Full-biodegradable material experiment system based on multi-scale simulation and machine learning

Abstract

The invention relates to the technical field of material informatics, and discloses a full biodegradable material experiment system based on multi-scale simulation and machine learning, which comprises an intelligent design and life prediction platform, an integrated data and interface layer, a multi-scale simulation layer, a machine learning fusion layer and an application layer. The multi-scale simulation layer calculates interface adsorption energy, a dispersion uniformity index and macrophysical field characteristics; and the machine learning fusion layer splices the physical characteristics and the process parameters, and constructs a performance and life prediction model by using transfer learning. The forward engine executes mapping from a formula to performance, and the reverse engine performs multi-target reverse optimization based on a genetic algorithm. According to the method, the physical mechanism feature is introduced to enhance the machine learning model, and the active learning mechanism and the long-short-term memory network are combined, so that the dynamic prediction of the degradation life of the full-biodegradable material and the closed-loop design of the formula are realized, the experiment cost is effectively reduced, and the research and development period is shortened.

Description

Technical Field

[0001] This invention relates to the field of materials informatics technology, specifically to an experimental system for fully biodegradable materials based on multi-scale simulation and machine learning. Background Technology

[0002] The development of fully biodegradable composite materials involves multivariate coupling, multi-scale correlation, and multi-objective synergistic optimization. With increasingly stringent environmental standards, developing high-performance composite materials that combine excellent mechanical properties with controllable degradation lifespan is a current technological trend. However, traditional materials research and development mainly relies on trial and error, where engineers repeatedly adjust formulations and processes based on experience and conduct experimental verification. Since the performance of fully biodegradable materials is influenced by multiple variables such as polymer matrix structure, filler type, interfacial compatibility, and processing technology, this screening model, which relies on numerous physical experiments, results in long development cycles, high experimental costs, and makes it difficult to determine the optimal solution within a vast parameter space.

[0003] In the prediction and mechanistic analysis of material properties, existing technologies struggle to establish an effective correlation between microscopic mechanisms and macroscopic properties. The macroscopic mechanical properties and degradation behavior of materials originate from interfacial interactions at the atomic scale and phase structure evolution at the mesoscopic scale, but current single-scale simulation methods cannot directly map cross-scale characteristics from micro to macro. On the other hand, while data-driven machine learning models have advantages in handling nonlinear relationships, they typically lack the support of physical mechanisms, resulting in insufficient generalization ability when faced with new formulations or processes outside the training data range, and difficulty in explaining the intrinsic causes of performance changes. Furthermore, traditional pure physics models often simplify multiphase composite systems to reduce computational complexity, making it difficult to accurately describe the dynamic behavior of complex real-world systems.

[0004] Furthermore, the key difference between fully biodegradable materials and traditional plastics lies in the controllability of their degradation behavior. Current lifespan assessment methods primarily rely on standard accelerated aging tests under laboratory conditions, which differ from the complex and variable real-world service environments. Existing technologies lack cross-scale lifespan prediction tools that can effectively couple environmental temporal characteristics such as light, temperature, and humidity, making it difficult to accurately predict the degradation process of materials in actual natural environments. This leads to premature failure or delayed degradation in practical applications. Simultaneously, the lack of a unified data management system results in the fragmented storage of multi-source data accumulated from historical experiments and expert experience, failing to form reusable and evolving data assets, thus hindering the improvement of new material research and development efficiency. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an experimental system for fully biodegradable materials based on multi-scale simulation and machine learning. This system solves the problems of long development cycles due to the reliance on numerous physical experiments in traditional fully biodegradable material development, weak generalization ability due to the lack of physical mechanism support in purely data-driven models, and difficulty in accurately predicting the degradation life of materials under complex service environments.

[0006] To achieve the above objectives, this invention provides the following technical solution: a fully biodegradable material experimental system based on multi-scale simulation and machine learning. This system includes a data and interface layer, a multi-scale simulation layer, a machine learning fusion layer, an application and interaction layer, and a forward prediction engine and a reverse design engine to drive the computational process. The platform's underlying layer is configured with a material gene library for storing basic physicochemical data, a process knowledge base for storing historical experimental rules and expert experience, and a model library for storing trained algorithms.

[0007] The data and interface layer is responsible for the input and preprocessing of multi-source data. It acquires experimental test, production process, field environment and literature data through the acquisition module, and uses the standardization module to clean and unify the format of heterogeneous data.

[0008] The multi-scale simulation layer is the core unit for generating physical mechanism features, covering three dimensions: microscopic, mesoscopic, and macroscopic.

[0009] Among them, the microscale simulation engine constructs an interface composite model and determines the interface adsorption energy by calculating the difference between the total potential energy of the composite system after equilibrium conformation and the sum of the independent energies of each component; the mesoscale simulation engine adopts a coarse-grained modeling strategy, divides the system into equal-volume three-dimensional grids and counts the local volume fraction of the target phase, and calculates the dispersion uniformity index based on the square root of the average of the sum of the squares of the differences between the volume fractions of all grid regions and the average volume fraction; the macroscale simulation engine integrates finite element analysis and computational fluid dynamics units, solves the mechanical equilibrium equations and fluid control equations respectively, and extracts the maximum von Mises stress, stress gradient distribution variance, flow field shear heat generation peak and average residence time as macroscopic physical field characteristics.

[0010] After obtaining the physical features, the machine learning fusion layer receives the original process formula parameters and the aforementioned physical simulation features through the feature fusion module.

[0011] This module standardizes each component and concatenates the physical simulation features as a supplementary dimension into the original parameter space to generate an enhanced feature vector. Based on this enhanced feature vector, the system constructs a dedicated machine learning model cluster that includes performance prediction and lifetime prediction. This model cluster employs a transfer learning strategy that combines simulation pre-training with experimental fine-tuning. Among them, the performance prediction model uses gradient boosting decision trees or deep neural network architectures to predict mechanical and thermal properties; the lifetime prediction model uses long short-term memory networks and uses forget gate, input gate and output gate structures to process the cumulative effect of external environmental time-series data and material internal characteristics, thereby calculating the percentage of remaining weight of the material and the retention rate of mechanical properties over time.

[0012] The system's computation and optimization process is implemented through a bidirectional engine: the forward prediction engine sequentially calls the multi-scale simulation layer and the machine learning fusion layer to perform the mapping calculation from design variables to performance indicators; the reverse design engine handles the reverse optimization from performance requirements to design variables, constructs a system objective function vector that includes cost, mechanical performance and degradation cycle deviation, uses a reference point-based non-dominated sorting genetic algorithm to generate candidate solutions, and calls the forward prediction engine to calculate fitness and screen the Pareto optimal solution set.

[0013] To enhance the robustness and interactivity of the system, this invention also integrates active learning, interpretability analysis, and visual interaction mechanisms.

[0014] When the model's prediction error is large or it is in a high-uncertainty region, the system uses a probabilistic model to quantify the uncertainty. It evaluates the sample value by calculating the sum of the predicted mean and weighted standard deviation, and recommends additional experimental points. Simultaneously, it uses an algorithm based on cooperative game theory to calculate the marginal contribution of feature components, quantifying the relative importance of the original parameters and physical features. At the application level, a reverse engineering workbench is used to display optimization schemes, a virtual experiment system graphically simulates the evolution of material properties, and a knowledge graph visualization interface intuitively displays the topological relationship between material structure and properties.

[0015] This invention provides an experimental system for fully biodegradable materials based on multi-scale simulation and machine learning. It offers the following advantages: 1. This invention improves the generalization ability and accuracy of material property prediction by integrating multi-scale simulation and machine learning techniques. The system utilizes microscopic interface adsorption energy, mesoscopic dispersion uniformity index, and macroscopic physical field characteristics to construct enhanced feature vectors, introducing explicit physical mechanisms into the data-driven model. This method effectively overcomes the shortcomings of traditional pure data-driven models, such as overfitting and lack of interpretability with small sample experimental data, ensuring that the model maintains high prediction accuracy when dealing with unseen formulation combinations.

[0016] 2. This invention utilizes active learning mechanisms and transfer learning strategies to reduce the reliance on physical experiments in the development of fully biodegradable materials. By pre-training and initializing model parameters using simulated data, and employing a data acquisition function to select high-value sample points for targeted experimental verification, the system achieves a research and development model that replaces physical experiments with virtual ones. This not only reduces the number of trial-and-error experiments and raw material consumption but also effectively shortens the research and development cycle from formulation design to product launch.

[0017] 3. This invention achieves dynamic prediction of material lifetime and reverse design of formulations under service environment coupling. By introducing a long short-term memory network to process environmental time-series data, the system can accurately simulate the cumulative degradation behavior of materials in complex natural environments, solving the pain point of difficulty in quantifying degradation lifetime. Combined with a reverse design engine based on a non-dominated sorting genetic algorithm, the system can deduce the optimal process formulation based on predetermined lifetime and mechanical performance targets, providing a closed-loop solution for the customized development of fully biodegradable materials. Attached Figure Description

[0018] Figure 1 This is a general framework diagram of the intelligent design and life prediction platform of the present invention; Figure 2 This is a schematic diagram illustrating the multi-scale simulation and feature extraction principle of the present invention; Figure 3 This is a diagram of the physical feature fusion and machine learning prediction model architecture of the present invention; Figure 4 This is a flowchart of the multi-objective reverse optimization based on the non-dominated sorting genetic algorithm of the present invention; Figure 5 This is a flowchart of the intelligent design method of the present invention; Figure 6 This is a comparison chart of the performance prediction accuracy of the present invention by fusing multi-scale features; Figure 7 This is an interface diagram of the reverse design case of the early spring weather-resistant mulch film of the present invention; Figure 8 This diagram illustrates the effectiveness of the active learning mechanism of the present invention.

[0019] The system comprises the following components: 10. Data and Interface Layer; 20. Multi-scale Simulation Layer; 21. Microscale Simulation Engine; 22. Mesoscale Simulation Engine; 23. Macroscale Simulation Engine; 30. Machine Learning Fusion Layer; 31. Feature Fusion Module; 32. Machine Learning Model Cluster; 33. Interpretability Analysis Module; 40. Application and Interaction Layer; 41. Reverse Engineering Workbench; 42. Virtual Experiment System; 43. Knowledge Graph Visualization Interface; 51. Materials Genome Library; 52. Process Knowledge Base; 53. Model Library; 61. Forward Prediction Engine; 62. Reverse Engineering Engine. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see the appendix Figure 1 This invention provides an experimental system for fully biodegradable materials based on multi-scale simulation and machine learning, including an intelligent design and lifespan prediction platform. The intelligent design and lifespan prediction platform logically comprises: a data and interface layer 10, a multi-scale simulation layer 20, a machine learning fusion layer 30, and an application and interaction layer 40. Furthermore, the intelligent design and lifespan prediction platform is also configured with a material gene library 51, a process knowledge base 52, and a model library 53 for data storage and retrieval, as well as a forward prediction engine 61 and a reverse design engine 62 for driving the computational process. The data and interface layer 10 is located at the bottom layer and is responsible for data input and preprocessing; the multi-scale simulation layer 20 is used to generate physical mechanism features; the machine learning fusion layer 30 is used to construct prediction models; and the application and interaction layer 40 provides an operating interface for users.

[0022] The data and interface layer 10 is equipped with a data acquisition module and a data standardization module. The data acquisition module is configured to connect to external devices or databases via a standard interface to collect experimental test data, production line process data, field service environment data, and publicly available literature data. Experimental test data includes tensile strength, elongation at break, and thermal performance indicators; production line process data includes extrusion temperature, screw speed, and traction speed; field service environment data includes soil temperature, humidity, and microbial activity indicators. The data standardization module cleans, denoises, and standardizes the format of the collected heterogeneous data before transmitting the processed data to the database or subsequent computation layers.

[0023] The multi-scale simulation layer 20 encompasses physical simulation capabilities from the microscopic to the macroscopic level, specifically including a microscale simulation engine 21, a mesoscale simulation engine 22, and a macroscale simulation engine 23. The microscale simulation engine 21 integrates first-principles calculation units based on density functional theory and molecular dynamics simulation units to calculate interfacial adsorption energies, electronic structures, and molecular chain conformational characteristics between material components. The mesoscale simulation engine 22 integrates dissipative particle dynamics calculation units and phase-field method calculation units to simulate phase separation behavior, filler dispersion states, and crystallization kinetics in multiphase systems, outputting dispersion uniformity indices and mesoscopic characteristics of phase region sizes. The macroscale simulation engine 23 integrates finite element analysis units and computational fluid dynamics units to simulate stress-strain distribution of materials under macroscopic external forces and rheological behavior during processing.

[0024] The machine learning fusion layer 30 is the core processing layer for achieving high-precision prediction, comprising a feature fusion module 31 and a machine learning model cluster 32. The feature fusion module 31 receives raw formulation process features from the data and interface layer 10, and physical features such as interfacial binding energy and filler dispersion index from the multi-scale simulation layer 20, concatenating and mapping them into a high-dimensional enhanced feature vector. The machine learning model cluster 32 includes a performance prediction model, a process optimization model, and a lifetime prediction model. The performance prediction model uses a gradient boosting decision tree or deep neural network architecture to predict the mechanical and thermal properties of the material based on the enhanced feature vector; the lifetime prediction model uses a temporal neural network architecture to predict the degradation curve of the material by combining environmental time-series data; the process optimization model is based on a reinforcement learning algorithm to search for the optimal combination of process parameters under a given performance objective.

[0025] The application and interaction layer 40 provides users with a visual operating terminal, including a reverse design workbench 41, a virtual experiment system 42, and a knowledge graph visualization interface 43. The reverse design workbench 41 allows users to input specific performance indicators, cost constraints, and degradation cycle requirements, and displays recommended formulations calculated by the system. The virtual experiment system 42 displays the microstructure evolution and performance prediction curves during the simulation process in graphical form. The knowledge graph visualization interface 43 is used to display the network of relationships between material composition, structure, properties, and processes.

[0026] The Materials Genome Base (51) stores digital fingerprint information of basic components, including molecular structure data and basic physical property parameters of polymer matrices, plant fibers, and functional fillers. The Process Knowledge Base (52) stores historical experimental data on process, structure, and performance correlation rules, as well as expert experience data. The Model Library (53) stores trained machine learning model files, surrogate model parameters, and corresponding hyperparameter configuration files, supporting model version management and online access.

[0027] The forward prediction engine 61 is configured to perform mapping calculations from design variables to performance indicators. When a set of defined material formulations and process parameters are input, the forward prediction engine 61 sequentially calls the multi-scale simulation layer 20 to generate physical features, and calls the performance prediction model and lifetime prediction model in the machine learning fusion layer 30 to output the performance prediction value and degradation lifetime prediction value of the scheme.

[0028] The reverse design engine 62 is configured to perform reverse optimization from performance requirements to design variables. This engine employs a multi-objective optimization algorithm, performing a global search within the composition space defined by the material gene library 51 and a preset process window. During the search, the reverse design engine 62 generates a group of candidate solutions and calls the forward prediction engine 61 to evaluate the fitness of each candidate solution, selecting the Pareto optimal solution set that satisfies multiple constraints, and providing feedback to the user through the application and interaction layer 40.

[0029] Please see the appendix Figure 5 This invention provides a smart design and lifetime prediction method for fully biodegradable materials, which is executed based on the aforementioned smart design and lifetime prediction platform. This method combines data-driven machine learning with physical mechanism-driven multi-scale simulation and introduces an active learning closed loop into the design process, achieving full-process digitalization from requirement definition to formulation implementation. The method mainly includes the following steps: S1. Execution Requirements Definition and Formalization. The application and interaction layer 40 receives user input regarding end-use indicators for fully biodegradable materials. These indicators typically include lower limits of mechanical properties (e.g., tensile strength, elongation at break), degradation lifespan range (e.g., full biodegradation time), upper limits of cost control, and processing limitations. The computational module within the intelligent design and lifespan prediction platform transforms the engineering requirements described in natural language into a multi-objective optimization problem in mathematical form. This process establishes a mathematical model containing an objective function vector and a set of constraints. The objective function vector encompasses the cost function, mechanical performance function, and degradation cycle deviation function, while the set of constraints includes component normalization constraints and physical boundary constraints of process parameters.

[0030] S2. Execution of Design Space Exploration and Initial Screening. The forward prediction engine 61 calls upon the component data in the material gene library 51 and delineates the initial boundary of the multidimensional design space based on the preset types of matrix materials, reinforcing fibers, and functional fillers. Within this initial boundary, the system generates sparse initial sample points using Latin hypercube sampling or random sampling algorithms. For these initial sample points, the forward prediction engine 61 runs a rapid evaluation process, predicting the performance of each sample point by calling low-precision surrogate models or historical data interpolation. Based on the prediction results, the system eliminates spatial regions that clearly do not meet the constraints (such as cost overruns or low performance), thereby outputting a reduced effective design space.

[0031] S3. Perform multi-objective reverse optimization. The reverse design engine 62 initiates a global optimization algorithm within the effective design space, preferably employing a non-dominated sorting genetic algorithm. During each generation of the optimization iteration, the reverse design engine 62 generates candidate schemes containing specific formulation ratios and process parameters. For each candidate scheme, the system automatically triggers the computation task of the multi-scale simulation layer 20. The microscale simulation engine 21 calculates the interfacial adsorption energy characteristics of the formulation system, and the mesoscale simulation engine 22 calculates the filler dispersion uniformity index characteristics of the system.

[0032] Subsequently, the feature fusion module 31 concatenates the simulated physical features with the original formulation and process parameters of the candidate solutions to construct an enhanced feature vector. The performance prediction model and lifetime prediction model in the machine learning fusion layer 30 receive this enhanced feature vector and output the predicted mechanical properties and degradation lifetime of the candidate solution. The reverse design engine 62 evaluates the fitness of the candidate solution based on the prediction results and sorts and selects the population according to the Pareto dominance relationship. This process iterates until a preset number of iterations or convergence criteria are reached, and finally outputs a set of non-dominated solutions located on the Pareto front.

[0033] S4. Implementation Scheme Recommendation and Interpretation. The intelligent design and life prediction platform post-processes the non-dominated solution set output in step S3. The system calculates the congestion distance and comprehensive score for each scheme in the solution set, sorts them according to the user's preference weights for each performance indicator, and selects the top-ranked recommended schemes. Simultaneously, the application and interaction layer 40 calls the interpretability analysis module 33 to calculate the contribution of each design variable and physical feature in the recommended scheme to the final performance prediction value using the SHAP method. The system generates a virtual experiment report containing the recommended formula, predicted performance curve, process parameter suggestions, and feature contribution analysis.

[0034] S5. Execute Experimental Verification and Active Learning Feedback. Users conduct physical prototyping and performance testing based on the recommended scheme in the virtual experiment report. The data and interface layer 10 collects the performance and degradation data obtained from the actual tests and feeds them back to the intelligent design and lifetime prediction platform. The intelligent design and lifetime prediction platform calculates the error between the predicted and measured values. If the error exceeds a preset threshold, or if the current scheme is in a region with high model prediction uncertainty, the system activates the active learning mechanism. This mechanism searches for the region with the highest information gain in the design space based on the acquisition function, generating supplementary experimental suggestion points for the next round, thereby improving the model's coverage of unknown regions by minimizing the number of experiments.

[0035] S6. Knowledge Accumulation and Model Evolution. When the verification results meet the design requirements, the intelligent design and life prediction platform encapsulates the successful formula, process parameters, microstructural characteristics, and corresponding performance data into standard data entries and stores them in the process knowledge base 52. Simultaneously, the intelligent design and life prediction platform triggers a model update command, using the newly stored data to perform incremental training or full retraining on the machine learning model cluster 32 in the model library 53. By updating the model weight parameters, the intelligent design and life prediction platform internalizes the newly acquired physical laws into algorithmic capabilities, achieving continuous evolution of prediction accuracy.

[0036] Please see the appendix Figure 2 In this invention, the microscale simulation engine 21, as a fundamental component of the multiscale simulation layer 20, is primarily responsible for revealing the interaction mechanisms between material components at the atomic and molecular levels. Specifically, this engine is configured to perform quantum mechanical calculations based on density functional theory or molecular dynamics simulations based on force fields. For fully biodegradable composite material systems, the core task of microscale simulation is to quantify the interfacial bonding ability between the matrix and filler, as well as the microscopic motion behavior of polymer chain segments, thereby generating high-level physical features that can be understood by machine learning models.

[0037] In the specific implementation process, the microscale simulation engine 21 first constructs an all-atom model. For polymer matrices, such as poly(butylene adipate / terephthalate), polylactic acid, or polypropylene carbonate and their copolymers, the system generates amorphous unit cells containing segments with specific degrees of polymerization using the Monte Carlo method, and performs energy minimization and pre-equilibrium under periodic boundary conditions. For reinforcing fillers, such as thermoplastic starch, cellulose nanocrystals, or calcium carbonate powder, the system constructs corresponding crystal surface models or nanocluster models. To simulate the two-phase interface, the system splices the matrix unit cell with the filler model to construct an interface composite model. The construction methods of the above molecular models, the allocation of force field parameters, and the geometric optimization process can be implemented by those skilled in the art using existing material simulation software. The specific algorithm implementation process is well-known in the field and will not be elaborated here.

[0038] The microscale simulation engine 21 calculates the interfacial adsorption energy based on a geometrically optimized interfacial composite model. This physical quantity is a key microscopic descriptor for evaluating the compatibility and macroscopic mechanical strength of multiphase materials. (Interfacial adsorption energy) The specific calculation formula is defined as follows: ; in, It represents the interfacial adsorption energy, and its value is usually negative. The larger the absolute value, the stronger the interfacial bonding effect, and the less likely the material is to undergo interfacial delamination when subjected to force. It represents the total potential energy of the composite system formed by the matrix and filler after reaching an equilibrium conformation; Represents the independent energy of the polymer matrix component in the composite system; This represents the independent energy of the filler components in the composite system. The calculation process considers not only van der Waals forces but also electrostatic interactions and hydrogen bond contributions. Once extracted, this eigenvalue is directly used as a numerical indicator reflecting component compatibility, providing a clear physical basis for subsequent machine learning models to predict tensile strength and elongation at break.

[0039] Furthermore, the microscale simulation engine 21 is also configured to perform long-term molecular dynamics simulations to calculate the mean square displacement and extract segment mobility characteristics. Under set temperature and pressure conditions, the system records the displacement trajectories of the polymer segment centroids or key atoms over time. By analyzing the slope of the mean square displacement over time, the self-diffusion coefficient of the segment can be derived. This parameter reflects the flexibility of the polymer chain and the free volume distribution within the system. This microdynamic characteristic is intrinsically correlated with the glass transition temperature of the material and directly affects the transport and diffusion rates of water molecules, oxygen, or degrading enzymes within the material matrix. Therefore, this segment mobility characteristic is used as an important physical input feature for predicting the material's biodegradation rate and aging resistance, and is fed into the feature fusion module 31.

[0040] Building upon microscopic simulations, the mesoscopic-scale simulation engine 22 further extends the simulation horizon to the micrometer level to analyze the phase morphology evolution of multiphase composite systems and the spatial distribution of fillers within the polymer matrix. This engine is configured to perform dissipative particle dynamics or phase field method calculations. For complex systems composed of different types of fully biodegradable polymers and inorganic fillers, mesoscopic simulations primarily address the challenges of quantifying the degree of microscopic phase separation and filler dispersion quality, which are difficult to quantify using traditional experiments.

[0041] During execution, the mesoscale simulation engine 22 employs a coarse-grained modeling strategy, simplifying polymer segments and filler particles into beads that possess not only mass but also interaction potentials. The system maps these thermodynamic parameters to conservative force repulsion parameters in the DPD simulation based on solubility parameters obtained from microscopic simulation calculations or Flory-Huggins interaction parameters obtained through the group contribution method. The specific algorithms and parameter conversion rules for this coarse-grained mapping are well-known techniques in this field and will not be elaborated upon here. Within a simulation box with periodic boundary conditions, the system simulates the trajectories of each component bead under Newton's laws of motion and random and dissipative forces until the system energy converges, forming a stable mesoscale phase structure.

[0042] To transform the intuitive morphological images obtained from simulation into numerical features that can be processed by machine learning models, the mesoscale simulation engine 22 incorporates a dispersion state quantization algorithm. This algorithm calculates a dispersion uniformity index using a gridded statistical method, serving as a key mesoscale descriptor for evaluating the homogeneity of composite materials. Specifically, the system divides the three-dimensional simulation box into... A three-dimensional grid of equal volume is constructed, and the local volume fraction of the target phase (such as filler particles or dispersed phase droplets) within each grid is statistically analyzed. Based on the statistical results, a dispersion uniformity index is calculated. The calculation formula is as follows: ; in, This represents the dispersion uniformity index, and its value typically ranges from 0 to 1. This represents the total number of grids into which the simulation system is divided. The number of grids is usually related to the feature length scale. Indicates the first Local volume fraction of filler or dispersed phase within each grid; It represents the dispersion uniformity index of the filler or local volume fraction within the entire simulation system; This represents the variance of the local volume fraction.

[0043] The square root term in the formula essentially calculates the standard deviation of the local volume fraction. When severe packing agglomeration occurs, the standard deviation of the volume fraction within a portion of the grid... It will be higher than the dispersion uniformity index. The remaining grid cells were below average, leading to an increase in the standard deviation. The value decreases; conversely, when the packing material is evenly dispersed at a certain height, the value within each grid increases. Approaching When the standard deviation approaches 0, at this point... The value approaches 1. This feature can effectively characterize whether there are stress concentration points or degradation weak points inside the material. The mesoscale simulation engine 22 will calculate the... The values ​​and average size features of the phase regions are transmitted to the feature fusion module 31 to assist in predicting the elongation at break and the disintegration behavior during the degradation process of the material, thereby supporting the technical solution in the claims regarding performance prediction based on mesoscopic dispersion features.

[0044] The macroscale simulation engine 23, as the top-level component of the multi-scale simulation layer 20, is mainly used to solve the problems of predicting the mechanical response of materials at the macroscopic solid level and analyzing the processing flow field. This engine integrates finite element analysis units and computational fluid dynamics units, aiming to establish a continuum mechanical relationship between the microstructure and macroscopic properties of materials, and to simulate the impact of actual processing technology on material properties.

[0045] In the simulation of macroscopic mechanical properties, the finite element analysis unit is configured to construct a geometric model of a representative volume element or standard test spline of the fully biodegradable material based on continuum mechanics theory. The system calls upon the constitutive equation library and sets the corresponding stress and strain relationship parameters according to the viscoelastic-plastic characteristics of the material. These constitutive parameters can be derived from fitting experimental data or transferred from the elastic constants obtained from the aforementioned microscopic and mesoscopic simulations. The finite element analysis unit simulates the deformation and failure process of the material under different working conditions by applying virtual tensile, bending, or impact loads to the geometric model. The system solves the mechanical equilibrium equations and outputs stress distribution contour maps, strain fields, and stress-strain curves throughout the process.

[0046] To transform macroscopic simulation results into features usable by machine learning models, this unit is specifically configured with a feature extraction algorithm to quantify stress distribution field characteristics. The system automatically identifies stress concentration regions in the simulation results and calculates the distribution variance of the maximum von Mises stress and stress gradient. These field feature data reflect the weakening effect of internal material defects on macroscopic strength and are transmitted as correction factors to the machine learning fusion layer 30 to calibrate the bias of the purely data-driven model in predicting complex nonlinear mechanical behaviors. The specific configuration of finite element mesh generation, boundary condition settings, and solver selection can be tailored to specific needs by those skilled in the art, and is well-known in the field, so it will not be elaborated further here.

[0047] In terms of process simulation, the computational fluid dynamics unit is configured to simulate the melt flow behavior of fully biodegradable materials in the extruder screw, die channel, and blown film processes. Considering that fully biodegradable materials are typically thermosensitive and their rheological behavior is highly sensitive to temperature and shear rate, this unit focuses on simulating the thermal and shear histories during processing. The system constructs a fluid domain model of the flow channel and sets the material viscosity parameters based on a non-Newtonian fluid rheological model.

[0048] By numerically solving the Navier-Stokes equations and energy equations, the system calculates the velocity distribution, pressure drop, temperature distribution, and shear rate distribution within the output flow field of the fluid dynamics unit. Specifically, the system calculates the residence time distribution of the melt within the flow channel, as excessively long residence times can lead to thermal degradation of biodegradable materials during processing, thereby reducing the mechanical properties of the final product. This unit extracts the peak shear heat generation and average residence time as key process features. These features, along with original process parameters such as extrusion temperature and screw speed, constitute the input to the machine learning model, enabling the platform to assess the impact of process conditions on the potential degradation risk of the material, thus supporting the technical solution in the claims regarding process optimization based on processing flow field simulation.

[0049] Please see the appendix Figure 3In this invention, the machine learning fusion layer 30 achieves deep integration of physical mechanism data and experimental statistical data through the feature fusion module 31, thereby constructing a high-precision hybrid-driven prediction model. The core function of this module is to construct an input vector containing multi-source information, solving the technical problem of low prediction accuracy and weak generalization ability of a single data source under small sample conditions.

[0050] The feature fusion module 31 is configured to receive raw process recipe parameters from the data and interface layer 10, and physical simulation features from the multi-scale simulation layer 20. The system defines the input vector for the feature fusion module 31 for model training and prediction. From the original process formula parameters Features of physical simulation Together they form a composition. Among them, the original process formulation parameters... It covers directly controllable design variables, specifically including the mass percentage of each component (such as polymer matrix ratio, reinforcing filler ratio, and functional additive ratio) and process settings during processing (such as extruder temperature in each zone, screw speed, and traction ratio). Physical simulation characteristics. This encompasses the mechanistic descriptors obtained through the aforementioned simulations, specifically including interfacial adsorption energies at the microscopic level. Dispersion uniformity index at the mesoscopic level And the characteristic values ​​of stress field distribution at the macroscopic level.

[0051] In the process of constructing the enhanced feature vector, the feature fusion module 31 first performs... and The components in the model are standardized to eliminate the order-of-magnitude differences between different physical dimensions. Then, the module performs a feature concatenation operation, embedding the standardized physical simulation features as a supplementary dimension into the original parameter space to form a high-dimensional enhanced feature vector. This enhanced feature vector not only includes the formula and process but also the internal microstructure evolution and interactions, thus providing physical constraints for subsequent nonlinear mapping.

[0052] Based on the enhanced feature vectors mentioned above, the machine learning fusion layer 30 constructs a nonlinear mapping relationship. This mapping relationship aims to establish the input vector. With target performance indicators The system establishes mathematical relationships between parameters such as tensile strength, elongation at break, modulus, and biodegradability. It employs gradient boosting decision trees or deep neural networks as the core algorithm to construct the prediction model. During the training phase, the model learns feature weights by minimizing the prediction loss function, since the input includes physically simulated features. The model can more easily capture the intrinsic laws governing the impact of component changes on performance, for example, by identifying interfacial adsorption energies. The positive correlation between tensile strength and the experimental properties is used to improve the prediction confidence for unseen formulations. This fusion strategy enables the present invention to maintain high prediction accuracy and robustness even in the context of small sample data, which is often encountered in the development of fully biodegradable materials.

[0053] In this invention, the machine learning model cluster 32 is configured to construct prediction models for static mechanical properties and lifetime prediction models for dynamic degradation behavior, respectively, depending on the application scenario. To achieve high-precision predictions under limited experimental data conditions in the development of fully biodegradable materials, the cluster employs a modeling strategy combining ensemble learning and deep learning, and introduces a specific training process for small sample data.

[0054] For predicting static performance indicators such as tensile strength, elongation at break, and modulus, the machine learning model cluster 32 preferably uses a gradient boosting decision tree algorithm to construct a regression model, such as XGBoost or LightGBM architecture. This model integrates multiple decision trees and iteratively fits the residuals of the previous prediction. In each iteration, the algorithm automatically selects the optimal splitting features and splitting points based on the feature splitting gain. Since the aforementioned feature fusion module 31 has already incorporated physical simulation features (such as interfacial adsorption energy)... As part of the input features, the GBDT model can explicitly utilize these physical quantities during tree splitting, thereby capturing the nonlinear correlation between material microstructure and macroscopic properties without brute-force learning through large-scale data. Furthermore, to handle high-dimensional nonlinear mappings, the system is configured with a deep neural network architecture, including an input layer, multiple fully connected hidden layers, and an output layer. The hidden layers use rectified linear units as activation functions to introduce nonlinear factors, while the output layer uses a linear activation function to output continuous performance predictions.

[0055] For predicting the degradation lifetime of fully biodegradable materials in natural environments (such as soil and compost), a machine learning model cluster 32 constructed a time-series prediction model based on a long short-term memory network. This model was designed to handle a mixture of static material features and dynamic environmental data as input. Specifically, the model's input receives two parts of data: one part is a fixed-dimensional material-enhanced feature vector generated by the feature fusion module 31, representing the material's inherent anti-degradation potential; the other part is time-series data of the external environment, including soil temperature sequences, humidity change sequences, and microbial activity index sequences. The LSTM unit, through its internal forget gate, input gate, and output gate structure, can remember and process the cumulative environmental effects over long time spans, thereby simulating the hysteresis effect and cumulative damage during the degradation process. The model's output generates a time-varying curve of the material's remaining weight percentage or mechanical property retention rate, thus calculating the specific time required to reach the degradation endpoint (e.g., a weight loss rate of 90%).

[0056] To address the risk of model overfitting due to the small experimental sample size (typically less than a thousand) in the development of fully biodegradable materials, the machine learning model cluster 32 employs a transfer learning strategy of simulated pre-training and experimental fine-tuning. In the pre-training phase, the system utilizes tens of thousands of sets of high-throughput simulated data (including virtual formulations and their corresponding physical calculation results) generated by the multi-scale simulation layer 20 to initially train the deep neural network, initializing the network's weight parameters and enabling it to learn basic physical response laws. In the fine-tuning phase, the system uses a small amount of actually collected high-precision experimental data to update the top-level parameters of the pre-trained model while freezing the bottom-level parameters. Furthermore, during training, the system uses mean squared error as the loss function, combined with L2 regularization and Dropout techniques to suppress model overfitting and ensure the model's generalization ability in the unknown design space. For determining model hyperparameters (such as learning rate, tree depth, and number of hidden layer nodes), the system integrates a Bayesian optimization algorithm for automatic optimization, seeking the globally optimal combination of hyperparameters. The specific optimization process utilizes Gaussian process regression to construct a surrogate model of the objective function, a conventional technique in this field, and will not be elaborated further here.

[0057] This invention includes an interpretability analysis module 33 (as a logical component of the machine learning fusion layer 30) configured in the machine learning fusion layer 30 for attribution analysis of the prediction results of the machine learning model cluster 32. This module aims to address the black-box problem inherent in deep neural networks and ensemble learning models, transforming complex nonlinear mapping relationships into human-understandable physical causal logic. The interpretability analysis module 33 employs the SHAP method, based on cooperative game theory, to calculate the marginal contribution of each feature component in the enhanced feature vector to the final prediction result.

[0058] Specifically, the interpretability analysis module 33 is configured to receive a trained prediction model and a specific sample input vector to be interpreted. For each feature in the input vector (e.g., a specific filler content or calculated interfacial adsorption energy), this module uses an algorithm to quantify the specific contribution weight of that feature in the model's predictive decision. This calculation process evaluates the degree of change in the model's predicted value relative to the baseline value when the feature is present or absent. Through this calculation of marginal contribution values, the system can clearly define the positive or negative impact of each design variable and physical feature on material properties (such as tensile strength or degradation life).

[0059] Based on the calculated feature contribution values, the interpretability analysis module 33 performs analysis output at two levels. At the local interpretation level, for a specific fully biodegradable material recommended formulation, the system generates a feature contribution map, visually showing which features improve performance indicators and which features cause performance degradation, thereby verifying the physical rationality of the model's prediction results. At the global importance ranking level, the system globally ranks the feature importance by statistically analyzing the feature contribution values ​​of all samples in the validation set. In particular, this module can distinguish and quantify the relative importance of the original process formulation parameters and the physical simulation features. When the global importance of physical simulation features (such as interfacial adsorption energy) ranks high, it indicates that the multi-scale simulation layer 20 introduced in this invention plays a substantial corrective role in the prediction logic, rather than merely relying on the statistical correlation of the data. This quantitative analysis directly supports the technical solution in the claims regarding performance prediction based on physical features, enabling users to adjust their design strategies according to clearly defined key component weights.

[0060] The reverse design engine 62 in this invention is configured to transform engineering requirements in natural language form input by the user at the application and interaction layer into a multi-objective optimization problem in mathematical form. The engine first defines a vector of design variables. This vector encompasses all controllable parameters in the formulation design and processing technology of fully biodegradable materials. Specifically, the design variable vector... It includes component variables and process variables. The component variables involve the mass percentage of polymer matrix, reinforcing materials and functional additives, while the process variables involve extrusion temperature distribution, screw speed and traction ratio parameters.

[0061] Please see the appendix Figure 4 To address the trade-off between performance, cost, and degradation cycle commonly encountered in the development of fully biodegradable materials, the reverse design engine 62 constructs a system objective function vector containing multiple sub-objective functions. This engine formalizes the reverse engineering task as finding the optimal input vector under specific constraints. The process makes the objective function vector To achieve minimization. The mathematical expression of the multi-objective optimization model is defined as follows: ; in, Represents the vector of multi-objective functions to be optimized; This represents an input vector containing the formula and process parameters; The cost function of the material is calculated by weighted summation of the unit price data of each component in the material gene library 51 and the component content in the design variables. This represents a function predicting the mechanical properties of a material (e.g., tensile strength or modulus). Since optimization algorithms typically perform a minimization search, a negative sign is used in the formula, meaning that minimizing... To maximize mechanical performance; This represents the degradation cycle deviation function of the material; This indicates the target degradation period set by the user (e.g., 90 days or 180 days); The term represents the absolute deviation between the predicted lifetime and the target lifetime, aiming to make the degradation behavior of the design scheme as close as possible to the actual needs of the application scenario.

[0062] In this process, the reverse engineering engine 62 does not directly solve the physics equations. and Instead, it is configured to invoke the cluster of machine learning models in the aforementioned machine learning fusion layer 30 as a proxy model. That is, for each potential solution vector in the design space... The engine inputs the data into a trained GBDT or LSTM model to quickly obtain the corresponding performance and lifetime predictions, thereby enabling rapid evaluation of the objective function.

[0063] Furthermore, to ensure the generated formulation is materially feasible and conforms to the actual capabilities of the processing equipment, the reverse engineering engine 62 sets a strict set of constraints when constructing the optimization model. These constraints mainly include equality constraints and inequality constraints. Equality constraints are used to limit the normalization conditions of the formulation components, i.e., the sum of the mass percentages of all components must strictly equal 100%. Inequality constraints are used to limit the physical boundaries of process parameters; for example, the extruder temperature setpoint must be between the material's melting point and thermal decomposition temperature, and the screw speed must be within the equipment's allowed rated speed range. Through the above mathematical modeling, this invention transforms complex material research and development empirical problems into numerical optimization problems that can be automatically solved by a computer.

[0064] After constructing the multi-objective optimization mathematical model, the reverse design engine 62 was configured to perform global optimization using a reference-point-based non-dominated sorting genetic algorithm. This algorithm was selected to handle three or more conflicting objective dimensions—cost, mechanical properties, and degradation lifespan—in the design of fully biodegradable materials. Compared to traditional weighted summation methods or bi-objective optimization algorithms, it can effectively maintain the diversity of solution set distribution in the high-dimensional objective space.

[0065] During the actual execution of the solution program, the reverse engineering engine 62 performs population initialization operations within the design space. The system employs a real-number encoding strategy, mapping each specific material formulation ratio and process parameter (such as temperature and rotation speed in each zone) to a decision vector in a multi-dimensional space, and generates an initial population containing multiple individuals through a random distribution algorithm. Each individual represents a potential fully biodegradable material design scheme, and all generated individuals undergo boundary checks to ensure they strictly satisfy the aforementioned component normalization constraints and process physical boundary constraints.

[0066] For the initial population and subsequent offspring populations generated during iterations, the reverse design engine 62 performs a rapid fitness assessment based on a surrogate model. At this stage, the system does not directly conduct physical experiments on each individual; instead, it uses the decoded data of each individual in the population as input to batch call the pre-trained performance prediction model and lifetime prediction model in the machine learning fusion layer 30. The system obtains the tensile strength, elongation at break, and biodegradation cycle prediction values ​​output by each model, and combines this with real-time raw material cost data to calculate the performance of each individual in the objective function vector. The system employs a multidimensional fitness value. This virtual screening mechanism, based on a machine learning model, enables the system to complete performance evaluations of tens of thousands of potential recipes in a relatively short time.

[0067] Based on the calculated fitness values, the reverse engineering engine 62 performs non-dominated sorting and environment selection on the population. According to the Pareto dominance definition, the system divides the population into several non-dominated levels, with the first level containing all optimal solutions in the current population that are not dominated by other individuals. To maintain the diversity of the solution set and avoid premature convergence during the selection process, the system introduces a pre-defined set of reference points. These reference points are structurally and uniformly distributed on a normalized hyperplane. The system calculates the perpendicular distance from each individual at each level to the reference points and associates each individual with the nearest reference point. In the environment selection step, when it is necessary to remove some individuals from a certain non-dominated level to maintain the population size, the system prioritizes retaining individuals with fewer associated reference points, i.e., those located in sparse regions of the target space, based on the association count of the reference points. This ensures that the final solution set uniformly covers the entire Pareto front.

[0068] After the selection process is completed, the reverse design engine 62 performs simulated binary crossover and polynomial mutation operations on the selected individuals to generate the next generation of offspring. This evolutionary iteration process is repeated until the preset maximum number of generations is reached or the Pareto improvement of the population is lower than the convergence threshold. After the algorithm terminates, the reverse design engine 62 outputs the final non-dominated solution set, which contains a series of recommended formulations and process combinations that achieve different trade-offs between low cost, high performance, and controllable lifetime. The specific crossover probability, mutation probability, and reference point generation strategy can be conventionally configured by those skilled in the art based on the variable dimensions and computational resources of the actual problem; these are well-known techniques in the field and will not be elaborated upon here.

[0069] Experimental verification and active learning feedback step S5 are crucial for achieving continuous model evolution. To address the scarcity of high-quality labeled data due to the high cost and long cycle of physical experiments in the development of fully biodegradable materials, this invention integrates an active learning strategy into the reverse engineering engine 62. This strategy is configured not to rely entirely on direct verification of the Pareto optimal solution set, but rather to intelligently select experimental points that are most informative for improving the model's prediction accuracy through algorithms, thereby achieving convergence of the model's generalization ability with the fewest experimental iterations.

[0070] In the specific screening process, the reverse engineering engine 62 employs an uncertainty-based sampling strategy. The system utilizes Gaussian process regression or a deep neural network based on Bootstrap aggregation as the probabilistic prediction model. This allows the model to quantify the uncertainty (i.e., prediction variance) of the output performance prediction value. Based on this, the system defines the sampling function. To evaluate the experimental value of each potential sample point in the design space. Acquisition function. The calculation formula is as follows: ; in, The input vector representing the sample The higher the score of the sampling function, the more worthy the sample point is of physical experimental verification. This represents the mean of the machine learning model's predictions for the target performance (such as tensile strength) of the sample point, and represents the strategy employed, namely, the tendency to select regions that the current model considers to have excellent performance. It represents the standard deviation or uncertainty measure of the model's prediction results for that sample point, and represents the exploration strategy, that is, it tends to select areas where the model has blind spots or where the prediction fluctuates greatly; This represents an adjustment parameter used to balance the weight between development and exploration; it is typically set to a larger value in the early stages of research and development. Values ​​are intended to encourage exploration of the unknown, but are reduced in the later stages of research and development. The value is focused on the best performance.

[0071] By maximizing the aforementioned acquisition function, the system generates a list of recommended experiments for the next round. After the user or automated experimental equipment completes the preparation and testing of the fully biodegradable material according to this recommended list, the new data, including the actual components, processes, and measured performance, is fed back to the process knowledge database through the data and interface layer 10. Once the new data is entered into the database, the system triggers a closed-loop evolution mechanism, calling the machine learning fusion layer 30 to execute an incremental learning program. This program uses the new samples to calculate the gradient of the loss function, only fine-tuning and updating the model parameters, or triggering full retraining after accumulating a certain amount of data. This closed-loop iterative mechanism of prediction, verification, and updating enables the prediction model of this invention to dynamically correct its cognitive biases regarding the relationship between material composition, structure, and performance as experimental data accumulates. Especially in the unknown design space involving novel biodegradable polymers or complex compound systems, it can quickly converge to a high-precision prediction state, thereby supporting the technical solution of model optimization based on active learning in the claims.

[0072] This embodiment demonstrates how a user can use the reverse design workbench 41 in the application and interaction layer 40 to complete the entire process of formula inversion and verification for the needs of early spring crops for high strength and specific cost control of mulch film.

[0073] Please see the appendix Figure 7 During implementation, users set specific engineering constraints in the input area of ​​the interface. For example... Figure 7 As shown on the left, the mechanical performance constraints include: a flexural modulus of less than or equal to 750 MPa at 5°C to ensure sufficient flexibility during low-temperature laying; and a tensile strength of greater than or equal to 32 MPa to meet the tensile requirements for mechanical mulching. Simultaneously, economic and degradation constraints are set: raw material costs are controlled to within 21 yuan / kg, and degradation is required to begin within 60 days (end of the induction period).

[0074] Based on the above constraints, the reverse design engine 62 performs multi-objective optimization in the material gene library 51. The final formulation optimization results output by the system are as follows: Figure 7 The middle section shows the recommended ternary blend system, with a specific formulation ratio of 40% poly(butylene adipate / terephthalate), 25% polypropylene carbonate, 15% polylactic acid, and 14% functional filler. This formulation effectively balances the toughness of PBAT and the rigidity of PLA by introducing the PPC component, while reducing the overall cost by utilizing functional filler.

[0075] The system performed a virtual verification of the recommended solution, and the verification data is as follows: Figure 7As shown in the upper right corner, the model predicted a flexural modulus of 698 MPa for the key indicator of 5℃ flexural modulus, strictly meeting the design requirement of less than 750 MPa. Subsequent physical experiments measured an actual value of 690 MPa, with a relative error of only -1.7% between the predicted and experimental values, demonstrating the accuracy of the forward prediction engine 61 in handling complex blend systems. Furthermore, the SHAP analysis module at the bottom of the interface visually demonstrates the model's interpretability. The graph shows that the PBAT content has the highest node connection density, indicating that PBAT content is the largest contributing factor to the mechanical strength of this system. Through this intelligent design process, the entire R&D cycle was compressed to 3 weeks.

[0076] To verify the technical effect of the feature fusion module 31 of the present invention, this embodiment constructs four different prediction models for comparative testing, and evaluates the three key mechanical indicators of tensile strength, elongation at break and impact strength respectively.

[0077] Please see the appendix Figure 6 The four comparison methods are: a classical mechanism model based solely on physical equations, a traditional machine learning model based solely on statistical laws, a model using only original feature input parameters of the formulation and process, and the fusion invention method proposed in this paper (i.e., a machine learning model that integrates microscopic interface energy and mesoscopic dispersion index). The vertical axis in the figure represents the model's coefficient of determination score on the test set; a higher value indicates higher prediction accuracy.

[0078] As can be seen from the comparison results of the bar chart, the fusion method used in this invention achieved the highest accuracy scores in all three dimensions: tensile strength, elongation at break, and impact strength. Particularly in the two nonlinear characteristic indicators, elongation at break and impact strength, compared to models using only the original features, the method of this invention... The improvement in value ranged from 0.15 to 0.25. This indicates that relying solely on component ratios (original features) is insufficient to capture the influence of complex phase morphology within the material on toughness. Introducing physical features obtained through multi-scale simulation effectively supplements the model's understanding of the material's microstructure and corrects prediction biases. This result fully supports the technical solution in the claims regarding constructing enhanced feature vectors to improve prediction accuracy.

[0079] This embodiment aims to verify the advantages of the active learning sampling strategy in the system over traditional randomized experimental design in terms of model convergence efficiency.

[0080] Please see the appendix Figure 8 This embodiment records the changing trend of the model's prediction error for the unknown design space as the number of experimental iterations increases. For example... Figure 8As shown, the dark solid line represents the active learning curve based on the uncertainty sampling strategy of this invention, and the light dashed line represents the control group curve using randomly selected experimental points.

[0081] In the initial stage (0 iterations), due to the scarcity of training data, both models exhibited high prediction errors (RMSE approximately 0.2). As the experiment progressed, the active learning model's RMSE showed a rapid decline. Specifically, by the 4th iteration, the active learning model's RMSE had decreased to below 0.05 (marked by the vertical dashed line in the figure), meeting the requirements for high-precision prediction. In contrast, the randomized experimental group maintained a high RMSE in the 4th iteration, with a slow decline. Even by the 10th iteration, the randomized experimental group's RMSE was approximately 0.08, still not reaching the accuracy level of the active learning group in the 4th iteration.

[0082] Data analysis shows that, while achieving the same prediction accuracy (e.g., RMSE = 0.05), the active learning mechanism of this invention reduces the number of physical experiments required by approximately 60%. This result confirms that by maximizing the acquisition function... Recommending high-value experimental sites can effectively avoid repeated experiments in low-information areas, thereby shortening the trial-and-error cycle of materials research and development, and supporting the technical solutions regarding closed-loop evolution and incremental learning in the claims.

Claims

1. An experimental system for fully biodegradable materials based on multi-scale simulation and machine learning, characterized in that, The system includes an intelligent design and life prediction platform, which comprises a data and interface layer (10), a multi-scale simulation layer (20), a machine learning fusion layer (30), an application and interaction layer (40), a forward prediction engine (61), and a reverse design engine (62). The data and interface layer (10) is used to process heterogeneous data and output standardized data; The positive prediction engine (61) is used to receive the material formulation and process parameters in the standardized data, sequentially call the multi-scale simulation layer (20) to generate physical mechanism features and pass them to the machine learning fusion layer (30) to generate performance prediction values ​​and degradation lifetime prediction values. The reverse design engine (62) is used to perform reverse optimization from performance requirements to design variables, generate candidate solutions and input them into the forward prediction engine (61), use the performance prediction value and degradation lifetime prediction value to calculate the fitness of the candidate solutions, and select the Pareto optimal solution set that meets the constraints.

2. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 1, characterized in that, The intelligent design and life prediction platform also includes a materials gene library (51), a process knowledge base (52), and a model library (53). The data and interface layer (10) is equipped with a data acquisition module and a data standardization module; The data acquisition module is used to collect the heterogeneous data, which includes experimental test data, production line process data, field service environment data, and publicly available literature data. The data standardization module cleans, denoises, and unifies the format of the collected heterogeneous data to generate the standardized data. The material gene library (51) stores the molecular structure data and basic physical property parameters of polymer matrix, plant fiber and functional filler extracted from the publicly available literature data and the experimental test data in the standardized data; The process knowledge base (52) stores the process, structure, performance association rules and expert experience data obtained from the standardized data based on the experimental test data and the production line process data mining. The model library (53) stores the prediction models and weight parameters generated by the machine learning fusion layer (30) using the standardized data.

3. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 1, characterized in that, The multi-scale simulation layer (20) includes a microscale simulation engine (21) and a mesoscale simulation engine (22). The microscale simulation engine (21) is used to construct an interface composite model based on the data and molecular structure data output by the interface layer (10), and calculate the interface adsorption energy by subtracting the sum of the independent energies of the polymer matrix component and the filler component in the composite system from the total potential energy of the composite system after reaching the equilibrium conformation. The mesoscale simulation engine (22) divides the simulation system into multiple three-dimensional grids of equal volume by adopting a coarse-grained modeling strategy, counts the local volume fraction of the target phase in each grid, and calculates the dispersion uniformity index based on the standard deviation of the local volume fraction.

4. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 1, characterized in that, The multi-scale simulation layer (20) also includes a macro-scale simulation engine (23); The macroscale simulation engine (23) integrates finite element analysis units and computational fluid dynamics units; The finite element analysis unit is used to calculate the distribution variance of the maximum von Mises stress and stress gradient through a feature extraction algorithm, as a feature of the stress distribution field. The computational fluid dynamics unit is used to calculate the peak shear heat generation and average residence time in the flow field by numerically solving the Navier-Stokes equations and energy equations.

5. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 3, characterized in that, The machine learning fusion layer (30) includes a feature fusion module (31); The feature fusion module (31) is used to receive process formula parameters from the standardized data, as well as physical simulation features from the multi-scale simulation layer (20); The feature fusion module (31) is used as the input vector for model training and prediction. The input vector is composed of the process formula parameter part and the physical simulation feature part. The feature fusion module (31) normalizes the components of the process formula parameter part and the physical simulation feature part, and embeds the processed physical simulation feature as a supplementary dimension into the process formula parameter space, and calculates and generates an enhanced feature vector through the feature splicing method.

6. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 5, characterized in that, The machine learning fusion layer (30) also includes a machine learning model cluster (32), which is configured to load the prediction model and weight parameters and update the weight parameters using a transfer learning strategy that combines simulated pre-training with experimental fine-tuning. The cluster of machine learning models (32) includes a performance prediction model and a lifetime prediction model; The performance prediction model employs a gradient boosting decision tree and a deep neural network architecture, utilizing the updated weight parameters and based on the enhanced feature vector to predict the mechanical and thermal properties of the material. The life prediction model adopts a time-series prediction model using a long short-term memory network. The input of the life prediction model receives the enhanced feature vector and the field service environment data that changes over time in the standardized data. The weight parameters are adjusted through forget gate, input gate and output gate structure to handle the cumulative effect of the environment and calculate the material remaining weight percentage and mechanical property retention rate curve that change over time.

7. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 1, characterized in that, The reverse design engine (62) constructs a system objective function vector containing multiple sub-objective functions, the system objective function vector covering the cost function, mechanical performance prediction function and degradation cycle deviation function; The reverse design engine (62) uses a reference point-based non-dominated sorting genetic algorithm to generate candidate solutions containing formula ratios and process parameters during the optimization iteration process; For each candidate solution, the system inputs the candidate solution into the positive prediction engine (61) and uses the performance prediction value and degradation lifetime prediction value output by the positive prediction engine (61) to calculate the fitness of the candidate solution.

8. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 1, characterized in that, The system is used to activate the active learning mechanism; When the error between the predicted value and the measured value exceeds the threshold and the current scheme is in the region where the model prediction uncertainty is higher than the threshold, the system uses a Gaussian process regression model and a Bayesian neural network model to quantify the uncertainty of the predicted value. The system defines an acquisition function to evaluate the experimental value of each potential sample point in the design space, and calculates the acquisition function by adding the mean of the machine learning model's prediction of the target performance of the current sample point to the adjustment parameter multiplied by the standard deviation of the model's prediction result for the current sample point. The system generates suggested points for the next round of supplementary experiments by maximizing the acquisition function.

9. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 5, characterized in that, The machine learning fusion layer (30) is also provided with an interpretability analysis module (33); The interpretability analysis module (33) is used to calculate the marginal contribution value of each feature component in the enhanced feature vector to the final prediction result using the SHAP algorithm based on cooperative game theory. The interpretability analysis module (33) distinguishes and quantifies the relative importance of the process formulation parameters and the physical simulation features, and generates a feature contribution map.

10. The experimental system for fully biodegradable materials based on multi-scale simulation and machine learning according to claim 2, characterized in that, The application and interaction layer (40) includes a reverse design workbench (41), a virtual experiment system (42), and a knowledge graph visualization interface (43). The reverse design workbench (41) is used to receive performance constraints input by the user and display the Pareto optimal solution set generated by the reverse design engine (62) in the form of interactive charts for the user to make decisions and filter. The virtual experiment system (42) is used to call the positive prediction engine (61) and render the dynamic process of material properties evolving over time in the form of curves and cloud maps to simulate alternative physical experiments; The knowledge graph visualization interface (43) is used to call the process knowledge base (52) to construct and display the topological relationship graph between material structure, process parameters and performance indicators.

Images