Method and system for designing high glass transition temperature chip underfill based on machine learning and application

Abstract

The invention belongs to the technical field of machine learning and chip manufacturing, and discloses a method and system for designing high glass transition temperature chip underfill glue based on machine learning and application, the method takes the high glass transition temperature Tg of an epoxy resin-curing agent system as a main target parameter, and comprises the following steps: S1, constructing polymer molecular structure parameters and process parameters; s2, screening a key feature set; s3, predicting based on machine learning; and S4, obtaining an overall optimization scheme. Experimental verification shows that the actual measurement Tg of the obtained optimal formula (such as a trifunctional epoxy / bisphenol A epoxy / anhydride naphthalene compound system) reaches 205-215 DEG C and is increased by 20% or above compared with a traditional formula, and the model prediction error is smaller than + / -6 DEG C. The bottleneck that the design of the epoxy resin matrix Tg in the underfill depends on a trial-and-error method is solved, the research and development period can be shortened by 60% or above, and the method is suitable for high-reliability packaging scenes such as 5G chips and power devices and is wide in application.

Description

Technical Field

[0001] The present invention relates to the technical fields of machine learning and chip manufacturing, and particularly relates to a method, system and application for designing a chip underfill with a high glass transition temperature based on machine learning. Background Art

[0002] In the field of high-density packaging of integrated circuits, as a key interface material, the chip underfill needs to maintain dimensional stability and mechanical strength during high-temperature reflow soldering (peak temperature 260 - 280 °C) and long-term service. Its glass transition temperature (Tg) directly determines the thermal fatigue resistance of the packaging structure. Epoxy resin has become the mainstream matrix material for underfill due to its excellent adhesion, chemical resistance, and adjustable curing network. In traditional technologies, on the one hand, to increase the Tg of epoxy resin, the molecular structure of monomers is designed. For example, introducing multi-functional epoxy monomers (such as tetra-functional TGDDM, hexa-functional HP-7200) to increase the crosslinking density, or using monomers with a rigid aromatic ring structure (such as biphenyl-type epoxy, naphthalene-ring modified epoxy) to restrict the movement of molecular chains. For example, the Tg of biphenyl epoxy (BPA type) can be increased by about 30 - 50 °C compared to standard bisphenol A epoxy (DGEBA). However, high-functional monomers tend to cause a sharp increase in the viscosity of the system, affecting the dispersion of fillers and the flowability of filling. On the other hand, by selecting a suitable curing agent or a modified curing agent. For example, using highly reactive aromatic amines (such as 4,4'-diaminodiphenyl sulfone DDS) or rigid anhydrides (such as methyl nadic anhydride MNA) as curing agents to increase the Tg by enhancing the interaction between molecular chains. However, the matching degree between the molecular weight distribution, steric hindrance effect (such as ortho-substituents) of the curing agent and the epoxy monomer needs to be verified by repeated experiments, and high-rigid curing agents often result in an increase in curing stress, leading to the risk of interface debonding. Finally, a multi-stage temperature rise program (such as preheating at 120 °C - main curing at 170 °C - post-curing at 200 °C) is used to promote the homogenization of the crosslinking network and reduce the residue of unreacted groups. However, there is a strong coupling effect between the process parameters (temperature gradient, holding time, pressure) and the material composition (monomer / curing agent ratio, catalyst content), and it is difficult to analyze their synergistic mechanism by traditional trial-and-error methods.

[0003] For example, the high-efficiency high-temperature heat-conducting bottom filling adhesive and its preparation method disclosed in CN201711069729.4 is a one-component epoxy resin adhesive, and its glass transition temperature is often lower than 130 °C, and there is a large fluctuation among various embodiments, which is unstable. Therefore, it is also difficult to balance the fluidity, viscosity and high-temperature resistance of the material.

[0004] Therefore, there are many bottlenecks in the existing technology. For example, there is a lack of an analytical model for the quantitative influence of molecular parameters such as the number of functional groups and aromatic ring density on Tg. The design relies on empirical rules (such as "the proportion of rigid groups > 40%"), making it difficult to accurately guide molecular modification. The R & D efficiency is low. The formulation screening requires hundreds of experiments, and the development cycle is as long as 6 - 12 months. Moreover, it is difficult to balance Tg and processability (such as viscosity, curing rate). In addition, the calculation cost is extremely high. Although molecular dynamics (MD) or first-principles simulation can predict the formation process of the crosslinked network, the computational resources consumed for complex multi-component systems are extremely large, and high-throughput design cannot be achieved. Summary of the Invention

[0005] The object of the present invention is to provide a method, system and application for designing a chip underfill with a high glass transition temperature based on machine learning, aiming at the above deficiencies in the existing technology. By organically combining machine learning with the components, ratios, processes, etc. in the specific technical field of chip manufacturing underfill, a machine learning algorithm is used to analyze the influence law of the molecular structure characteristics of epoxy resin monomers and curing agents on the glass transition temperature (Tg), and it is applied to the design of chip underfills with a high Tg temperature, obtaining an efficient design method and system that integrates multi-scale features and has both prediction accuracy and physical interpretation ability, breaking through the technical difficulties in optimizing the Tg of the epoxy resin matrix, realizing the efficient optimization of complex formulations and processes under multiple constraints, and being easy to achieve high-throughput design to solve the above technical problems.

[0006] To achieve the above object, the technical solution provided by the present invention is as follows: A method for designing a chip underfill with a high glass transition temperature based on machine learning, which takes the high glass transition temperature Tg of the epoxy resin-curing agent system as the main target parameter, and includes the following steps: S1: Construct polymer molecular structure parameters and process parameters Using a publicly available polymer database, obtain the molecular structure parameters of epoxy resin monomers and curing agents and the measured Tg values corresponding thereto. The molecular structure parameters include the number of functional groups (nEpoxy), aromatic ring density (Ar-density), molecular weight (MW) of the epoxy monomer, as well as the amine hydrogen equivalent (AHE) and steric hindrance parameter (Steric-index) of the curing agent, to obtain the corresponding molecular parameter descriptor set and process parameter descriptor set; S2: Screen the key feature set Adopt the recursive feature elimination (RFE) combined with the XGBoost algorithm to screen the key feature set from the molecular parameter descriptor set and the process parameter descriptor set, where the process parameters include the curing step temperature and the curing time; S3: Prediction based on machine learning Based on the filtered feature set, a Gaussian Process Regression (GPR) model is constructed to predict the Tg of the epoxy resin-curing agent system, obtaining candidate solutions for components, ratios, and process parameters; then the SHAP method is used to analyze the contribution degree of each feature to Tg, and the candidate components, ratios, and process parameters are optimized; S4: Obtain the overall optimization solution Taking Tg≥180°C and viscosity≤5000 cP as constraints, the Bayesian optimization algorithm is used to search for the Pareto optimal solution in the epoxy monomer-curing agent ratio space, obtaining the optimization solution of the epoxy resin-curing agent system with the high glass transition temperature of the epoxy resin-curing agent system as the target parameter, including the overall optimization solution of components, ratios, and process parameters.

[0007] A system for designing a chip underfill with a high glass transition temperature based on machine learning, which is used to implement the foregoing method, includes: 1) Data input module: used to import the molecular structures of epoxy monomers and curing agents and process parameters; 2) Feature engineering module: perform RFE-XGBoost feature screening to generate a key feature set; 3) Model calculation module: run the GPR model to predict Tg and output the feature contribution degree through the SHAP interpreter; 4) Optimization output module: generate a combination of formulations and process parameters that meet the Tg and viscosity constraints based on Bayesian optimization.

[0008] The optimization output module is connected to the molecular dynamics simulation software interface to verify the crosslinked network morphology of the predicted formulation and output a candidate system with a crosslinking degree≥85%.

[0009] A chip underfill with a high glass transition temperature, characterized in that it is designed by the foregoing method and is made of components in the following proportions: bisphenol A epoxy (DGEBA) 40-55 wt%, tetrafunctional epoxy (TGDDM) 15-35 wt%, naphthalene ring modified epoxy (NE) 5-10 wt%, and 20-30 wt% boron nitride (BN) is added as a thermal conductive filler, with a glass transition temperature Tg≥195°C and a viscosity≤4500 cP (25°C); its curing process is three-stage gradient heating: 120°C / 1 h - 170°C / 3 h - 200°C / 1 h, with a heating rate of 2°C / min.

[0010] Compared with the prior art, the present invention at least includes the following beneficial effects: 1. The present invention provides a method and system for designing an epoxy resin-curing agent system chip underfill based on the high glass transition temperature (Tg) as the core target parameter by machine learning. It mainly quantifies the influence laws of molecular structure and process parameters on Tg through machine learning to realize the generation of candidate solutions and subsequent efficient optimization. The present invention rapidly predicts and optimizes the Tg of the epoxy resin matrix through a machine learning model, avoiding a large number of repeated experiments of the traditional trial-and-error method, significantly shortening the R & D cycle, and improving the R & D efficiency.

[0011] 2. The machine learning with a multi-parameter and multi-constraint system adopted by the present invention reduces the number of experiments and material consumption, lowers the R & D cost, and at the same time avoids the waste of resources caused by improper experimental conditions. It can comprehensively explore the complex formulation and process parameter space, screen out the optimal combination of formulation and process parameters, significantly improve the glass transition temperature of the chip underfill, enhance its thermal stability and reliability, and meet the requirements of high-performance underfill for chip packaging. The method and system provided by the present invention collect and analyze a large amount of experimental data, construct a machine learning model, and provide a method for predicting the glass transition temperature of epoxy resin, which can accelerate the discovery of new epoxy resins with high glass transition temperature and significantly improve the glass transition temperature of the chip underfill.

[0012] 3. Actual experiments and verification show that the measured Tg of the preferred formulation (such as the ternary functional group epoxy / bisphenol A epoxy / naphthalic anhydride complex system) obtained by the present invention reaches 205 - 215 °C, which is more than 20% higher than the traditional formulation, and the model prediction error is < ±6 °C.

[0013] 4. The present invention adopts a new technical idea, uses RFE nested XGBoost to screen features, uses the XGBoost algorithm to build a model, and uses the SHAP method to explain the relationship between structure and performance to predict epoxy resins with high Tg values; the method and system have low operating costs, are simple and efficient, have complete and accurate data, and are environmentally friendly and pollution-free throughout the process. It solves the bottleneck that the design of the Tg of the epoxy resin matrix depends on the trial-and-error method, can shorten the R & D cycle by more than 60%, and is applicable to high-reliability packaging scenarios such as 5G chips and power devices.

[0014] 5. The entire design process of the present invention does not involve complex chemical experiments and high-energy-consuming equipment, conforms to the development concept of green environmental protection, and is easy to realize high-throughput design and industrialized preparation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the system composition and method flow for predicting and optimizing the glass transition temperature of epoxy resin in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, so that those skilled in the art can implement it according to the description in the specification.

[0017] It should be understood that the terms such as "having", "comprising", and "including" used herein do not exclude the presence or addition of one or more other elements or their combinations.

[0018] Unless otherwise specifically stated, various raw materials, reagents, instruments, and equipment used in the present invention can be obtained through market purchase or can be prepared by existing methods.

[0019] Basic Embodiment Refer to the attached Figure 1 , a method for designing a high glass transition temperature chip underfill adhesive based on machine learning provided by an embodiment of the present invention takes the high glass transition temperature Tg of the epoxy resin-curing agent system as the main target parameter, and includes the following steps: S1: Construct polymer molecular structure parameters and process parameters Using a public polymer database, obtain the molecular structure parameters of epoxy resin monomers and curing agents and the corresponding measured Tg values. The molecular structure parameters include the number of functional groups (nEpoxy), aromatic ring density (Ar-density), molecular weight (MW) of epoxy monomers, as well as the amine hydrogen equivalent (AHE) and steric hindrance parameter (Steric-index) of the curing agent, to obtain the corresponding molecular parameter descriptor set and process parameter descriptor set; S2: Screen the key feature set Adopt the recursive feature elimination (RFE) combined with the XGBoost algorithm to screen the key feature set from the molecular parameter descriptor set and the process parameter descriptor set, where the process parameters include the curing step temperature and curing time; S3: Prediction based on machine learning Based on the screened feature set, construct a Gaussian process regression (GPR) model to predict the Tg of the epoxy resin-curing agent system, and obtain candidate solutions for components, ratios, and process parameters; then use the SHAP method to analyze the contribution degree of each feature to Tg, and optimize the candidate components, ratios, and process parameters; S4: Obtain the overall optimization plan Taking Tg≥180°C and viscosity≤5000 cP as constraints, use the Bayesian optimization algorithm to search for the Pareto optimal solution in the epoxy monomer-curing agent ratio space, and obtain an optimization plan for the epoxy resin-curing agent system with the high glass transition temperature of the epoxy resin-curing agent system as the target parameter, including the overall optimization plan for components, ratios, and process parameters.

[0020] The epoxy monomer in step S1 is at least one of bisphenol A epoxy (DGEBA), tetrafunctional epoxy (TGDDM), and naphthalene ring modified epoxy, and the compounding ratio satisfies: the proportion of bifunctional epoxy is 40-70 wt%, and the proportion of polyfunctional epoxy is 30-60 wt%.

[0021] The curing agent in step S1 is an aromatic amine or acid anhydride compound, and its steric hindrance parameter Steric-index is obtained by calculating the superposition volume of atomic van der Waals radii through molecular dynamics simulation, and the value range is 0.6-0.9.

[0022] The curing agent is selected from at least one of 4,4'-diaminodiphenyl sulfone (DDS), methyl nadic anhydride (MNA), and benzophenone tetracarboxylic dianhydride (BTDA), and the amine hydrogen equivalent (AHE) is 80-120 g / eq.

[0023] The optimization of process parameters in step S3 includes: dividing the stepwise curing program into 3-5 stages, with an initial curing temperature of 80-120 °C, a final stage curing temperature of 160-200 °C, and a total curing duration of 2-8 h.

[0024] In the SHAP analysis process of step S3, the total contribution weight of the number of epoxy monomer functional groups (nEpoxy) and the aromatic ring density (Ar-density) ≥ 60%, and the absolute value of the negative contribution weight of the curing agent steric hindrance parameter (Steric-index) ≤ 15%.

[0025] A system for designing a high glass transition temperature chip underfill based on machine learning, which is used to implement the foregoing method, includes: 1) Data input module: used to import the molecular structures of epoxy monomers and curing agents and process parameters; 2) Feature engineering module: perform RFE-XGBoost feature screening to generate a key feature set; 3) Model calculation module: run the GPR model to predict Tg and output the feature contribution degree through the SHAP interpreter; 4) Optimization output module: generate a combination of formulations and process parameters that meet the Tg and viscosity constraints based on Bayesian optimization.

[0026] The optimization output module is connected to the molecular dynamics simulation software interface to verify the crosslinked network morphology of the predicted formulation and output a candidate system with a crosslinking degree ≥ 85%.

[0027] A chip underfill adhesive with a high glass transition temperature is designed by the aforementioned method and is made of components in the following proportions: bisphenol A epoxy (DGEBA) 40 - 55 wt%, tetrafunctional epoxy (TGDDM) 15 - 35 wt%, naphthalene ring modified epoxy (NE) 5 - 10 wt%, and 20 - 30 wt% boron nitride (BN) is added as a thermal conductive filler. The glass transition temperature Tg ≥ 195 °C, and the viscosity ≤ 4500 cP (25 °C); its curing process is three-stage gradient heating: 120 °C / 1 h - 170 °C / 3 h - 200 °C / 1 h, with a heating rate of 2 °C / min.

[0028] The above technical solutions will be further described below in conjunction with specific examples. The preferred embodiments of the present invention are described in detail as follows: Example 1

[0029] The method, system and application for designing a chip underfill adhesive with a high glass transition temperature based on machine learning provided in the embodiments of the present invention are specific implementations of the foregoing basic embodiments, and the differences are as follows: The method for designing a chip underfill adhesive with a high glass transition temperature based on machine learning takes the high glass transition temperature Tg of the epoxy resin-curing agent system as the main target parameter, collects and preprocesses data on factors affecting the prediction of the glass transition temperature of epoxy resin, and then performs prediction, analysis and optimization. The specific steps of step S1 are as follows: (1) Obtain 200 groups of epoxy resin-curing agent system data from PolyInfo and experiments, covering monomers such as bisphenol A epoxy (DGEBA), tetrafunctional epoxy (TGDDM), naphthalene ring modified epoxy (NE), and curing agents such as DDS, MNA, and BTDA. Include Tg (120 - 250 °C, DMA test, tanδ peak method), and simultaneously record the viscosity (25 °C) as a processability constraint; (2) Use ChemAxon to optimize the molecular structure, extract descriptors for epoxy monomers and curing agents. For epoxy monomers, the number of functional groups of DGEBA, TGDDM, and NE are set to 2, 4, and 3 respectively; the aromatic ring density is determined by the number of aromatic ring atoms in a unit molecule. For example, DGEBA = 0.032 atoms / g / mol; calculate the polar surface area (TPSA) through Dragon software (such as DGEBA = 45.2 Ų); the amine hydrogen equivalent (AHE) of curing agents DDS and MNA are 107 g / eq and 178 g / eq respectively; the steric hindrance parameter is based on the atomic packing volume calculated by molecular dynamics simulation (DDS = 0.72, MNA = 0.65); (3) Collect process parameters. For example: the number of stepped heating segments is divided into 2 - 5 segments, the final segment temperature is set to 160 °C - 220 °C, and the curing time is set to 1 - 6 h; (4)For the collection of filler information, it is mainly boron nitride filler with a content of 20 - 40 wt% and a particle size of 1 - 10 μm.

[0030] (5)The first specific raw material components and process for the underfill adhesive are as follows: bisphenol A epoxy (DGEBA) 50 wt%, tetrafunctional epoxy (TGDDM) 22 wt%, naphthalene ring modified epoxy (NE) 8 wt%, and 20 wt% boron nitride (BN) is added as a thermal conductive filler. The measured Tg ≥ 195 °C and the viscosity ≤ 4500 cP (25 °C); its curing process is three-stage gradient heating: 120 °C / 1 h - 170 °C / 3 h - 200 °C / 1 h, and the heating rate is 2 °C / min; The epoxy resin monomer in this example is bisphenol A epoxy (DGEBA), and the compounding ratio satisfies: the proportion of bifunctional epoxy is 40 wt%, and the proportion of polyfunctional epoxy is 60 wt%; the compounding ratios of other examples can also satisfy: the proportion of bifunctional epoxy is 50 wt%, and the proportion of polyfunctional epoxy is 50 wt%; or the proportion of bifunctional epoxy is 70 wt%, and the proportion of polyfunctional epoxy is 30 wt%.

[0031] Example 2 The method, system and application for designing a high glass transition temperature chip underfill adhesive based on machine learning provided by the embodiment of the present invention are specific to the foregoing basic embodiment. It is basically the same as Embodiment 1, and the difference is that step S2 for screening the key feature set specifically further includes the following steps: Carry out model training and verification for predicting the glass transition temperature of epoxy resin, including the following steps: (1)A total of 85 features including epoxy monomers, curing agents and process parameters; (2)Using recursive feature elimination (RFE) combined with the XGBoost algorithm to screen the top 8 key features, which are: the number of functional groups, aromatic ring density, polar surface area, amine hydrogen equivalent, steric hindrance, final temperature, curing time and boron nitride content; (3)Using Gaussian process regression (GPR) with a kernel function of Matern 5 / 2; (4)Dividing 160 groups (80%) into the training set and 40 groups (20%) into the test set, and using 5-fold cross-validation; (5)The R² of the model training set = 0.93 and the MAE = 5.8 °C; the R² of the test set = 0.89 and the MAE = 7.1 °C; (6) SHAP value analysis was adopted for interpretability analysis. The positive contributions included the number of functional groups (+42%), the aromatic ring density (+28%), and the final stage time (+15%). The negative contribution was the steric hindrance parameter (-10%, excessive value led to hindered crosslinking). For each increase of 1 in the functionality of the epoxy monomer, the Tg increased by an average of 32 ± 5 °C (confidence level 95%).

[0032] (7) The second specific raw material components and process of the underfill were as follows: bisphenol A type epoxy (DGEBA) 40 wt%, tetrafunctional epoxy (TGDDM) 35 wt%, naphthalene ring modified epoxy (NE) 5 wt%, and 20 wt% boron nitride (BN) was added as a thermal conductive filler. The measured Tg ≥ 195 °C and the viscosity ≤ 4500 cP (25 °C). Its curing process was three-stage gradient heating: 120 °C / 1 h - 170 °C / 3 h - 200 °C / 1 h, with a heating rate of 2 °C / min.

[0033] Example 3

[0034] The method, system, and application for designing a high glass transition temperature chip underfill based on machine learning provided by the embodiments of the present invention are specific implementations of the foregoing basic embodiments. They are basically the same as Embodiment 1 and Embodiment 2. The differences are that the formulation optimization and experimental verification for predicting the glass transition temperature of epoxy resin, and the specific steps of step S3 based on machine learning prediction further include the following steps: (1) The optimization target parameter was to maximize Tg (≥180 °C), and the constraint was that the viscosity ≤ 5000 cP (25 °C); The content of the filler boron nitride was fixed at 35 wt% (particle size 5 μm); (2) The best formulation obtained by Bayesian optimization was that when the epoxy monomer component was 40% TGDDM + 30% NE + 30% DGEBA, the curing agent component was DDS (amine hydrogen equivalent was 107 g / eq, and the steric hindrance parameter was 0.72), and the process parameters were three-stage curing of (120 °C / 1 h → 170 °C / 3 h → 200 °C / 1 h); (3) According to the above for prediction, the predicted Tg = 218 ± 6 °C and the viscosity = 4800 ± 300 cP.

[0035] (4) The epoxy monomer and DDS were mixed according to the ratio, the BN filler was added, and vacuum degassing was carried out for 30 min; curing was carried out step by step according to the optimized process to obtain an underfill sample. The measured Tg was 215 °C (DMA, tanδ peak); the viscosity was 4650 cP (25 °C, Brookfield DV2T); the prediction error was a Tg deviation of -1.4% (<5% threshold, the model was valid); Compared with the traditional sample DGEBA 100% + MNA (Tg = 165°C), the optimized formulation has a 30.3% improvement.

[0036] (6)The third specific raw material components and process of the underfill are as follows: bisphenol A epoxy (DGEBA) 40wt%, tetrafunctional epoxy (TGDDM) 20wt%, naphthalene ring modified epoxy (NE) 10wt%, and 30wt% boron nitride (BN) is added as a thermal conductive filler. The measured Tg ≥ 195°C and the viscosity ≤ 4500 cP (25°C); its curing process is three-stage gradient heating: 120°C / 1h - 170°C / 3h - 200°C / 1h, with a heating rate of 2 °C / min.

[0037] Example 4

[0038] The method, system and application for designing a high glass transition temperature chip underfill based on machine learning provided by the embodiments of the present invention are specific implementations of the foregoing basic embodiments. They are basically the same as Embodiments 1 - 3, except that the step of verifying the robustness of the model for predicting the glass transition temperature of epoxy resin includes the following content: (1)Input the new epoxy monomer (pentafunctional PX-100, nEpoxy = 5, Ar_density = 0.045), and predict Tg = 242°C; the measured Tg = 230°C, with an error of 4.9% (verifying the generalization ability of the model).

[0039] (2)The curing final stage temperature fluctuates by ±10°C (190 - 210°C): predict that the Tg changes by ±8°C, and the measured value is ±7 - 9°C; the model accurately captures the process sensitivity and guides the temperature control accuracy of the production line (required ±5°C).

[0040] (3)The fourth specific raw material components and process of the underfill are as follows: bisphenol A epoxy (DGEBA) 55wt%, tetrafunctional epoxy (TGDDM) 15wt%, naphthalene ring modified epoxy (NE) 7.5wt%, and 22.5wt% boron nitride (BN) is added as a thermal conductive filler. The measured Tg ≥ 195°C and the viscosity ≤ 4500 cP (25°C); its curing process is three-stage gradient heating: 120°C / 1h - 170°C / 3h - 200°C / 1h, with a heating rate of 2 °C / min.

[0041] Example 5 In this embodiment, based on Embodiments 1 - 4, a failure case and feedback iteration step for predicting the glass transition temperature of epoxy resin are further provided, including the following content: (1)Input (50%) NE + (Steric_index = 0.65) MNA, predicted Tg = 188 °C; measured Tg = 162 °C, with an error of 14.3%, thus triggering re-screening of features; (2)Based on the above analysis, the reason may be that the compatibility (polarity difference) between MNA and NE was not considered, and a new feature "epoxy-curing agent polarity matching degree" was added; after subsequent iteration, the R² of the model was improved to 0.91, and MAE = 6.2 °C.

[0042] In summary, the methods for predicting, optimizing, and verifying the glass transition temperature of epoxy resins provided by the above embodiments generally include the following steps: (1)Using a publicly available polymer database, obtain the molecular structure parameters of epoxy resin monomers and curing agents and their corresponding measured Tg values. The molecular structure parameters include the number of functional groups (nEpoxy), aromatic ring density (Ar-density), molecular weight (MW) of epoxy monomers, as well as the amine hydrogen equivalent (AHE) and steric hindrance parameter (Steric-index) of the curing agent; (2)Adopt the recursive feature elimination (RFE) combined with the XGBoost algorithm to screen the key feature set from the molecular structure parameters and process parameters. The process parameters include the curing step temperature and curing time; (3)Based on the screened feature set, construct a Gaussian process regression (GPR) model to predict the Tg of the epoxy resin-curing agent system, and use the SHAP method to analyze the contribution degree of each feature to Tg; (4)Taking Tg ≥ 180 °C and viscosity ≤ 5000 cP as constraints, use the Bayesian optimization algorithm to search for the Pareto optimal solution in the epoxy monomer-curing agent ratio space; the epoxy resin monomer is at least one of bisphenol A epoxy (DGEBA), tetrafunctional epoxy (TGDDM), and naphthalene ring-modified epoxy, and the compounding ratio satisfies: the proportion of bifunctional epoxy is 40 - 70 wt%, and the proportion of polyfunctional epoxy is 30 - 60 wt%; The curing agent is an aromatic amine or acid anhydride compound, and its steric hindrance parameter Steric-index is obtained by calculating the superposition volume of atomic van der Waals radii through molecular dynamics simulation, with a value range of 0.6 - 0.9; the curing agent is selected from at least one of 4,4'-diaminodiphenyl sulfone (DDS), methyl nadic anhydride (MNA), and benzophenone tetracarboxylic dianhydride (BTDA), and the amine hydrogen equivalent (AHE) is 80 - 120 g / eq; the optimization of the process parameters includes: the step curing program is divided into 3 - 5 stages, the initial curing temperature is 80 - 120 °C, the final stage curing temperature is 160 - 200 °C, and the total curing duration is 2 - 8 h; (5) In the SHAP analysis, the sum of the contribution weights of the number of epoxy monomer functional groups (nEpoxy) and the aromatic ring density (Ar-density) ≥ 60%, and the absolute value of the negative contribution weight of the steric hindrance parameter (Steric-index) of the curing agent ≤ 15%.

[0043] The design methods provided by the above embodiments avoid the process of repeated experiments and continuous trial and error. By using Chem3D to optimize molecules, ChemAxon software to generate structure descriptors, and through feature screening and modeling by the RFE nested XGBoost method, the Tg value of epoxy resin can be predicted in advance. The operation process is simple, low-cost, and pollution-free, and can be completed by only one person. In the above embodiment methods, SHAP and reference documents are used to explain features and construct a QSPR model to better explain the relationship between molecular structure and performance, and try to solve the shortcoming of the black box of most methods. The above embodiment methods use XGBoost to establish a model for quickly predicting the Tg value of epoxy resin, which can greatly shorten the R & D time of new epoxy resins and reduce the R & D cost.

[0044] The key points of the above embodiments of the present invention are: first, construct a molecular descriptor set of epoxy resin monomers (number of functional groups, aromatic ring density) and curing agents (amine hydrogen equivalent, steric hindrance parameter); then use recursive feature elimination (RFE) combined with the XGBoost algorithm to screen key features, establish a Gaussian process regression (GPR) model to predict Tg, and analyze the structure contribution degree by the SHAP method; then combine Bayesian optimization to screen candidate systems with high Tg (> 180 °C) and low viscosity (< 5000 cP) in the synergistic space of epoxy-curing agent ratio and step curing process; finally, optimize and verify the scheme.

[0045] Experimental verification shows that the measured Tg of the optimized formula (such as the ternary functional group epoxy / bisphenol A epoxy / naphthalic anhydride complex system) reaches 205 - 215 °C, which is more than 20% higher than the traditional formula, and the model prediction error < ±6 °C.

[0046] The above embodiments of the present invention have successfully designed an epoxy resin matrix chip underfill with a high glass transition temperature through machine learning technology, significantly improving the thermal stability and reliability of the underfill, providing new ideas and methods for the R & D of high-performance chip packaging materials, solving the bottleneck of the Tg design of epoxy resin matrix relying on the trial and error method, shortening the R & D cycle by more than 60%, and being applicable to high-reliability packaging scenarios such as 5G chips and power devices. The present invention not only improves the R & D efficiency, reduces the R & D cost, but also conforms to the development concept of green environmental protection and has broad application prospects.

[0047] It should be specifically noted that within the scope of the components, ratios, and process parameters described in the present invention, other technical solutions obtained through specific selections can all achieve the technical effects of the present invention, so they will not be listed one by one. Other technical solutions obtained by using components, ratios, preparation methods, and applications equivalent to those described in the present invention are all included in the protection scope of the present invention.

[0048] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for designing a high glass transition temperature chip underfill based on machine learning, characterized in that: Taking the high glass transition temperature Tg of the epoxy resin-curing agent system as the main target parameter, it includes the following steps: S1: Constructing polymer molecular structure parameters and process parameters Using a public polymer database, the molecular structure parameters of epoxy resin monomers and curing agents and the corresponding Tg measured values ​​are obtained. The molecular structure parameters include the number of functional groups (nEpoxy), aromatic ring density (Ar-density), molecular weight (MW) of the epoxy monomer, and the amine hydrogen equivalent (AHE) and steric hindrance parameter (Steric-index) of the curing agent, and the corresponding molecular parameter descriptor set and process parameter descriptor set are obtained; S2: Screening key feature sets Recursive feature elimination (RFE) combined with XGBoost algorithm was used to screen key feature sets from the molecular parameter descriptor set and the process parameter descriptor set, where the process parameters included curing step temperature and curing time. S3: Prediction based on machine learning Based on the screened feature set, a Gaussian process regression (GPR) model was constructed to predict the Tg of the epoxy resin-curing agent system, and candidate solutions including components, ratios and process parameters were obtained. The SHAP method was then used to analyze the contribution of each feature to Tg, and the candidate components, ratios and process parameters were optimized. S4: Get the overall optimization solution With Tg ≥ 180 ℃ and viscosity ≤ 5000 cP as constraints, the Bayesian optimization algorithm was used to search for the Pareto optimal solution in the epoxy monomer-curing agent ratio space, and an optimization scheme for the epoxy resin-curing agent system with a high glass transition temperature of the epoxy resin-curing agent system as the target parameter was obtained, including an overall optimization scheme for components, ratios and process parameters.

2. The method for designing a high glass transition temperature chip underfill based on machine learning according to claim 1, characterized in that: The epoxy resin monomer in step S1 is at least one of bisphenol A epoxy (DGEBA), tetrafunctional epoxy (TGDDM), and naphthalene ring modified epoxy, and the compounding ratio satisfies: the bifunctional epoxy accounts for 40-70wt% and the multifunctional epoxy accounts for 30-60wt%.

3. The method for designing a high glass transition temperature chip underfill based on machine learning according to claim 1, characterized in that: In the step S1, the curing agent is an aromatic amine or anhydride compound, and the steric hindrance parameter Steric-index is obtained by calculating the atomic van der Waals radius superposition volume through molecular dynamics simulation, and the value range is 0.6-0.

9.

4. The method for designing a high glass transition temperature chip underfill based on machine learning according to claim 3, characterized in that: The curing agent is selected from at least one of 4,4'-diaminodiphenyl sulfone (DDS), methyl nadic anhydride (MNA), and benzophenone tetracarboxylic dianhydride (BTDA), and has an amine hydrogen equivalent (AHE) of 80-120 g / eq.

5. The method for designing a high glass transition temperature chip underfill based on machine learning according to claim 1, characterized in that: The optimization of the process parameters in step S3 includes: dividing the step curing procedure into 3-5 stages, the initial curing temperature is 80-120° C., the final curing temperature is 160-200° C., and the total curing time is 2-8 hours.

6. The method for designing a high glass transition temperature chip underfill based on machine learning according to claim 1, characterized in that: During the SHAP analysis of step S3, the total weight of the contribution of the number of epoxy monomer functional groups (nEpoxy) and the aromatic ring density (Ar-density) is ≥60%, and the absolute value of the negative weight of the contribution of the curing agent steric hindrance parameter (Steric-index) is ≤15%.

7. A system for designing high glass transition temperature chip underfill based on machine learning, characterized in that: It is used to implement the method according to any one of claims 1 to 6, comprising: 1) Data input module: used to import epoxy monomer, curing agent molecular structure and process parameters; 2) Feature Engineering Module: Perform RFE-XGBoost feature screening to generate key feature sets; 3) Model calculation module: Run the GPR model to predict Tg and output feature contribution through the SHAP interpreter; 4) Optimization output module: Generates formula and process parameter combinations that meet Tg and viscosity constraints based on Bayesian optimization.

8. The system for designing high glass transition temperature chip underfill based on machine learning according to claim 7, characterized in that: The optimization output module is connected to the molecular dynamics simulation software interface to verify the cross-linked network morphology of the predicted formula and output a candidate system with a cross-linking degree of ≥85%.

9. A chip underfill with a high glass transition temperature, characterized in that: The thermal conductive filler is designed by the method described in any one of claims 1 to 6, and is made of components in the following proportions: 40-55wt% of bisphenol A epoxy (DGEBA), 15-35wt% of tetrafunctional epoxy (TGDDM), 5-10wt% of naphthalene ring modified epoxy (NE), and 20-30wt% of boron nitride (BN) is added as a thermal conductive filler, with a glass transition temperature Tg ≥ 195°C and a viscosity ≤ 4500cP (25°C).

10. The chip underfill according to claim 9, characterized in that: The curing process is a three-stage gradient heating process: 120°C / 1h-170°C / 3h-200°C / 1h, with a heating rate of 2°C / min.

Images