A machine learning based dynamic optimization method for radiation tolerant chip shielding package enhancement scheme

Abstract

The application relates to a kind of dynamic optimization methods of radiation-resistant chip shielding packaging enhancement scheme based on machine learning, comprising the following steps: S1, obtaining ray type, energy flux, temperature field distribution;S2, shielding filler formula and P-Zeta potential modification parameter are generated with material genetic engineering model;S3, kneading process is executed in packaging production line and temperature, shear, pressurization are adjusted in real time with reinforcement learning;S4, real-time process data is collected, and chip radiation damage and life are predicted by digital twin system;S5, according to the prediction result, the final packaging scheme is optimized and solidified in reverse.Compared with the prior art, the application takes material genetic engineering model as the core, and directly maps environmental parameters such as ray type, energy flux and temperature field into shielding filler formula and P-Zeta potential modification scheme, significantly shortens the development cycle and reduces the test cost.

Description

Technical Field

[0001] This invention relates to the field of radiation-resistant chip shielding and packaging technology, and in particular to a dynamic optimization method for radiation-resistant chip shielding and packaging enhancement schemes based on machine learning. Background Technology

[0002] In high-radiation environments such as nuclear energy, aerospace, and deep space exploration, memory chips are exposed to high-energy particles and gamma rays for extended periods, making them highly susceptible to radiation damage such as single-event upsets, threshold drift, and gate oxide breakdown, leading to system failure. Traditional radiation-resistant designs rely on process node redundancy or circuit-level hardening, which is costly, time-consuming, and suffers from "over-design" issues in low-to-medium dose scenarios. Furthermore, current molding compounds are primarily based on general-purpose epoxy systems, lacking radiation-specific shielding fillers and interface control methods. The encapsulation layer itself becomes a "secondary source" of radiation damage, failing to meet the urgent needs of next-generation equipment for highly reliable, miniaturized memory chips.

[0003] The emerging "shielded packaging" approach attempts to add an irradiation shielding layer to conventional chips, but it faces two major bottlenecks: First, the synergistic design of filler formulations and resin systems lacks theoretical tools, often relying on trial and error, making it difficult to balance shielding effectiveness and process feasibility. Second, the process window is narrow; even small fluctuations in temperature, shear, and pressure parameters during the kneading-molding process can cause filler agglomeration and potential drift, thereby reducing shielding consistency. Current technologies have not yet established a closed-loop optimization system encompassing "environmental parameters—material formulation—process parameters—service life," making it impossible to dynamically adjust the packaging scheme according to the target scenario, thus limiting the engineering implementation of shielded packaging technology.

[0004] The rise of multi-scale physical simulations and materials genome engineering technologies has provided new avenues for the design of shielding fillers. Tools such as VESTA, Materials Studio, and VASP can predict the interaction between radiation and matter, interfacial binding energy, and filler dispersion stability at the atomic-mesoscopic scale, thereby establishing a quantitative correlation between shielding effectiveness and microstructure. However, significant discrepancies remain between these simulation results and actual process conditions: on the one hand, simulations are typically based on ideal interfaces and do not consider the influence of temperature gradients and shear history on the P-Zeta potential of the filler surface during kneading; on the other hand, the lack of an effective feedback mechanism for real-time process data causes models to remain in the "offline optimization" stage for extended periods, unable to adaptively correct for production line fluctuations.

[0005] The integration of reinforcement learning and digital twin technology offers a feasible solution to the aforementioned challenges. By mapping real-time process data, environmental parameters, and package-level performance to the same digital twin framework, the "process-performance" relationship can be rapidly iterated in virtual space, and the optimal strategy can be distributed to the production line controller to achieve millisecond-level closed-loop regulation. However, current research mainly focuses on homogeneous materials or macroscopic structures, lacking cross-scale twin models for multiphase systems such as "filler-resin-chip," and failing to establish a smooth transition mechanism from "laboratory formulation" to "small-batch customization."

[0006] Therefore, there is an urgent need to establish a dynamic optimization method for radiation-resistant chip shielding and packaging based on machine learning and materials genome engineering, so as to connect the entire chain of design-manufacturing-verification and promote the large-scale application of shielding and packaging technology in the fields of nuclear energy, aerospace and military. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a dynamic optimization method for radiation-resistant chip shielding and packaging enhancement based on machine learning. With a materials genome engineering model as the core, environmental parameters such as radiation type, energy flux, and temperature field are directly mapped to shielding filler formulation and P-Zeta potential modification scheme. This eliminates the lengthy process of repeated sampling in traditional methods, opens up the entire chain of design-manufacturing-verification, significantly shortens the R&D cycle and reduces testing costs.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] This invention provides a dynamic optimization method for radiation-resistant chip shielding and packaging enhancement schemes based on machine learning, characterized by the following steps:

[0010] S1. Obtain target irradiation environment parameters, including radiation type, energy flux, and temperature field distribution;

[0011] S2. Based on the environmental parameters, a shielding filler formulation and corresponding P-Zeta potential modification parameters are generated through a materials genome engineering model, wherein the materials genome engineering model is obtained by training integrated multi-scale physical simulation data.

[0012] S3. Based on the shielding filler formulation and P-Zeta potential modification parameters, the packaging production line performs a kneading process, and the process parameters are adjusted in real time through a reinforcement learning algorithm. The process parameters include temperature control, shear rate, and pressurization sequence.

[0013] S4. Collect real-time process data during the packaging process, combine it with the environmental parameters and input it into the digital twin verification system to generate chip irradiation damage distribution and packaging failure lifetime prediction.

[0014] S5. Based on the irradiation damage distribution and encapsulation failure lifetime prediction, determine the encapsulation scheme that meets the shielding requirements of the target scenario.

[0015] Furthermore, in S1, the specific process of obtaining the target irradiation environment parameters includes:

[0016] S1-1. Call the Phy-X / PSD database and VESTA / Materials Studio interface to retrieve the ray type-energy flux coupling curve corresponding to the target application scenario;

[0017] S1-2. Using the atomic-level temperature field distribution calculated by VASP as the initial temperature field, the temperature field distribution is mapped to the chip packaging-level mesh through an interpolation algorithm to obtain the temperature field distribution.

[0018] S1-3. Encapsulate the ray type, energy flux, and temperature field distribution into a unified format environmental parameter vector.

[0019] Furthermore, in S2, the specific process of generating the shielding filler formulation and corresponding P-Zeta potential modification parameters based on the environmental parameters using a materials genome engineering model includes:

[0020] S2-1. Input the environmental parameter vector obtained in S1-3 into the material genome engineering model. The model uses the multi-scale physical simulation dataset output by the shielding filler calculation and design platform as the training set. The dataset includes the radiation absorption cross section, the filler-resin interface energy, the P-Zeta potential, and the dispersion stability correlation curve.

[0021] S2-2. During the training process, a Bayesian optimizer is used to sample the filler composition space and iteratively update the ternary formulation of filler, resin and modifier until the predicted shielding effectiveness index and interface potential index converge simultaneously.

[0022] S2-3. Output the converged shielding filler formulation and the corresponding P-Zeta potential modification parameters. The formulation includes the proportion of special resin for shielding filler, the type and amount of dispersant modifier, and the P-Zeta potential modification parameters include the surface potential adjustment range and pH window range.

[0023] Furthermore, in S2, the process of acquiring the materials genome engineering model through training by integrating multi-scale physical simulation data includes:

[0024] Using the shielding filler computational design platform as the data source, VESTA is called sequentially to establish the atomic-scale crystal structure of the filler, Materials Studio is used to construct the filler-resin interface model, and VASP is used to calculate the ray-matter interaction cross section, generating multi-scale physical simulation data at the atomic-mesoscopic scale.

[0025] The multi-scale physical simulation data are labeled according to ray type, energy flux, and temperature field distribution to form a training sample library;

[0026] A graph neural network architecture is adopted, with the filler component vector, interface atomic coordinates and P-Zeta potential as input nodes, and shielding effectiveness and dispersion stability as output nodes. The sample library is iteratively trained until the prediction error is less than the preset value, thereby obtaining the material genome engineering model.

[0027] Furthermore, in S3, the specific process of performing the kneading process on the encapsulation production line according to the shielding filler formulation and P-Zeta potential modification parameters includes:

[0028] S3-1. Based on the shielding filler formula output by S2-3, the shielding filler special resin and dispersant modifier are fed into the AB mixing unit in a set ratio to form a premix.

[0029] S3-2. Using the P-Zeta potential modification parameter, the surface potential of the premix is ​​controlled in the kneader, the potential value is monitored in real time, and the pH window range is adjusted in a closed loop.

[0030] S3-3. During the kneading process, a reinforcement learning algorithm is used to receive real-time feedback from temperature sensors, shear rate sensors, and pressure sensors, generate dynamic process instructions for temperature control, shear rate, and pressurization timing, and adjust the kneading machine actuator at the millisecond level through the controller until a homogeneous shielding molding compound is obtained.

[0031] Furthermore, in S3, the specific process of adjusting process parameters in real time using reinforcement learning algorithms includes:

[0032] The obtained real-time values ​​of P-Zeta potential, instantaneous values ​​of shielding molding compound viscosity, and instantaneous values ​​of filler dispersion are used to construct a state vector, which is then used as the input to the reinforcement learning agent.

[0033] The agent uses temperature control value, shear rate value, and pressurization timing value as its action space. It updates its strategy in the continuous action-state space according to the reward function, which takes maximizing the uniformity of filler dispersion and minimizing potential drift as weighted objectives.

[0034] Within each batch kneading cycle, action commands are output at a millisecond frequency to the kneader temperature control module, shear drive module, and pressurization servo module to form a closed-loop real-time adjustment until the state vector converges to the preset threshold range, thereby completing the dynamic optimization of process parameters.

[0035] Furthermore, in S4, the specific process of collecting real-time process data during the packaging process, combining it with the environmental parameters, and inputting it into the digital twin verification system to generate chip irradiation damage distribution and packaging failure lifetime prediction includes:

[0036] S4-1. After the shielding molding compound formed in S3 is injected into the mold cavity and the chip is encapsulated in the second stage, the temperature, residual stress and filler distribution images are collected in real time by distributed optical fiber sensors and MEMS temperature-stress sensors to form real-time process data.

[0037] S4-2. Input the real-time process data and the environmental parameter vector output in S1-3 into the digital twin verification system. The system has a built-in irradiation-mechanical coupling model based on VASP calibration and calculates the energy deposition field after the interaction between the ray and the packaging material through a finite element solver.

[0038] S4-3. Based on the energy deposition field results, the failure criterion library output by the materials genome engineering model is called to map the irradiation damage distribution inside the chip, and the Miner cumulative damage criterion is combined to predict the packaging failure lifetime, outputting the chip irradiation damage distribution map and the packaging failure lifetime curve.

[0039] Furthermore, in S5, the specific process of determining the packaging scheme that meets the shielding requirements of the target scenario based on irradiation damage distribution and packaging failure lifetime prediction includes:

[0040] S5-1. Import the chip irradiation damage distribution map and the packaging failure lifetime curve output by S4-3 into the shielding packaging scheme decision engine. The decision engine uses the total irradiation dose threshold and the minimum service life threshold of the target application scenario as hard constraints.

[0041] S5-2. The decision engine calls the formulation-lifetime mapping matrix in the material genome engineering model to reversely retrieve the shielding filler formulation range, P-Zeta potential modification parameter range and corresponding process parameter range that meet the hard constraint conditions, and generates a set of candidate encapsulation schemes.

[0042] S5-3. Perform multi-objective optimization on the candidate encapsulation scheme set, with the optimization objectives of maximizing shielding effectiveness and minimizing quality increment. Use a non-dominated sorting genetic algorithm to iteratively solve the problem and output the optimal solution of the Pareto front.

[0043] S5-4. The final encapsulation scheme is formed by curing the shielding filler formulation corresponding to the Pareto front, the P-Zeta potential modification parameters, temperature control, shear rate, and pressure timing. This scheme is then fed back to the material genome engineering model in S2 and the reinforcement learning agent in S3 for sample expansion and strategy update in the next iteration.

[0044] Furthermore, once the P-Zeta potential modification parameters are determined, in the kneading process of S3, only closed-loop pH adjustment is allowed for potential fine-tuning, without changing the direction of the adjustment amplitude of the solidified surface potential.

[0045] The result of the potential fine-tuning is written into the state vector of the reinforcement learning agent in real time as an additional dimension to ensure that the target of minimizing potential drift continues to be effective.

[0046] Furthermore, in S4-1, the collected packing distribution image is automatically converted into a packing volume fraction field by the digital twin verification system and then directly used as the input boundary condition for calculating the energy deposition field.

[0047] The volume fraction field is synchronously mapped to the formulation-lifetime mapping matrix during reverse retrieval in S5-2, which is used to quickly screen out candidate encapsulation schemes whose filler distribution does not meet the homogeneity requirement.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1) This invention uses a materials genome engineering model as its core, directly mapping environmental parameters such as radiation type, energy flux, and temperature field to shielding filler formulation and P-Zeta potential modification scheme. This eliminates the lengthy process of repeated sampling in traditional methods, significantly shortens the R&D cycle, and reduces testing costs. At the same time, the model can enable general-purpose chips to quickly obtain high-level radiation resistance without changing the wafer process, achieving a good balance between performance and cost.

[0050] 2) In the manufacturing process, reinforcement learning algorithms receive multi-source data such as temperature, shear, and pressure in real time and perform millisecond-level closed-loop adjustment of the kneading process to ensure the uniformity of the shielding molding compound and the stability of the interface potential, effectively suppressing filler agglomeration and potential drift; the digital twin verification system synchronously iterates the real-time process data with the irradiation-mechanical coupling model, so that the prediction results are continuously aligned with the actual service performance, thereby stably outputting high-quality packaging products.

[0051] 3) In terms of engineering implementation, this method supports the rapid generation and solidification of packaging solutions for different application scenarios on the same platform, forming a reusable and scalable process-performance knowledge base; the supporting blockchain evidence storage mechanism establishes a traceable digital fingerprint for each solution, which facilitates subsequent maintenance, upgrades and standard promotion, and significantly accelerates the large-scale deployment of radiation-resistant memory chips in nuclear energy, aerospace and military equipment. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the dynamic optimization method for radiation-resistant chip shielding and packaging enhancement based on machine learning in this invention.

[0053] Figure 2 This is a flowchart illustrating step 1 in this invention;

[0054] Figure 3 This is a flowchart illustrating step 2 in this invention;

[0055] Figure 4 This is a flowchart illustrating step 3 in this invention;

[0056] Figure 5 This is a flowchart illustrating step 5 in this invention;

[0057] Figure 6 This is a flowchart illustrating step 6 of the present invention. Detailed Implementation

[0058] The technical terms used in this invention are explained below:

[0059] Phy-X / PSD: A publicly accessible online database of nuclear reactions and radiation shielding, used for quickly querying the cross-section, attenuation coefficient, and dose conversion factor of various rays at different energies.

[0060] VESTA: Crystal structure visualization and modeling software that can read first-principles calculation results and generate atomic-level 3D models, providing geometric and electronic structure inputs for subsequent interface simulations.

[0061] Materials Studio: A multi-scale materials simulation platform that can build filler-resin interfaces at the mesoscale, calculate interfacial energy, diffusion coefficient, and compatibility, and provide thermodynamic basis for formulation optimization.

[0062] VASP: A first-principles calculation program based on density functional theory, used to solve for ray-matter interactions, defect formation energies, and electronic structures at the atomic scale, providing high-precision data for model training.

[0063] P-Zeta potential: The surface charge density of filler particles in the liquid phase, which determines the repulsive force between particles and the dispersion stability, and directly affects the uniformity of kneading and the reliability of the final shielding layer.

[0064] AB mixing unit: A continuous twin-screw mixing unit that synchronously delivers resin and modifier in a set ratio to achieve pre-dispersion of fillers and form a uniform premix, laying the foundation for subsequent kneading.

[0065] Non-dominated sorting genetic algorithm: a multi-objective optimization algorithm that finds the Pareto optimal solution between shielding effectiveness and quality increment by simulating the biological evolution process, thus avoiding the bias caused by human weight setting.

[0066] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, circuit structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0067] Example 1

[0068] This invention provides a dynamic optimization method for radiation-resistant chip shielding and packaging enhancement schemes based on machine learning. (See also...) Figure 1 This includes the following steps:

[0069] S1. Obtain target irradiation environment parameters, including radiation type, energy flux, and temperature field distribution. See [link to relevant documentation]. Figure 2 .

[0070] In specific implementation, the process of obtaining target irradiation environment parameters in S1 includes:

[0071] S1-1. Call the Phy-X / PSD database and VESTA / Materials Studio interface to retrieve the ray type-energy flux coupling curve corresponding to the target application scenario;

[0072] S1-2. Using the atomic-level temperature field distribution calculated by VASP as the initial temperature field, the temperature field distribution is mapped to the chip packaging-level mesh through an interpolation algorithm to obtain the temperature field distribution.

[0073] S1-3. Encapsulate the ray type, energy flux, and temperature field distribution into a unified format environmental parameter vector.

[0074] In principle, this step selects radiation type, energy flux, and temperature field distribution as the core parameters of the target irradiation environment because they jointly determine the entire physical-chemical process of irradiation damage within the encapsulation layer: radiation type determines the dominant mechanism of interaction with matter (ionization, displacement, or nuclear reaction); energy flux quantifies the energy density deposited per unit time, directly related to defect yield and cumulative dose; temperature field distribution affects defect migration, recombination, and interfacial reaction rates through thermal activation, thereby altering the microstructural stability of the shielding filler and resin. By obtaining the radiation type-energy flux coupling curve through the Phy-X / PSD database and the VESTA / Materials Studio interface, the most representative irradiation spectrum can be quickly identified at the macroscopic level. Using the atomic-level temperature field output by VASP as the source, it is interpolated and mapped to the encapsulation-level mesh, seamlessly extending microscopic thermodynamic information to the macroscopic scale, solving the problem of cross-scale information fragmentation. Encapsulating all three into a vector preserves the integrity of the physical meaning and provides a directly readable and scalable input format for subsequent materials genome engineering models, ensuring that subsequent formulation design, process optimization, and lifetime prediction are always anchored to real-world service scenarios.

[0075] S2. Based on the environmental parameters, a shielding filler formulation and corresponding P-Zeta potential modification parameters are generated through a materials genome engineering model, wherein the materials genome engineering model is trained by integrating multi-scale physical simulation data.

[0076] In specific implementation, S2, based on the aforementioned environmental parameters, includes the following process for generating the shielding filler formulation and corresponding P-Zeta potential modification parameters through a materials genome engineering model:

[0077] S2-1. Input the environmental parameter vector obtained in S1-3 into the material genome engineering model. The model uses the multi-scale physical simulation dataset output by the shielding filler calculation and design platform as the training set. The dataset includes the radiation absorption cross section, the filler-resin interface energy, the P-Zeta potential, and the dispersion stability correlation curve.

[0078] S2-2. During the training process, a Bayesian optimizer is used to sample the filler composition space and iteratively update the ternary formulation of filler, resin and modifier until the predicted shielding effectiveness index and interface potential index converge simultaneously.

[0079] S2-3. Output the converged shielding filler formulation and the corresponding P-Zeta potential modification parameters. The formulation includes the proportion of special resin for shielding filler, the type and amount of dispersant modifier, and the P-Zeta potential modification parameters include the surface potential adjustment range and pH window range.

[0080] In specific implementation, in S2, the process of acquiring the materials genome engineering model through training by integrating multi-scale physical simulation data includes:

[0081] Using the shielding filler computational design platform as the data source, VESTA is called sequentially to establish the atomic-scale crystal structure of the filler, Materials Studio is used to construct the filler-resin interface model, and VASP is used to calculate the ray-matter interaction cross section, generating multi-scale physical simulation data at the atomic-mesoscopic scale.

[0082] The multi-scale physical simulation data are labeled according to ray type, energy flux, and temperature field distribution to form a training sample library;

[0083] A graph neural network architecture is adopted, with the filler component vector, interface atomic coordinates and P-Zeta potential as input nodes, and shielding effectiveness and dispersion stability as output nodes. The sample library is iteratively trained until the prediction error is less than the preset value, thereby obtaining the material genome engineering model.

[0084] In terms of specific principles, this step starts with environmental parameter vectors and utilizes a materials genome engineering model fully trained with multi-scale physical simulation data to couple ray-matter interactions, interfacial thermodynamics, and electrochemical stability within the same prediction framework. First, atomic-to-mesoscale ray absorption cross-sections, interfacial energies, and P-Zeta potential correlation data are generated sequentially using VESTA, Materials Studio, and VASP, and uniformly encoded with scene tags to form a high-fidelity training sample library covering ray types, energy fluxes, and temperature field distributions. Then, a graph neural network is used to map filler composition, interfacial atomic arrangement, and surface potential to shielding effectiveness and dispersion stability, achieving end-to-end prediction from microstructure to macroscopic performance. In the online design stage, a Bayesian optimizer iteratively samples in a high-dimensional composition space, using the dual convergence of shielding effectiveness and interfacial potential as criteria to quickly lock the optimal ratio of filler-resin-modifier, and simultaneously provides the surface potential control direction and pH window to ensure that the subsequent kneading process obtains a uniform, stable, and highly shielded encapsulation system under a given environmental load.

[0085] S3. Based on the shielding filler formulation and P-Zeta potential modification parameters, the packaging production line performs a kneading process, and the process parameters are adjusted in real time through a reinforcement learning algorithm. The process parameters include temperature control, shear rate, and pressurization sequence.

[0086] In specific implementation, S3, the specific process of performing the kneading process on the encapsulation production line according to the shielding filler formulation and P-Zeta potential modification parameters includes:

[0087] S3-1. Based on the shielding filler formula output by S2-3, the shielding filler special resin and dispersant modifier are fed into the AB mixing unit in a set ratio to form a premix.

[0088] S3-2. Using the P-Zeta potential modification parameter, the surface potential of the premix is ​​controlled in the kneader, the potential value is monitored in real time, and the pH window range is adjusted in a closed loop.

[0089] S3-3. During the kneading process, a reinforcement learning algorithm is used to receive real-time feedback from temperature sensors, shear rate sensors, and pressure sensors, generate dynamic process instructions for temperature control, shear rate, and pressurization timing, and adjust the kneading machine actuator at the millisecond level through the controller until a homogeneous shielding molding compound is obtained.

[0090] In specific implementation, the process of adjusting process parameters in real time through reinforcement learning algorithms in S3 includes:

[0091] The obtained real-time values ​​of P-Zeta potential, instantaneous values ​​of shielding molding compound viscosity, and instantaneous values ​​of filler dispersion are used to construct a state vector, which is then used as the input to the reinforcement learning agent.

[0092] The agent uses temperature control value, shear rate value, and pressurization timing value as its action space. It updates its strategy in the continuous action-state space according to the reward function, which takes maximizing the uniformity of filler dispersion and minimizing potential drift as weighted objectives.

[0093] Within each batch kneading cycle, action commands are output at a millisecond frequency to the kneader temperature control module, shear drive module, and pressurization servo module to form a closed-loop real-time adjustment until the state vector converges to the preset threshold range, thereby completing the dynamic optimization of process parameters.

[0094] In specific implementation, once the P-Zeta potential modification parameters are determined, in the kneading process of S3, only closed-loop pH adjustment is allowed for potential fine-tuning, without changing the direction of the adjustment amplitude of the solidified surface potential.

[0095] The result of the potential fine-tuning is written into the state vector of the reinforcement learning agent in real time as an additional dimension to ensure that the target of minimizing potential drift continues to be effective.

[0096] The core of S3 lies in transforming static formulations into dynamically malleable homogeneous shielding molding compounds. First, based on the shielding filler formulation given in S2-3, the special resin for shielding filler and the dispersant modifier are fed into the AB mixing unit in a predetermined ratio to form a uniform premix, ensuring no deviation in macroscopic composition. Then, P-Zeta potential modification parameters are introduced into the kneader. The potential value is monitored in real time and fine-tuned using closed-loop pH adjustment, which preserves the adjustment direction of the solidified surface potential while suppressing drift and maintaining the stability of the interfacial charge. Most importantly, the reinforcement learning agent encapsulates the P-Zeta potential, viscosity, and filler dispersion into a multi-dimensional state vector in real time. With temperature, shear rate, and pressurization sequence as the continuous action space, it generates and issues action commands to the temperature control, shear, and pressurization modules at millisecond frequencies. With the help of the reward function, it continuously balances the two major objectives of maximizing dispersion uniformity and minimizing potential drift, forming a process-performance closed loop. Each potential fine-tuning result is immediately written back to the state vector, allowing the agent to sense and compensate for new disturbances in the next control cycle. Finally, a homogeneous, stable shielding molding compound that is precisely matched with the subsequent irradiation scenario is output throughout the entire kneading cycle.

[0097] In this embodiment, the agent employs the Proximal Policy Optimization (PPO) algorithm, the core formula of which is as follows:

[0098]

[0099] The parameters are explained as follows:

[0100] State space s:

[0101]

[0102] Real-time P-Zeta potential value (unit: mV), reflecting the surface charge stability of the filler.

[0103] η: Instantaneous viscosity of the shielding molding compound (unit: Pa·s), characterizing flowability and dispersion difficulty.

[0104] D: Instantaneous value of packing dispersion (unit: %), calculated using X-ray imaging, packing distribution variation coefficient. Design rationale: Potential. The viscosity η determines the interfacial bonding strength (affecting the risk of delamination after irradiation), the viscosity η is related to process feasibility, and the dispersion D directly determines the shielding uniformity. These three factors together constitute the core indicators of process stability.

[0105] Action space a:

[0106]

[0107] ΔT: Temperature adjustment value (unit: °C), range ±10 °C

[0108] Shear rate adjustment value (unit: s) -1 ), range ±50s -1

[0109] ΔP: Pressure timing adjustment value (unit: s), range ±0.5s. Design rationale: Temperature affects resin curing and filler movement, shear rate controls the dispersion force field, and pressure timing determines the venting and densification process.

[0110] Advantage function A π (s,a):

[0111] A π (s,a)=Q(s,a)-V(s)

[0112] Where Q(s, a) is the action value function and V(s) is the state value function, calculated using generalized advantage estimation (GAE):

[0113]

[0114] γ: Discount factor (taken as 0.99).

[0115] λ: Bias-variance tradeoff coefficient (taken as 0.95).

[0116] The reward function is designed as follows:

[0117]

[0118] Parameter explanation:

[0119] Dispersion uniformity term CV -1 :

[0120] (Coefficient of variation), σ μ denoted as the standard deviation of the packing distribution, and μ as the mean.

[0121] When this value is maximized, the filler distribution is most uniform (and the shielding effectiveness is optimal).

[0122] ω1: Weighting coefficient (taken as 0.6), reflecting the priority of shielding effectiveness.

[0123] Potential stability term

[0124] S2-3 output target potential value

[0125] This penalty potential drift ensures interface charge stability (preventing aggregation).

[0126] ω2: Weighting coefficient (taken as 0.3),

[0127] Process robustness item e -|ΔT| :

[0128] The exponential decay function penalizes temperature fluctuations.

[0129] ω3: Weighting coefficient (taken as 0.1).

[0130] S4. Collect real-time process data during the packaging process, combine it with the environmental parameters and input it into the digital twin verification system to generate chip irradiation damage distribution and packaging failure lifetime prediction.

[0131] In S4, the specific process of collecting real-time process data during the packaging process, combining it with the environmental parameters, and inputting it into the digital twin verification system to generate chip irradiation damage distribution and packaging failure lifetime prediction includes:

[0132] S4-1. After the shielding molding compound formed in S3 is injected into the mold cavity and the chip is encapsulated in the second stage, the temperature, residual stress and filler distribution images are collected in real time by distributed optical fiber sensors and MEMS temperature-stress sensors to form real-time process data.

[0133] S4-2. Input the real-time process data and the environmental parameter vector output in S1-3 into the digital twin verification system. The system has a built-in irradiation-mechanical coupling model based on VASP calibration and calculates the energy deposition field after the interaction between the ray and the packaging material through a finite element solver.

[0134] S4-3. Based on the energy deposition field results, the failure criterion library output by the materials genome engineering model is called to map the irradiation damage distribution inside the chip, and the Miner cumulative damage criterion is combined to predict the packaging failure lifetime, outputting the chip irradiation damage distribution map and the packaging failure lifetime curve.

[0135] At the principle level, in S4, distributed optical fibers and MEMS temperature-stress sensors converge in real time images of temperature, residual stress, and filler distribution to form a process fingerprint. This fingerprint, together with the environmental parameter vectors of S1-3, is injected into the digital twin verification system. The system is based on the radiation-mechanical coupling model calibrated by VASP. The finite element solver first calculates the energy deposition field of the rays in the package, and then calls the failure criterion library provided by the material genome engineering model to map the energy deposition field into the radiation damage distribution inside the chip. At the same time, the system accumulates the defect evolution according to the Miner cumulative damage criterion, and finally outputs an intuitive radiation damage distribution map and package failure lifetime curve.

[0136] S5. Based on the irradiation damage distribution and encapsulation failure lifetime prediction, determine the encapsulation scheme that meets the shielding requirements of the target scenario.

[0137] In specific implementation, in S5, the process of determining the packaging scheme that meets the shielding requirements of the target scenario based on irradiation damage distribution and packaging failure lifetime prediction includes:

[0138] S5-1. Import the chip irradiation damage distribution map and the packaging failure lifetime curve output by S4-3 into the shielding packaging scheme decision engine. The decision engine uses the total irradiation dose threshold and the minimum service life threshold of the target application scenario as hard constraints.

[0139] S5-2. The decision engine calls the formulation-lifetime mapping matrix in the material genome engineering model to reversely retrieve the shielding filler formulation range, P-Zeta potential modification parameter range and corresponding process parameter range that meet the hard constraint conditions, and generates a set of candidate encapsulation schemes.

[0140] S5-3. Perform multi-objective optimization on the candidate encapsulation scheme set, with the optimization objectives of maximizing shielding effectiveness and minimizing quality increment. Use a non-dominated sorting genetic algorithm to iteratively solve the problem and output the optimal solution of the Pareto front.

[0141] S5-4. The final encapsulation scheme is formed by curing the shielding filler formulation corresponding to the Pareto front, the P-Zeta potential modification parameters, temperature control, shear rate, and pressure timing. This scheme is then fed back to the material genome engineering model in S2 and the reinforcement learning agent in S3 for sample expansion and strategy update in the next iteration.

[0142] In practice, in S4-1, the collected packing distribution image is automatically converted into a packing volume fraction field by the digital twin verification system and then directly used as the input boundary condition for calculating the energy deposition field.

[0143] The volume fraction field is synchronously mapped to the formulation-lifetime mapping matrix during reverse retrieval in S5-2, which is used to quickly screen out candidate encapsulation schemes whose filler distribution does not meet the homogeneity requirement.

[0144] At the principle level, in S5, the irradiation damage distribution map and failure lifetime curve output by S4-3 are first used as input. With the total irradiation dose threshold and minimum service life threshold of the target scenario as hard constraints, the formulation-lifetime mapping matrix of the materials genome engineering model is searched in reverse to quickly locate the shielding filler formulation range, P-Zeta potential modification range, and corresponding temperature, shear, and pressurization process range that simultaneously meet the lifetime and dose windows. Then, a non-dominated sorting genetic algorithm is used to perform multi-objective trade-offs between shielding effectiveness and mass increment to generate the Pareto front optimal solution, so that the final encapsulation scheme can provide sufficient protection without excessive weight increase. After the formulation, potential parameters, and process curves corresponding to the optimal solution are solidified, they are immediately used for production line execution and simultaneously fed back to the materials genome engineering model and reinforcement learning agent to become high-confidence samples for the next round of training, forming a closed loop of continuous evolution. At the same time, the digital twin system converts the packing distribution image collected on-site into a volume fraction field, which is directly used as the boundary condition for energy deposition calculation. It also acts as a pre-screen during the S5-2 reverse retrieval, eliminating any schemes with non-uniform volume fraction fields in advance, ensuring that the candidate schemes entering the optimization cycle always have an achievable uniform microstructure.

[0145] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A dynamic optimization method for radiation-resistant chip shielding and packaging enhancement schemes based on machine learning, characterized in that, Includes the following steps: S1. Obtain target irradiation environment parameters, including radiation type, energy flux, and temperature field distribution; S2. Based on the environmental parameters, a shielding filler formulation and corresponding P-Zeta potential modification parameters are generated through a materials genome engineering model, wherein the materials genome engineering model is obtained by training integrated multi-scale physical simulation data. S3. Based on the shielding filler formulation and P-Zeta potential modification parameters, the packaging production line performs a kneading process, and the process parameters are adjusted in real time through a reinforcement learning algorithm. The process parameters include temperature control, shear rate, and pressurization sequence. S4. Collect real-time process data during the packaging process, combine it with the environmental parameters and input it into the digital twin verification system to generate chip irradiation damage distribution and packaging failure lifetime prediction. S5. Based on the irradiation damage distribution and encapsulation failure lifetime prediction, determine the encapsulation scheme that meets the shielding requirements of the target scenario.

2. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, In S1, the specific process of obtaining the target irradiation environment parameters includes: S1-1. Call the Phy-X / PSD database and VESTA / Materials Studio interface to retrieve the ray type-energy flux coupling curve corresponding to the target application scenario; S1-2. Using the atomic-level temperature field distribution calculated by VASP as the initial temperature field, the temperature field distribution is mapped to the chip packaging-level mesh through an interpolation algorithm to obtain the temperature field distribution. S1-3. Encapsulate the ray type, energy flux, and temperature field distribution into a unified format environmental parameter vector.

3. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, In S2, the specific process of generating the shielding filler formulation and corresponding P-Zeta potential modification parameters based on the environmental parameters using a materials genome engineering model includes: S2-1. Input the environmental parameter vector obtained in S1-3 into the material genome engineering model. The model uses the multi-scale physical simulation dataset output by the shielding filler calculation and design platform as the training set. The dataset includes the radiation absorption cross section, the filler-resin interface energy, the P-Zeta potential, and the dispersion stability correlation curve. S2-2. During the training process, a Bayesian optimizer is used to sample the filler composition space and iteratively update the ternary formulation of filler, resin and modifier until the predicted shielding effectiveness index and interface potential index converge simultaneously. S2-3. Output the converged shielding filler formulation and the corresponding P-Zeta potential modification parameters. The formulation includes the proportion of special resin for shielding filler, the type and amount of dispersant modifier, and the P-Zeta potential modification parameters include the surface potential adjustment range and pH window range.

4. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, In S2, the process of acquiring the materials genome engineering model through training by integrating multi-scale physical simulation data includes: Using the shielding filler computational design platform as the data source, VESTA is called sequentially to establish the atomic-scale crystal structure of the filler, Materials Studio is used to construct the filler-resin interface model, and VASP is used to calculate the ray-matter interaction cross section, generating multi-scale physical simulation data at the atomic-mesoscopic scale. The multi-scale physical simulation data are labeled according to ray type, energy flux, and temperature field distribution to form a training sample library; A graph neural network architecture is adopted, with the filler component vector, interface atomic coordinates and P-Zeta potential as input nodes, and shielding effectiveness and dispersion stability as output nodes. The sample library is iteratively trained until the prediction error is less than the preset value, thereby obtaining the material genome engineering model.

5. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, In S3, the specific process of performing the kneading process on the encapsulation production line according to the shielding filler formulation and P-Zeta potential modification parameters includes: S3-1. Based on the shielding filler formula output by S2-3, the shielding filler special resin and dispersant modifier are fed into the AB mixing unit in a set ratio to form a premix. S3-2. Using the P-Zeta potential modification parameter, the surface potential of the premix is ​​controlled in the kneader, the potential value is monitored in real time, and the pH window range is adjusted in a closed loop. S3-3. During the kneading process, a reinforcement learning algorithm is used to receive real-time feedback from temperature sensors, shear rate sensors, and pressure sensors, generate dynamic process instructions for temperature control, shear rate, and pressurization timing, and adjust the kneading machine actuator at the millisecond level through the controller until a homogeneous shielding molding compound is obtained.

6. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 5, characterized in that, In S3, the specific process of adjusting process parameters in real time using reinforcement learning algorithms includes: The obtained real-time values ​​of P-Zeta potential, instantaneous values ​​of shielding molding compound viscosity, and instantaneous values ​​of filler dispersion are used to construct a state vector, which is then used as the input to the reinforcement learning agent. The agent uses temperature control value, shear rate value, and pressurization timing value as its action space. It updates its strategy in the continuous action-state space according to the reward function, which takes maximizing the uniformity of filler dispersion and minimizing potential drift as weighted objectives. Within each batch kneading cycle, action commands are output at a millisecond frequency to the kneader temperature control module, shear drive module, and pressurization servo module to form a closed-loop real-time adjustment until the state vector converges to the preset threshold range, thereby completing the dynamic optimization of process parameters.

7. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, In S4, the specific process of collecting real-time process data during the packaging process, combining it with the environmental parameters, and inputting it into the digital twin verification system to generate chip irradiation damage distribution and packaging failure lifetime prediction includes: S4-1. After the shielding molding compound formed in S3 is injected into the mold cavity and the chip is encapsulated in the second stage, the temperature, residual stress and filler distribution images are collected in real time by distributed optical fiber sensors and MEMS temperature-stress sensors to form real-time process data. S4-2. Input the real-time process data and the environmental parameter vector output in S1-3 into the digital twin verification system. The system has a built-in irradiation-mechanical coupling model based on VASP calibration and calculates the energy deposition field after the interaction between the ray and the packaging material through a finite element solver. S4-3. Based on the energy deposition field results, the failure criterion library output by the materials genome engineering model is called to map the irradiation damage distribution inside the chip, and the Miner cumulative damage criterion is combined to predict the packaging failure lifetime, outputting the chip irradiation damage distribution map and the packaging failure lifetime curve.

8. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, In S5, the specific process of determining the packaging scheme that meets the shielding requirements of the target scenario based on irradiation damage distribution and packaging failure lifetime prediction includes: S5-1. Import the chip irradiation damage distribution map and the packaging failure lifetime curve output by S4-3 into the shielding packaging scheme decision engine. The decision engine uses the total irradiation dose threshold and the minimum service life threshold of the target application scenario as hard constraints. S5-2. The decision engine calls the formulation-lifetime mapping matrix in the material genome engineering model to reversely retrieve the shielding filler formulation range, P-Zeta potential modification parameter range and corresponding process parameter range that meet the hard constraint conditions, and generates a set of candidate encapsulation schemes. S5-3. Perform multi-objective optimization on the candidate encapsulation scheme set, with the optimization objectives of maximizing shielding effectiveness and minimizing quality increment. Use a non-dominated sorting genetic algorithm to iteratively solve the problem and output the optimal solution of the Pareto front. S5-4. The final encapsulation scheme is formed by curing the shielding filler formulation corresponding to the Pareto front, the P-Zeta potential modification parameters, temperature control, shear rate, and pressure timing. This scheme is then fed back to the material genome engineering model in S2 and the reinforcement learning agent in S3 for sample expansion and strategy update in the next iteration.

9. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 1, characterized in that, Once the P-Zeta potential modification parameters are determined, in the kneading process of S3, only closed-loop pH adjustment is allowed for potential fine-tuning, without changing the direction and magnitude of the already solidified surface potential adjustment. The result of the potential fine-tuning is written into the state vector of the reinforcement learning agent in real time as an additional dimension to ensure that the target of minimizing potential drift continues to be effective.

10. The method for dynamic optimization of radiation-resistant chip shielding and packaging enhancement scheme based on machine learning according to claim 8, characterized in that, In S4-1, the collected packing distribution image is automatically converted into a packing volume fraction field by the digital twin verification system and then directly used as the input boundary condition for the energy deposition field calculation. The volume fraction field is synchronously mapped to the formulation-lifetime mapping matrix during reverse retrieval in S5-2, which is used to quickly screen out candidate encapsulation schemes whose filler distribution does not meet the homogeneity requirement.

Images