process causes the central region temperature to naturally rise to 60-65°C within 6-8 minutes, close to the lower activation temperature of the latent thermal initiator. The magnetic field duty cycle, defined as the ratio of the alternating magnetic field on-time to the total cycle time, is adjusted to control the average heating power in the central region. In practice, the magnetic field duty cycle in the central region is set to 30-40%, meaning that the magnetic field is on for 3-4 seconds and off for 6-7 seconds within each 10-second cycle, maintaining an appropriate average heating power level. The activation temperature zone of the latent thermal initiator refers to the temperature range in which azobisisobutyronitrile effectively decomposes, which is 68-88°C as determined by DSC. Temperature maintenance is achieved through a closed-loop control system. When the infrared thermal imager detects that the temperature in the central area is lower than 68°C, the magnetic field duty cycle is automatically increased to 45-50%; when the temperature exceeds 88°C, the duty cycle is reduced to 25-30% or the magnetic field output is suspended. The formation of the central deep curing zone is determined when the degree of curing reaches 85-90%, which is monitored by infrared spectroscopy at 1650cm -1 The disappearance of the double bond stretching vibration peak at the center of the chip was confirmed by the degree of disappearance. For example, in the central 14×14mm area of ​​a 28×28mm GPU chip, the temperature of the area was stably maintained in the range of 75-80°C for 12 minutes through magnetic field control with a 35% duty cycle. The degree of cure in the deep central curing zone increased from the initial 65% to 87%. This heating strategy, which combines thermal diffusion with magnetic fields, fully utilizes the heat from the peripheral area while avoiding the risk of overheating in the central area through precise duty cycle control.

[0049] The third stage alternating magnetic field treatment employs a decreasing amplitude and phase modulation technique to promote stress relaxation and rearrangement of resin molecular segments. The decreasing amplitude refers to a control process that gradually reduces the magnetic field strength from an initial value to a final value. The decreasing curve adopts an exponential decay mode, with the initial magnetic field strength being 12 kA / m and decaying to 2 kA / m according to an exponential law with a time constant of 180 seconds. Phase modulation refers to periodically changing the phase relationship of the magnetic field while maintaining the same frequency. The modulation frequency is set to 0.1-0.5 Hz, and the phase shift amplitude is ±30°. The relaxation of resin molecular segments refers to the transition process of polymer molecular chains from stress concentration state to equilibrium state under the action of temperature and weak vibration field. The gradual reduction of magnetic field amplitude reduces the heating power of the magnetic induction heated nanoparticles, and the resin temperature slowly decreases from 85°C to 50°C. The entire cooling process lasts for 25-30 minutes. The periodic magnetic field changes generated by phase modulation form a micro-vibration effect around the nanoparticles, with a vibration amplitude of about nanometers and a frequency equal to the modulation frequency. This micro-vibration promotes the micro-motion of molecular segments and accelerates the rearrangement and stress release of molecular chains. The redistribution of internal residual stress is achieved through the stress relaxation process. The shrinkage stress accumulated during the curing process is gradually released and redistributed to the lower energy configuration under the dual action of temperature and micro-vibration. The residual stress level is evaluated by measuring the lattice strain using X-ray diffraction. The residual stress before treatment is 15-20 MPa, and after treatment it is reduced to 3-5 MPa. For example, in the overall treatment process of a 28x28mm GPU chip, the third stage magnetic field treatment lasts for 30 minutes, the magnetic field strength decays from 12 kA / m to 2 kA / m, and a phase modulation of 0.2 Hz is applied. The final internal residual stress of the entire packaging structure is reduced by 75%, significantly improving the long-term reliability of the package.

[0050] The stop condition judgment ensures the complete termination of the reaction process through heat flux density monitoring. Heat flux density refers to the heat transfer rate per unit area, which is monitored in real time by a heat flow sensor array. The sensors are arranged at key positions of the packaging structure with a detection accuracy of ±0.1 W / m 2 ·min. The steady state judgment criterion is that the rate of change of heat flux is less than 0.05 W / (m 2 ·min) for 5 consecutive minutes, indicating that there is no significant chemical reaction exotherm inside the system. The exotherm change monitoring is confirmed by DSC scanning. When there is no obvious exotherm peak on the DSC curve, it is confirmed that all latent heat initiators have completely reacted. The quality standards of the completely cured resin layer include: a curing degree greater than 95%, confirmed by infrared spectroscopy and impedance spectroscopy; a glass transition temperature higher than 120°C, ensuring high temperature performance; and a shear strength greater than 35 MPa, meeting the mechanical reliability requirements. For example, in the treatment process of a 28x28mm GPU chip, when the heat flux density stabilizes at 0.8 W / m 2When the performance index of the fully cured resin layer does not change for 8 minutes, the system automatically stops the alternating magnetic field output, at which time the comprehensive performance index of the fully cured resin layer all reaches the design requirements. This stop control strategy based on the thermodynamic endpoint determination ensures the complete completion of the curing reaction and avoids the incomplete reaction or over-treatment problems caused by time control.

[0051] In one embodiment of the present application, the third stage alternating magnetic field continues to be applied to the peripheral deep curing zone and the central deep curing zone, gradually reduces the magnetic field amplitude and introduces phase modulation, which promotes the completion of the relaxation of the resin molecular chain segments and redistributes the internal residual stress, including: generating a differential magnetic field intensity distribution based on the curing degree difference between the peripheral deep curing zone and the central deep curing zone, applying a strong magnetic field mode to the area whose curing degree does not reach the preset curing standard, and applying a weak magnetic field mode to the area whose curing degree has reached the preset curing standard, the magnetic field strength of the weak magnetic field mode being less than that of the strong magnetic field mode; through the phase period modulation of the magnetic field, the magnetic induction heating nanoparticles produce pulsating heating, the periodic temperature fluctuation induces the cooperative thermal motion of the resin molecular chain segments, and the curing shrinkage stress is released; the magnetic field amplitude decay curve is set according to the thermal expansion matching requirement of the multi-layer heterogeneous interface, so that the thermal shrinkage process of the resin layer, the die and the substrate is kept synchronous, and the interface shear stress concentration is inhibited; when the magnetic field amplitude decreases to the lower limit value, it is switched to a constant low amplitude maintenance mode, the particles produce weak heating until the temperature change rate and the stress change rate are lower than the steady state determination standard, and the redistribution of the residual stress is completed.

[0052] The following is a specific description of the steps involved in the above embodiments: The generation of differential magnetic field intensity distribution is based on the real-time detection data of the curing degree of the peripheral deep curing zone and the central deep curing zone. The curing degree difference is obtained by simultaneous detection of a multi-point impedance spectrum analyzer, which is equipped with 36 detection probes arranged at 18 points in the peripheral area and 18 points in the central area, with a detection frequency of 1 kHz, and the curing degree is reflected by the change in dielectric constant. The preset curing standard is defined as a curing degree of 88%, corresponding to a dielectric constant value of 6.5-7.0. The differential magnetic field intensity distribution refers to the spatial variation of the magnetic field intensity pattern set according to the curing degree difference in different areas, which is realized by regional adjustment through a programmable magnetic field control system. The strong magnetic field mode is set to a magnetic field intensity of 16-20 kA / m, a frequency of 4.2 MHz, and a duty cycle of 60%, which is used for areas with a curing degree below 88%; the weak magnetic field mode is set to a magnetic field intensity of 8-12 kA / m, a frequency of 3.8 MHz, and a duty cycle of 30%, which is used for areas with a curing degree above 88%. The spatial adjustment of the magnetic field intensity distribution is realized by a ring-shaped segmented coil, which is divided into 8 segments in the peripheral coil and 4 segments in the central coil, with each segment independently controlling the magnetic field parameters. In the data processing process, the control system collects impedance data from each detection point every 30 seconds, calculates the corresponding curing degree, and then automatically adjusts the magnetic field intensity of each coil segment according to the preset threshold. For example, at a certain moment of a 28x28mm GPU chip processing, 12 of the 18 detection points in the peripheral area have a curing degree of 90-92%, and 6 have a curing degree of 85-87%; 8 of the 18 points in the central area have a curing degree of 89-91%, and 10 have a curing degree of 84-86%. The system automatically sets the coil segments corresponding to the 6 points with low curing degree in the peripheral area to the strong magnetic field mode (18 kA / m), and the remaining 12 points to the weak magnetic field mode (10 kA / m); the coil segments corresponding to the 10 points with low curing degree in the central area are set to the strong magnetic field mode (17 kA / m), and the remaining 8 points are set to the weak magnetic field mode (9 kA / m). This adaptive differential magnetic field distribution technology realizes precise heating control based on real-time curing state, avoids material degradation caused by excessive heating in areas with high curing degree, and ensures that areas with low curing degree obtain sufficient activation energy, ultimately achieving high consistency of curing degree in the entire packaging area.

[0053] The magnetic field phase periodic modulation technology produces pulsating heating effect by changing the phase relationship of the magnetic field, and induces the synergistic thermal motion of the resin molecular chain segment. The phase periodic modulation refers to periodically changing the phase offset of the magnetic field waveform while keeping the frequency and intensity of the magnetic field basically stable. The modulation period is set to 8-12 seconds, and the phase offset amplitude is ±45°. Pulsating heating refers to the periodic temperature change of the magnetic nanoparticles under the action of the phase modulation magnetic field. The temperature fluctuation amplitude is controlled within ±8-12°C. In specific implementation, a double-channel magnetic field generator is used to achieve the modulation effect by changing the phase difference between the two channels. When the two channels are in phase (phase difference 0°), the magnetic field strength reaches the maximum value, and the heating power of the nanoparticles is the highest. When the two channels are in anti-phase (phase difference 180°), the magnetic field strength decreases to the minimum value, and the heating power decreases accordingly. The periodic temperature fluctuation is monitored and verified in real time by an infrared thermal imager, and the temperature change presents a regular sinusoidal waveform. Synergistic thermal motion refers to the regular thermal vibration of polymer molecular chain segments under periodic temperature excitation. This vibration helps the molecular chain to transform from a high-energy configuration to a low-energy configuration. The release of curing shrinkage stress is realized through the stress relaxation mechanism. When the temperature rises, the activity of the molecular chain segment increases, and the internal stress is partially released. When the temperature decreases, the molecular chain segment rearranges to a relatively stable configuration. For example, under the condition of a phase modulation period of 10 seconds, the temperature of the resin around the magnetic nanoparticles periodically changes between 75-87°C. Through X-ray stress analysis, it is determined that the residual stress can be released by 3-5% in each modulation period. After 50 modulation periods, the overall residual stress release rate reaches 60-70%. This phase modulation technology creates a "thermal massage" effect, which promotes the micro-rearrangement of molecular chain segments through periodic temperature stimulation, and realizes efficient stress release without compromising material performance.

[0054] Thermal expansion matching control realizes the coordinated shrinkage of the multi-layer heterogeneous interface by setting a specific magnetic field amplitude decay curve. The multi-layer heterogeneous interface refers to the bonding interface between different material layers in the GPU packaging structure, mainly including the resin layer-die interface, resin layer-substrate interface and die-substrate interface. The thermal expansion matching requirement refers to the size change of each material layer during temperature change should be coordinated to avoid interface stress concentration caused by the difference in expansion coefficient. The linear expansion coefficient of the GPU die is 2.6x10 -6 / °C, the linear expansion coefficient of the substrate is 13-17x10 -6 / °C, and the linear expansion coefficient of the cured resin is 45-55x10 -6°C. The magnetic field amplitude decay curve is set by a multi-stage control strategy: the first stage decays from 20 kA / m to 15 kA / m for 8 minutes, corresponding to a temperature drop from 90 °C to 75 °C; the second stage decays from 15 kA / m to 8 kA / m for 12 minutes, corresponding to a temperature drop from 75 °C to 55 °C; the third stage decays from 8 kA / m to 3 kA / m for 15 minutes, corresponding to a temperature drop from 55 °C to 35 °C. The decay rate is set based on the shrinkage characteristics of each material: the fast cooling stage mainly releases the shrinkage stress of the resin, the moderate cooling stage coordinates the shrinkage difference between the die and the substrate, and the slow cooling stage realizes the final stress balance. The interface shear stress suppression is verified by stress monitoring sensors, which are installed at key interface positions to monitor the shear stress changes in real time. For example, in the processing of a 28x28 mm GPU chip, after the decay curve is controlled according to the set curve, the maximum shear stress of the resin layer-die interface is reduced from 25 MPa to 8 MPa, and the shear stress of the resin layer-substrate interface is reduced from 30 MPa to 12 MPa, effectively suppressing the interface stress concentration phenomenon. This precise temperature control technology based on material thermal expansion matching solves the key reliability problem in multi-material packaging structure, and avoids interface cracking and delamination failure by coordinating the shrinkage process of each layer of material.

[0055] The constant low amplitude maintaining mode realizes the final balance of the residual stress through the precise temperature and stress double parameter control. The lower limit value is set to the magnetic field strength of 3 kA / m, and when the magnetic field amplitude decays to this value, the control system automatically switches to the constant mode. The constant low amplitude maintaining mode refers to maintaining the magnetic field strength in the range of 3±0.2 kA / m stable output, and the frequency is maintained at 3.5 MHz, to produce weak and continuous heating effect. The role of weak heating is to maintain the local temperature around the magnetic induction heating nanoparticles slightly higher than the ambient temperature 5-8°C, to provide enough thermal energy for the molecular chain segment to make micro-adjustment. The temperature change rate is monitored by a distributed thermocouple array, which contains 16 measuring points, and the sampling frequency is 0.1 Hz. The steady state judgment criterion is defined as the temperature change rate less than 0.2°C / min within 10 minutes, which indicates that the thermal equilibrium has been reached. The stress change rate is monitored by a strain gauge sensor, which is pasted on the key stress position of the packaging structure, and the detection accuracy is ±0.1με. The steady state judgment criterion of the stress change rate is that the stress change rate is less than 0.5 MPa / min within 15 minutes. The judgment of the completion of the redistribution of the residual stress needs to meet the two steady state conditions of temperature and stress at the same time. For example, in the final processing stage of the 28×28 mm GPU chip, the constant low amplitude maintaining mode lasted for 45 minutes, during which the average temperature was stable at 32±1°C, and the temperature change rate was kept at 0.08°C / min within the last 15 minutes; the stress monitoring showed that the stress value of the main stress concentration point changed at a rate of 0.3 MPa / min within the last 20 minutes, meeting the steady state judgment condition. After the processing is completed, the residual stress distribution of the whole packaging structure is highly uniformized, and the maximum stress value is reduced from the initial 28 MPa to 4 MPa, and the stress distribution standard deviation is reduced from 8.5 MPa to 1.2 MPa. This double parameter steady state control technology ensures the sufficiency and stability of the stress release process, avoids the long-term reliability hidden danger caused by insufficient processing, and provides reliable guarantee for the long-term stable operation of the GPU package in the harsh working environment.

[0056] Please continue to refer to Figure 1 Gradient cooling from the substrate side to the die side and isothermal aging are performed on the fully cured resin layer, so that the curing shrinkage stress is redistributed and stabilized, and a GPU chip packaging structure is obtained.

[0057] In one embodiment of the present application, the gradient cooling from the substrate side to the die side is performed on the fully cured resin layer and the isothermal aging is carried out, so that the curing shrinkage stress is redistributed and stabilized, and a GPU chip packaging structure is obtained, including: establishing a temperature gradient curve according to the thickness of the fully cured resin layer and the difference in the linear expansion coefficient of the die and the substrate, setting the starting temperature of the substrate side and the target temperature of the die side, and generating a gradient cooling parameter set; implementing segmented cooling from the substrate side to the die side according to the gradient cooling parameter set, maintaining a constant temperature platform at the end of each temperature segment by using the residual heat of the resin curing exothermic, and forming a multi-stage gradient cooling process; maintaining the constant temperature section after the gradient cooling is completed until the heat flux density inside the resin is lower than the preset threshold, and applying a micro-vibration mechanical wave to the packaging structure in the isothermal stage to promote chain segment relaxation and stress redistribution; confirming that the temperature gradient in the resin layer is less than the temperature uniformity standard and the stress change rate tends to zero at the end of the isothermal stage, completing the isothermal aging, and forming a GPU chip packaging structure.

[0058] The following is a specific description of the steps involved in the above-mentioned embodiment: The establishment of the temperature gradient curve is based on the thickness measurement of the fully cured resin layer and the difference analysis of the linear expansion coefficients of each component material. The thickness of the fully cured resin layer is accurately measured by an ultrasonic thickness gauge, with a measurement accuracy of ±2 μm, and a typical thickness of 65-75 μm. The difference in the linear expansion coefficient refers to the difference in the relative expansion amount per unit length of different materials when the temperature changes. The linear expansion coefficient of the GPU die (silicon material) is 2.6×10 -6 / °C, the linear expansion coefficient of the substrate (organic material) is 15-17×10 -6 / °C, and the linear expansion coefficient of the fully cured resin layer is 48-52×10 -6°C. The temperature gradient curve refers to the temperature distribution function from the substrate side to the die side, and the optimal gradient form is determined by numerical calculation. In the data processing, the packaging structure is divided into 10 equally spaced layers in the thickness direction, and the thermal stress and deformation of each layer are calculated to optimize the target temperature distribution of each layer. The starting temperature of the substrate side is set to 85 °C, making full use of the residual heat in the magnetic field processing stage, and the target temperature of the die side is set to 45 °C, forming a total temperature difference of 40 °C and a temperature gradient of about 6 °C / mm. The gradient cooling parameter set includes parameters such as the target temperature, cooling rate, holding time, etc. of each temperature control point, and a parameter matrix is generated through a special algorithm. The specific parameters are set as follows: the target temperature of the first layer of the substrate side is 85 °C, and the cooling rate is 3 °C / min; the target temperature of the third layer is 70 °C, and the cooling rate is 4 °C / min; the target temperature of the fifth layer is 60 °C, and the cooling rate is 5 °C / min; the target temperature of the eighth layer is 50 °C, and the cooling rate is 4 °C / min; the target temperature of the tenth layer of the die side is 45 °C, and the cooling rate is 3 °C / min. For example, in the processing of a 28x28 mm GPU chip package, through the established temperature gradient curve, the shrinkage stress at the substrate interface is preferentially released, avoiding the generation of excessive tensile stress at the die interface, and the calculation results show that the peak value of the interfacial shear stress is reduced from 35 MPa in uniform cooling to 18 MPa in gradient cooling. This gradient design based on material thermal expansion matching realizes the ordered release of multi-material interface stress, and through precise temperature distribution control, the interfacial cracking and delamination failure caused by thermal shrinkage mismatch are avoided.

[0059] The segmented cooling process implements precise space-time temperature control according to the gradient cooling parameter set, and fully utilizes the residual exothermic characteristics of resin curing. The segmented cooling refers to dividing the entire cooling process into multiple independent temperature segments, each segment having a specific start and end temperature and cooling rate. The temperature control uses a zoned heating / cooling system, with a resistance heater and forced air cooling device on the substrate side, and a semiconductor refrigerator on the die side, with each zone temperature adjusted by an independent PID controller. The first temperature segment decreases from 85°C to 70°C, with a cooling rate of 3°C / min on the substrate side and 2°C / min on the die side, and a duration of 8 minutes; the second temperature segment decreases from 70°C to 55°C, with a cooling rate of 4°C / min on the substrate side and 3°C / min on the die side, and a duration of 6 minutes; the third temperature segment decreases from 55°C to 45°C, with a cooling rate of 3°C / min on the substrate side and 2.5°C / min on the die side, and a duration of 5 minutes. The residual exothermicity of resin curing refers to the weak exothermicity generated by the continuous reaction of part of the latent heat initiator during the early stage of the magnetic field deep curing process, and the residual exothermicity is determined by DSC analysis to be 8-12 J / g. The isothermal plateau maintenance refers to pausing active cooling at the end of each temperature segment and using residual exothermicity to maintain the current temperature for 2-3 minutes, so that the temperature distribution is further homogenized. The multi-stage gradient cooling process is monitored in real time by 16 temperature sensors distributed at different depths, ensuring that the actual temperature distribution deviates from the set curve by less than ±2°C. For example, at the end of the second temperature segment, the substrate side temperature is 55.2°C and the die side temperature is 57.8°C, at which point active cooling is stopped and 0.8 J / g of residual exothermicity is used to maintain an isothermal plateau for 3 minutes, with both sides eventually equalizing to 56.5°C. This segmented gradient cooling technology achieves step-by-step release of internal stress in the packaging structure and high homogenization of the temperature field through precise space-time temperature control and utilization of residual heat, avoiding thermal shock and stress concentration that can be caused by rapid cooling.

[0060] The isothermal aging treatment achieves the final relaxation of molecular chain segments and stress balance through the synergistic effect of heat flux monitoring and mechanical vibration. The isothermal section refers to the stage of maintaining the temperature stability of the packaging structure after gradient cooling is completed, with a target temperature set to 40±1°C, maintained by a precision temperature control box. Heat flux monitoring uses a heat flux sensor array, with 12 sensors evenly distributed on the packaging surface, with a detection accuracy of ±0.05 W / m 2 , which monitors the heat changes inside the packaging in real time. The pre-set threshold is set to a heat flux change rate of less than 0.1 W / (m 2• min), indicating that the internal chemical reactions and physical relaxation processes have essentially stopped. The micro-vibration mechanical waves are generated by piezoelectric ceramic actuators, with a frequency setting of 25-35 kHz and an amplitude control of 3-5 pm, and the actuators are installed at the four corners of the package structure. Segment relaxation refers to the process of polymer molecular chain transformation from a high-stress state to a low-energy configuration under the dual action of temperature and micro-vibration. Stress redistribution is monitored by a strain gauge sensor array, with 20 strain gauges attached to key stress locations to record stress change trends in real time. The mechanism of micro-vibration is to promote the microscopic movement of molecular segments through periodic mechanical excitation, accelerating the stress relaxation process without damaging the polymer network structure that has been formed. The duration of isothermal aging is determined by both the heat flux and the stress change rate, and the process is terminated when both parameters meet the steady-state condition. For example, in the isothermal aging process of a 28x28 mm GPU chip, the initial stage heat flux is 2.8 W / m 2 , and the stress change rate is 1.2 MPa / min; after 6 hours of isothermal treatment, the heat flux decreases to 0.6 W / m 2 , and the stress change rate decreases to 0.05 MPa / min, meeting the preset threshold requirements. Micro-vibration treatment further releases residual stress by 15-20%, and the final residual stress level stabilizes at 2-3 MPa. This isothermal aging technology achieves deep release of internal stress and significant improvement in long-term stability of the package structure through the synergistic effect of thermodynamic equilibrium and mechanical relaxation.

[0061] The final quality confirmation verifies the integrity of the package structure by dual standards of temperature uniformity and stress stability. The temperature uniformity standard is defined as the temperature difference between any two points in the resin layer being less than 3°C, which is verified by a distributed thermocouple array measurement. The thermocouple array contains 25 measurement points, which are distributed in a 5x5 grid in the package area, with a measurement accuracy of ±0.1°C and a data acquisition frequency of 1 Hz. The temperature gradient is calculated by dividing the temperature difference between adjacent measurement points by the distance. When the temperature gradient of all measurement point pairs is less than 1°C / mm, it is determined that the temperature distribution meets the uniformity standard. Stress change rate monitoring is achieved by a high-precision strain measurement system, which contains 32 resistance strain gauges to monitor the changes in tensile, compressive and shear stresses. The criterion for the stress change rate tending to zero is that the stress change rate of all measurement points is less than 0.02 MPa / min within 30 consecutive minutes. The comprehensive criterion for the completion of the constant temperature aging requires that both the temperature uniformity and the stress stability are met. The final formation of the GPU chip package structure marks the completion of the entire multi-physical field collaborative curing process, and the formed package structure has a high degree of uniform internal performance and excellent long-term reliability. For example, in the 28x28mm GPU chip package structure that completes the constant temperature aging, the maximum temperature difference of the 25 temperature measurement points is 2.1°C, and the average temperature gradient is 0.6°C / mm; the average stress change rate of the 32 stress measurement points within the last 60 minutes is 0.008 MPa / min, and the maximum value is 0.015 MPa / min, both of which meet the set standards. The finally formed GPU chip package structure passes the 1000 times -65°C to +150°C thermal cycle test, with no detectable interface debonding or crack defects, and the shear strength remains above 95% of the initial value. This dual-standard-based quality confirmation technology ensures that the package structure meets the expected performance indicators, providing reliable protection for the long-term stable operation of the GPU chip in harsh working environments, and verifying the effectiveness and advancement of the entire black workshop multi-physical field collaborative packaging process.

[0062] The above describes the GPU chip packaging method in the embodiments of the present application, and the following describes the GPU chip forming equipment in the embodiments of the present application. Please refer to Figure 2The GPU chip forming device according to an embodiment of the present application comprises: a functional resin injection module 101, configured to inject a bottom filling resin containing ultraviolet initiators, latent thermal initiators, up-conversion luminescent nanoparticles and magnetic induction heating nanoparticles into a gap between a die and a substrate, to form a fully filled functional resin layer; a circumferential ultraviolet curing module 102, configured to perform ultraviolet irradiation on a peripheral area of the functional resin layer, so that the ultraviolet initiators are polymerized and heat is released, and the latent thermal initiators are activated, to generate a thermal polymerization reaction front that advances from the outside to the inside of the resin layer, and to obtain a peripheral cured resin area; a near-infrared transmission excitation module 103, configured to use near-infrared light to transmit through the die and excite the up-conversion luminescent nanoparticles to release ultraviolet light in a light-shielded area, and to cooperatively induce resin polymerization in the light-shielded area with the thermal polymerization reaction front, to obtain a continuous primary cured resin layer; an alternating magnetic field deep curing module 104, configured to apply an alternating magnetic field to the continuous primary cured resin layer, to excite the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent thermal initiators to complete deep curing, and to synchronously release internal residual stress, to obtain a fully cured resin layer; and a gradient cooling aging module 105, configured to perform gradient cooling from the substrate side to the die side on the fully cured resin layer and to perform isothermal aging, to redistribute and stabilize the curing shrinkage stress, to obtain a GPU chip packaging structure.

[0063] The above description is only preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation, direct / indirect application in other related technical fields, or the like, within the inventive concept of the present application, using the content of the present application specification and drawings, is included in the patent protection scope of the present application.

Claims

1. A GPU chip packaging method, characterized in that: include: Injecting a bottom filling resin containing ultraviolet light initiator, latent thermal initiator, upconversion luminescent nanoparticles, and magnetic induction heating nanoparticles into the gap between the bare chip and the substrate to form a completely filled functional resin layer; Applying ultraviolet light to the peripheral area of ​​the functionalized resin layer to cause the ultraviolet light initiator to polymerize and release heat and activate the latent thermal initiator, thereby generating a thermal polymerization reaction front that advances from the outside to the inside along the resin layer, thereby obtaining a peripheral cured resin area; Utilizing near-infrared light transmitted through the bare chip to excite the upconversion luminescent nanoparticles to release ultraviolet light in the light-shielded area, and cooperating with the thermal polymerization reaction front to initiate polymerization of the resin in the light-shielded area to obtain a continuous primary cured resin layer; Applying an alternating magnetic field to the continuous initially cured resin layer to stimulate the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent thermal initiator to complete deep curing and simultaneously release the internal residual stress to obtain a completely cured resin layer; The fully cured resin layer is subjected to gradient cooling from the substrate side to the die side and is subjected to constant temperature aging to redistribute and stabilize the curing shrinkage stress, thereby obtaining a GPU chip packaging structure.

2. The GPU chip packaging method according to claim 1, wherein: The method of injecting a bottom filling resin containing ultraviolet light initiator, latent thermal initiator, upconversion luminescent nanoparticles, and magnetic induction heating nanoparticles into the gap between the bare chip and the substrate to form a completely filled functional resin layer includes: Calculating a shading coefficient based on the die area and the HBM stack height, setting a molar ratio of the ultraviolet light initiator to the latent thermal initiator based on the shading coefficient, preparing an epoxy resin matrix and obtaining a resin premix; Performing hydroxylation surface modification on the upconversion luminescent nanoparticles to form hydroxyl sites on the particle surfaces that can condense with epoxy groups, adding the modified particles to the resin premix and dispersing them under constant shear conditions to obtain an upconversion dispersion; adding magnetic induction heating nanoparticles to the upconversion dispersion, performing shear homogenization and vacuum exhaust combined treatment to remove suspended bubbles and uniformly distribute the two types of functional particles in the system, thereby obtaining a functionalized bottom filling resin; Maintaining a negative pressure in the gap between the die and the substrate in a constant temperature environment, applying a local infrared radiation heat field at the outer edge of the gap to establish a temperature gradient and a viscosity gradient that decreases from the outer edge to the center, and injecting the functionalized underfill resin into the outer edge and center of the gap in two stages according to the viscosity gradient to form a primary filling resin layer; Maintain negative pressure and replenish the functionalized bottom filling resin in a step-by-step manner, while applying micro-vibration mechanical waves on both sides of the gap to migrate residual bubbles to the outer edge and be discharged, thereby forming a completely filled functionalized resin layer.

3. The GPU chip packaging method according to claim 2, wherein: The upconversion luminescent nanoparticles are subjected to hydroxylation surface modification to form hydroxyl sites on the particle surfaces that can condense with epoxy groups, and the modified particles are added to the resin premix and dispersed under constant shear conditions to obtain an upconversion dispersion, comprising: By treating the nanoparticles with silane coupling agents, the coordination environment of rare earth ions on their surfaces is regulated, and a directional hydroxyl functional layer is constructed on the particle surface, so that the hydroxyl density reaches a preset standard that meets the conversion efficiency requirements of near-infrared light to ultraviolet light in a completely dark environment. The concentration gradient of the upconversion luminescent nanoparticles in the resin is determined based on the spatial distribution characteristics of the GPU chip's light-shielding area. By controlling the spatial non-uniformity of the shear field, the modified particles are able to form a concentration distribution in the resin premix that matches the light-shielding intensity. During the constant shear dispersion process, an ultrasonic field is applied synchronously to break up particle agglomerates and activate the hydroxylated surface through cavitation, forming pre-crosslinking bonds with epoxy groups and enhancing the particle-resin interface bonding strength. By adjusting the corresponding relationship between shear rate and ultrasonic power to control the pre-crosslinking density of the particle surface, the upconversion luminescent nanoparticles can maintain stable light conversion performance during the subsequent excitation process to obtain an upconversion dispersion.

4. The GPU chip packaging method according to claim 1, wherein: The method of irradiating the peripheral area of ​​the functionalized resin layer with ultraviolet light to cause the ultraviolet initiator to polymerize and release heat and activate the latent thermal initiator, thereby generating a thermal polymerization reaction front that advances from the outside to the inside along the resin layer to obtain a peripheral cured resin area, comprises: Performing circumferential partition modulation on the ultraviolet light according to the outer edge profile of the die and the thickness of the functional resin layer to obtain a circumferentially uniform circumferential ultraviolet light field; Implementing pulse timing control on the annular ultraviolet light field, setting the pulse width and interval, so that the polymerization exothermic peak temperature is in the latent thermal initiator activation temperature zone, and obtaining a stable exothermic pulse sequence; The intensity of the ultraviolet light is adjusted in a radial gradient according to the stable exothermic pulse sequence to obtain a balanced energy distribution along the thickness direction of the resin layer, thereby forming a thermal polymerization reaction front advancing from the outside to the inside; The thermal polymerization reaction front is tracked in real time. When the front reaches a set distance from the light shielding boundary of the bare chip, the illumination parameters are kept constant until heat release is completed, forming a peripheral solidified resin area.

5. The GPU chip packaging method according to claim 1, wherein: The method utilizes near-infrared light transmitted through the bare chip to excite the up-conversion luminescent nanoparticles to release ultraviolet light in the light-shielded area, and cooperates with the thermal polymerization reaction front to initiate polymerization of the resin in the light-shielded area to obtain a continuous primary cured resin layer, including: Based on the GPU die layout structure data, optical simulation is performed on the silicon-based wiring layer, through-hole and metal interconnect structure to obtain the coordinate corresponding wavelength transmittance matrix and obtain the near-infrared transmission spectrum parameter matrix; constructing a multi-wavelength irradiation parameter set according to the near-infrared transmission spectrum parameter matrix, and aligning the multi-wavelength irradiation parameter set with the spatiotemporal coordinates of the thermal polymerization reaction front to obtain a scanning control matrix; According to the scanning control matrix, near-infrared light beams of different wavelengths and powers are sequentially projected onto the front surface of the die to stimulate the upconversion luminescent nanoparticles at the corresponding coordinates to release ultraviolet light in the light-shielded area, thereby forming a spatially uniform endogenous ultraviolet light field; The endogenous ultraviolet light field is kept overlapping synchronously with the thermal polymerization reaction front until the resin conversion rate in the light-shielded area reaches a preset conversion rate standard, thereby forming a continuous primary cured resin layer.

6. The GPU chip packaging method according to claim 5, wherein: The method of sequentially projecting near-infrared light beams of different wavelengths and powers onto the front surface of the die according to the scanning control matrix to excite up-conversion luminescent nanoparticles at corresponding coordinates to release ultraviolet light in the light-shielded area, thereby forming a spatially uniform endogenous ultraviolet light field, includes: A dynamic focusing processing strategy for the near-infrared beam is established based on the scanning control matrix, so that the beam focus is positioned at key structural layers at different depths within the die to compensate for the difference in attenuation of the near-infrared light by each metal layer; Adjust the beam dwell time and repetition frequency according to the distribution density of the heat generating units inside the chip to suppress local temperature rise and ensure that the particles receive sufficient excitation energy; Multi-beam parallel scanning is used to synchronously irradiate the shading area in different zones, and the intensity of the internal source UV light in each zone is balanced through phase modulation between beams to eliminate the phenomenon of asynchronous curing. The source UV light field distribution in each scanning area is collected in real time. When the light intensity deviation exceeds the light intensity uniformity standard, the corresponding beam parameters are automatically corrected to maintain the spatiotemporal stability of the source UV light field in the shading area.

7. The GPU chip packaging method according to claim 1, wherein: The method of applying an alternating magnetic field to the continuous initially cured resin layer to stimulate the magnetic induction heating nanoparticles to generate induction heat to trigger the remaining latent thermal initiator to complete deep curing and simultaneously release the internal residual stress to obtain a completely cured resin layer includes: Calculating a magnetic field parameter set based on volume distribution data of magnetic induction heating nanoparticles in a continuous primary cured resin layer and resin thermal conductivity data, wherein the magnetic field parameter set includes magnetic field frequency and field intensity, and setting an alternating magnetic field based on the magnetic field parameter set; applying a first-stage alternating magnetic field to the peripheral area of ​​the continuous primary cured resin layer according to the magnetic field parameter set, so that the peripheral magnetic induction heating nanoparticles generate induction heat and activate the remaining latent thermal initiator in the peripheral area to obtain a peripheral deep cured area; Applying a second-stage alternating magnetic field to the central region of the continuous primary cured resin layer using the thermal diffusion gradient generated by the peripheral deep curing zone, adjusting the magnetic field duty cycle to maintain the resin temperature in the central region within the latent thermal initiator activation temperature range, thereby forming a central deep curing zone; The third stage alternating magnetic field is continuously applied to the peripheral deep curing zone and the central deep curing zone, the magnetic field amplitude is gradually reduced and phase modulation is introduced to promote the relaxation of the resin molecular segments and redistribute the internal residual stress; When the heat flux density reaches a steady state and there is no further heat release change, the alternating magnetic field is stopped to form a completely cured resin layer.

8. The GPU chip packaging method according to claim 7, wherein: The third stage of applying the alternating magnetic field to the peripheral deep curing zone and the central deep curing zone, gradually reducing the magnetic field amplitude and introducing phase modulation, so as to promote the relaxation of the resin molecular segments and redistribute the internal residual stress, includes: generating a differential magnetic field intensity distribution based on the difference in curing degree between the peripheral deep curing zone and the central deep curing zone, applying a strong magnetic field mode to the area where the curing degree does not reach the preset curing standard, and applying a weak magnetic field mode to the area where the curing degree reaches the preset curing standard, wherein the magnetic field intensity of the weak magnetic field mode is less than the magnetic field intensity of the strong magnetic field mode; By periodically modulating the magnetic field phase, the magnetic induction heating nanoparticles generate pulsating heat, and the periodic temperature fluctuations are used to induce the coordinated thermal motion of the resin molecular segments to release the curing shrinkage stress. The magnetic field amplitude attenuation curve is set according to the thermal expansion matching requirements of the multi-layer heterogeneous interface to synchronize the thermal contraction processes of the resin layer, bare chip and substrate, thereby suppressing the concentration of shear stress on the interface. When the magnetic field amplitude drops to the lower limit, it switches to a constant low-amplitude maintenance mode, maintaining a slight heating of the particles until both the temperature change rate and the stress change rate are lower than the steady-state judgment standard, completing the residual stress redistribution.

9. The GPU chip packaging method according to claim 1, wherein: The fully cured resin layer is subjected to gradient cooling from the substrate side to the die side and constant temperature aging to redistribute and stabilize the curing shrinkage stress, thereby obtaining a GPU chip packaging structure, including: A temperature gradient curve is established based on the thickness of the fully cured resin layer and the difference in linear expansion coefficient between the die and substrate, and the starting temperature on the substrate side and the target temperature on the die side are set to generate a gradient cooling parameter set. Implementing segmented cooling from the substrate side to the die side according to the gradient cooling parameter set, and maintaining a constant temperature platform at the end of each temperature segment by utilizing the residual heat released by resin curing, thereby forming a multi-stage gradient cooling process; After the gradient cooling is completed, the constant temperature section is maintained until the heat flux density inside the resin is lower than the preset threshold, and micro-vibration mechanical waves are applied to the packaging structure during the constant temperature stage to promote chain segment relaxation and stress redistribution; At the end of the constant temperature stage, it is confirmed that the temperature gradient in the resin layer is less than the temperature uniformity standard and the stress change rate tends to zero, and the constant temperature aging is completed to form the GPU chip packaging structure.

10. A GPU chip forming device, characterized in that: The GPU chip molding device adopts the GPU chip packaging method according to any one of claims 1 to 9, and the GPU chip molding device includes: Functionalized resin injection module, used to inject bottom filling resin containing ultraviolet light initiator, latent thermal initiator, upconversion luminescent nanoparticles, and magnetic induction heating nanoparticles into the gap between the bare chip and the substrate to form a completely filled functionalized resin layer; a circumferential UV curing module for irradiating the peripheral area of ​​the functionalized resin layer with UV light, causing the UV initiator to polymerize and release heat and activate the latent thermal initiator, thereby generating a thermal polymerization reaction front that advances from the outside to the inside along the resin layer, thereby obtaining a peripheral cured resin area; A near-infrared transmission excitation module is used to utilize near-infrared light to transmit through the bare chip to excite the upconversion luminescent nanoparticles to release ultraviolet light in the light-shielded area, and cooperate with the thermal polymerization reaction front to initiate polymerization of the resin in the light-shielded area to obtain a continuous primary cured resin layer; An alternating magnetic field deep curing module is used to apply an alternating magnetic field to the continuous primary cured resin layer, stimulating the magnetic induction heating nanoparticles to generate induced heat to trigger the remaining latent thermal initiator to complete deep curing and simultaneously release internal residual stress to obtain a completely cured resin layer; The gradient cooling and aging module is used to perform gradient cooling from the substrate side to the die side on the fully cured resin layer and perform constant temperature aging to redistribute and stabilize the curing shrinkage stress to obtain a GPU chip packaging structure.

Images