A photovoltaic material structure performance relationship analysis method based on a graph neural network

Abstract

The application discloses a photovoltaic material structure performance relationship analysis method based on a graph neural network, and belongs to the technical field of the cross of artificial intelligence and new energy materials. The method comprises the following steps: acquiring degradation process data of a photovoltaic material under the coupling action of multiple working conditions, wherein the degradation process data comprises a multi-scale structure snapshot sequence and corresponding working condition parameter information; obtaining structure feature representation results corresponding to each structure snapshot, fusing the structure feature representation results with the corresponding working condition parameter information, and obtaining a working condition self-adaptive molecular graph sequence; extracting multi-scale structure features representing the degradation behavior of the photovoltaic material, and forming a corresponding time sequence feature sequence; obtaining time sequence state features representing the degradation evolution process of the photovoltaic material; constructing a photovoltaic material structure performance relationship model; obtaining corresponding performance attenuation results; and identifying key defect structures, key degradation stages and dominant degradation mechanisms affecting performance attenuation in the degradation process of the photovoltaic material.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of artificial intelligence and new energy materials, and more specifically, to a method for analyzing the relationship between the structure and performance of photovoltaic materials based on graph neural networks. Background Technology

[0002] As the core functional materials of solar cells, photovoltaic materials directly affect carrier transport, light absorption efficiency, and long-term device stability due to their microstructure characteristics. With the rapid development of new photovoltaic technologies such as high-efficiency crystalline silicon solar cells, perovskite solar cells, and tandem solar cells, the relationship between factors such as the internal crystal structure, defect distribution, interface state, and compositional evolution of materials and device performance has become increasingly complex. Especially during long-term operation, photovoltaic materials are subjected to the coupling effects of various environmental factors such as light, temperature, humidity, and electric bias voltage, which trigger a series of degradation phenomena such as defect generation, ion migration, interface degradation, and lattice distortion. This leads to a continuous decline in photoelectric conversion efficiency, seriously affecting the service life and reliability of photovoltaic devices.

[0003] In existing technologies, research on the relationship between the structure and performance of photovoltaic materials mainly relies on experimental characterization, physical model analysis, and statistical methods. Among these, experimental characterization methods can obtain information about the microstructure of materials at specific moments, but they are usually difficult to continuously characterize the dynamic degradation process of materials under complex operating conditions. Although physical models have a certain degree of interpretability, they often require a large number of prior assumptions and are difficult to accurately describe the complex nonlinear relationships between multi-scale structures. Statistical or traditional machine learning methods are mainly based on manually designed features for analysis, which makes it difficult to fully explore the topological relationships between atoms, grains, and defect structures within the material, resulting in limited ability to express structural features.

[0004] In recent years, graph neural networks have attracted widespread attention due to their ability to effectively represent complex graph structure data. By abstracting material structures into graph data, graph neural networks can be used to learn the relationships between the internal components of materials, providing a new technical approach for predicting material properties. However, most existing graph neural network-based material analysis methods focus on the mapping relationship between static structure and performance, and do not adequately consider the dynamic degradation evolution process of photovoltaic materials under multi-condition coupling. At the same time, they lack the ability to jointly analyze key defect structures, key degradation stages, and dominant degradation mechanisms during the degradation process, making it difficult to establish a structure-performance relationship model with time-series evolution characteristics and mechanistic explanation capabilities. This limits their application effectiveness in predicting photovoltaic material performance degradation and optimizing material design.

[0005] In summary, how to fully utilize the multi-scale structural evolution information of photovoltaic materials under multi-condition coupling, accurately establish a structural performance relationship model with time-series evolution characteristics and mechanism explanation capabilities, realize accurate prediction of material performance degradation process, and identify key defect structures, key degradation stages and dominant degradation mechanisms that affect performance degradation has become an urgent technical problem to be solved. Summary of the Invention

[0006] To overcome a series of shortcomings in existing technologies, the purpose of this application is to provide a method for analyzing the structural-performance relationship of photovoltaic materials based on graph neural networks, comprising the following steps: Data on the degradation process of photovoltaic materials under multi-condition coupling are obtained, including multi-scale structural snapshot sequences and corresponding condition parameter information. Feature extraction is performed on the multi-scale structural snapshot sequence to obtain the structural feature representation results corresponding to each structural snapshot. The structural feature representation results are then fused with the corresponding operating condition parameter information to obtain the operating condition adaptive molecular diagram sequence. The working condition adaptive molecular graph sequence is input into the defect perception graph neural network for structural characterization, extracting multi-scale structural features that characterize the degradation behavior of photovoltaic materials, and organizing the multi-scale structural features in chronological order to form the corresponding temporal feature sequence. Multi-timescale temporal coding and correlation modeling are performed on the temporal feature sequences to obtain temporal state features characterizing the degradation and evolution process of photovoltaic materials; Based on multi-scale structural features and temporal state features, a structural performance relationship model for photovoltaic materials is constructed. Based on the photovoltaic material structure-performance relationship model, the photoelectric conversion efficiency degradation rate, open-circuit voltage drift and short-circuit current degradation of photovoltaic materials are predicted, and the corresponding performance degradation results are obtained. Based on performance degradation results and multi-scale structural characteristics, we identify the key defect structures, key degradation stages, and dominant degradation mechanisms that affect performance degradation during the degradation process of photovoltaic materials.

[0007] In some embodiments, the method for feature extraction of multi-scale structural snapshot sequences is as follows: Based on the spatial distribution information of atoms in the multi-scale structural snapshot sequence, a multi-scale local neighborhood is constructed for each atom, and local chemical environment descriptors at different neighborhood scales are extracted. Scale fusion is performed on local chemical environment descriptors to obtain atomic features that characterize the local environment of atoms; Based on the interatomic interaction relationships in the structural snapshot, chemical bond type features, bond length distribution features, bond angle distribution features, and bond dynamic evolution features are extracted to obtain chemical bond structural features that characterize structural connection relationships. Defects are identified based on the structural deviation between the structural snapshot and the reference crystal structure, and defect density features, defect spatial distribution features, and defect interaction network features are extracted to obtain defect features. By correlating and fusing atomic features, chemical bond structure features, and defect features, structural feature characterization results corresponding to each structural snapshot are obtained.

[0008] In some embodiments, the method for obtaining the working condition adaptive molecular map sequence is as follows: Feature encoding is performed on the operating condition parameter information to obtain the corresponding operating condition features; Based on the operating condition characteristics, the atomic features are adaptively mapped and enhanced to obtain the operating condition adaptive atomic features; Based on the characteristics of operating conditions, the chemical bond structure characteristics and defect characteristics are dynamically adjusted and enhanced to obtain operating condition adaptive structural characteristics. By using the working condition adaptive atomic features as node attributes and the working condition adaptive structural features as edge attributes and graph-level attributes, a working condition adaptive molecular graph corresponding to each structural snapshot is constructed. The adaptive molecular diagrams for each operating condition are arranged according to the time sequence of the structural snapshots to form a sequence of adaptive molecular diagrams for each operating condition.

[0009] In some embodiments, the method for extracting multi-scale structural features is as follows: Based on the node embedding vectors output by each layer of the defect-aware graph neural network, the atomic local scale features are obtained by reading out and splicing them layer by layer. Hierarchical pooling is performed based on node embedding vectors, and nodes are clustered and aggregated according to chemical environment similarity to obtain grain-scale features; Based on the set of interface nodes in the adaptive molecular diagram, feature difference information of nodes on both sides of the interface is extracted to obtain the device interface scale features. A multi-scale structural feature representation is obtained by fusing and encoding atomic local scale features, grain scale features, and device interface scale features; By jointly encoding the multi-scale structural feature representation with the corresponding structural hierarchy label information, we obtain multi-scale structural features that characterize the degradation behavior of photovoltaic materials.

[0010] In some embodiments, the method for performing multi-timescale temporal coding and correlation modeling on temporal feature sequences is as follows: Parallel convolutional encoding of temporal feature sequences is performed based on causal temporal convolutional channels at multiple different time scales to extract temporal evolution features at the corresponding time scales, thereby obtaining temporal encoded features at multiple time scales. The time-series coding features corresponding to each time scale are normalized to obtain normalized time-series coding features; The normalized temporal coding features are weighted and fused based on the channel attention mechanism to obtain multi-timescale temporal coding vectors. Global temporal dependencies are established based on multi-timescale temporal coding vectors to obtain global temporal association features; Temporal state features are generated based on global temporal correlation features to characterize local time-varying patterns and global evolution trends during the degradation process.

[0011] In some embodiments, the method for obtaining the performance degradation result is as follows: Obtain the performance characterization features output by the photovoltaic material structure-performance relationship model; Based on performance characterization features, we establish branches for predicting photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation. Based on the photoelectric conversion efficiency attenuation rate prediction branch, a predicted value for the photoelectric conversion efficiency attenuation rate is generated. Based on the open-circuit voltage drift prediction branch and the short-circuit current degradation prediction branch, the predicted values ​​of open-circuit voltage drift and short-circuit current degradation are generated respectively. Consistency verification and joint constraint analysis are performed on the predicted values ​​of photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation to generate corresponding performance degradation results.

[0012] In some embodiments, the method for identifying critical defect structures is as follows: Based on the performance degradation results and multi-scale structural features, the performance contribution of each structural feature node is determined. Candidate defect sites that are highly correlated with performance degradation are selected based on their performance contribution. Establish a causal relationship between candidate defect sites and performance degradation results to identify key defect sites that have a significant impact on performance degradation. Based on the defect attribute information, spatial distribution information and structural hierarchy information of key defect sites, the key defect sites are aggregated and characterized to generate key defect structure identification results.

[0013] In some embodiments, the method for identifying critical degradation stages is as follows: The cosine distance between temporal features at adjacent time points is calculated based on the temporal feature sequence to obtain the degradation behavior change sequence. Statistical analysis of the degradation behavior change sequence is performed based on a sliding time window, and an adaptive change threshold is generated; The moment of abrupt change in the degradation state is identified by comparing the sequence of changes in degradation behavior with the adaptive change threshold, and the moment of abrupt change in the degradation state is used as the boundary of the degradation stage. The temporal feature sequence is segmented based on the degradation stage boundary to obtain multiple degradation stages, and the stage representative features corresponding to each degradation stage are extracted. The correlation strength between representative characteristics of each stage and photoelectric conversion efficiency attenuation rate, open-circuit voltage drift and short-circuit current degradation; Based on the correlation strength, each degradation stage is screened to determine the key degradation stage, and its corresponding start and end times and cumulative performance loss percentage are output.

[0014] In some embodiments, the method for identifying the dominant degradation mechanism is as follows: The pre-built candidate degradation mechanism library is retrieved, and the key defect structure, key degradation stage and performance degradation result corresponding to the current degradation process are obtained; the candidate degradation mechanism library includes multiple degradation mechanism prototypes jointly characterized by structural evolution mode, key defect characteristics and performance degradation behavior; The key defect structure, key degradation stage and performance degradation results are matched with each degradation mechanism prototype in the candidate degradation mechanism library in a multimodal manner to obtain the comprehensive matching score of each degradation mechanism prototype. The prototypes of each degradation mechanism are ranked based on the comprehensive matching score, and candidate dominant degradation mechanisms that meet the preset conditions are selected based on the comprehensive matching score. Contribution analysis and synergy analysis are performed on the candidate dominant degradation mechanisms to determine the dominant degradation mechanism and output the corresponding mechanism contribution information.

[0015] In some embodiments, the photovoltaic material structure-performance relationship analysis method further includes: Determine whether the dominant degradation mechanism meets the preset mechanism parsing integrity condition; If not, supplementary degradation process data under the corresponding working conditions will be collected, and the structural characterization, temporal modeling and degradation analysis processes will be re-executed until the integrity condition of mechanism analysis is met. If the conditions are met, then based on the identified key defect structures, key degradation stages, and dominant degradation mechanisms, corresponding material modification directions and structural optimization suggestions will be generated.

[0016] Compared with the prior art, this application has the following beneficial effects: This application enables high-precision prediction of performance degradation during the degradation process of photovoltaic materials, and simultaneously identifies key defect structures and dominant degradation mechanisms, thereby providing a reliable quantitative basis for the optimization of photovoltaic material structure and the improvement of lifespan. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for analyzing the structural-performance relationship of photovoltaic materials based on graph neural networks, as disclosed in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the defect-aware graph neural network structure in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of multi-timescale temporal coding and correlation modeling in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some embodiments of this invention, but not all embodiments.

[0021] 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.

[0022] The embodiments and directional terms described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0023] like Figure 1 As shown, a method for analyzing the structural-performance relationship of photovoltaic materials based on graph neural networks includes the following steps: Data on the degradation process of photovoltaic materials under the coupled effects of illumination, damp heat and bias conditions were obtained. The degradation process data included multi-scale structural snapshot sequences and corresponding operating condition parameter information. Feature extraction is performed on the multi-scale structural snapshot sequence to obtain the structural feature representation results corresponding to each structural snapshot. The structural feature representation results are then fused with the corresponding operating condition parameter information to obtain the operating condition adaptive molecular diagram sequence. The working condition adaptive molecular graph sequence is input into the defect perception graph neural network for structural characterization, extracting multi-scale structural features that characterize the degradation behavior of photovoltaic materials, and organizing the multi-scale structural features in chronological order to form the corresponding temporal feature sequence. Multi-timescale temporal coding and correlation modeling are performed on the temporal feature sequences to obtain temporal state features characterizing the degradation and evolution process of photovoltaic materials; Based on multi-scale structural features and temporal state features, a structural performance relationship model for photovoltaic materials is constructed. Based on the photovoltaic material structure-performance relationship model, the photoelectric conversion efficiency degradation rate, open-circuit voltage drift and short-circuit current degradation of photovoltaic materials are predicted, and the corresponding performance degradation results are obtained. Based on performance degradation results and multi-scale structural characteristics, we identify and analyze the key defect structures, key degradation stages, and dominant degradation mechanisms that affect performance degradation during the degradation process of photovoltaic materials. Determine whether the dominant degradation mechanism meets the preset mechanism parsing integrity condition; If not, supplementary degradation process data under the corresponding working conditions will be collected, and the structural characterization, temporal modeling and degradation analysis processes will be re-executed until the integrity condition of mechanism analysis is met. If the conditions are met, then based on the identified key defect structures, key degradation stages, and dominant degradation mechanisms, corresponding material modification directions and structural optimization suggestions will be generated.

[0024] The aforementioned graph neural network-based method for analyzing the structural-performance relationship of photovoltaic materials achieves a comprehensive characterization of photovoltaic material degradation behavior through multi-scale structural feature extraction and temporal modeling driven by coupled degradation data under multiple operating conditions. By leveraging the synergistic effect of a defect-aware graph neural network and multi-timescale temporal coding, it accurately captures the structural evolution patterns during degradation. Based on this, a photovoltaic material structural-performance relationship model is constructed to predict key performance degradation indicators, simultaneously identifying key defect structures, key degradation stages, and dominant degradation mechanisms. Iterative verification ensures the completeness of the mechanism analysis, thus providing systematic guidance for the modification and structural optimization of photovoltaic materials, significantly improving the accuracy and interpretability of photovoltaic material degradation mechanism analysis.

[0025] In some embodiments, the method for obtaining degradation process data is as follows: Construct a coupled degradation environment encompassing illumination, humidity and heat, and bias conditions, and deploy an online in-situ detection system and a distributed sensor network within this degradation environment; Based on the changes in the degradation rate of photovoltaic materials, a non-uniform time step sampling strategy is determined to generate the corresponding data acquisition time sequence; Based on the data acquisition time sequence, the operating condition parameter sequence is obtained by collecting information on light intensity, temperature and humidity gradient and electric field intensity through a distributed sensor network; Based on the data acquisition time sequence, the crystal structure information, nanostructure information and surface morphology information of photovoltaic materials are simultaneously acquired through an online in-situ detection system to form a corresponding multi-scale structural snapshot sequence. The operating condition parameter sequence and the multi-scale structural snapshot sequence are time-stamp aligned and correlated and fused, and then fused according to a unified time base to obtain the degradation process data of photovoltaic materials under the coupled effects of illumination, damp heat and bias conditions.

[0026] For example, taking a certain perovskite photovoltaic thin film material as an example, an accelerated degradation platform was built in the laboratory, where the illumination conditions were set to simulate AM1.5G sunlight irradiation (1000W / m).2 The system was configured with a constant 85℃ / 85%RH environment for the damp heat condition and a +0.8V external electric field for the bias condition, achieving synchronous coupling loading of the three conditions. Based on the material's degradation pattern—relatively slow in the initial stage (0–24h), abrupt structural changes in the middle stage (24–120h), and stable performance degradation in the later stage (after 120h)—the system adaptively generated a non-uniform sampling time sequence. For example, sampling was performed every 6 hours from 0–24h, every hour from 24–120h, and every 3 hours after 120h. At each sampling time, the distributed sensor network synchronously recorded an illumination intensity of approximately 1000W / m². 2 The system measures temperature (approximately 85℃), relative humidity (approximately 85%), and local electric field intensity distribution changes. Simultaneously, an online in-situ detection system acquires crystal structure phase transition information, nanocrystal evolution characteristics, and surface roughness variation maps through in-situ XRD, in-situ TEM, and in-situ AFM, respectively, forming a multi-scale structural snapshot sequence at corresponding time points. Subsequently, the above-mentioned operating condition parameter sequence and the structural snapshot sequence are aligned based on a unified timestamp, for example, by normalizing time using the "degradation initiation time" as the zero point. Interpolation and resampling are then used to match and fuse data from different sampling frequencies, ultimately obtaining degradation process data that reflects the entire process of this perovskite material under the coupled effects of light, humidity, heat, and bias voltage, from lattice distortion and interface defect generation to macroscopic performance degradation.

[0027] The degradation process data acquisition method described in this application achieves comprehensive real-time monitoring of the photovoltaic material degradation process by constructing a multi-condition coupled degradation environment and deploying an online in-situ detection system and a distributed sensor network. Through a non-uniform time step sampling strategy, the sampling density is increased during periods of rapid degradation rate change, while the sampling frequency is appropriately reduced during stable periods, effectively minimizing data acquisition redundancy. Furthermore, the operating condition parameter sequence and multi-scale structural snapshot sequence are time-stamp aligned and correlated to ensure consistency of multi-source heterogeneous data in the time dimension, thereby providing a high-quality, multi-scale, and time-complete degradation process data foundation for subsequent structural performance relationship analysis.

[0028] In some embodiments, the method for feature extraction of multi-scale structural snapshot sequences is as follows: Based on the spatial distribution information of atoms in the multi-scale structural snapshot sequence, a multi-scale local neighborhood is constructed for each atom, and local chemical environment descriptors at different neighborhood scales are extracted. Scale fusion is performed on local chemical environment descriptors to obtain atomic features that characterize the local environment of atoms; Based on the interatomic interaction relationships in the structural snapshot, chemical bond type features, bond length distribution features, bond angle distribution features, and bond dynamic evolution features are extracted to obtain chemical bond structural features that characterize structural connection relationships. Defects are identified based on the structural deviation between the structural snapshot and the reference crystal structure, and defect density features, defect spatial distribution features, and defect interaction network features are extracted to obtain defect features. By correlating and fusing atomic features, chemical bond structure features, and defect features, structural feature characterization results corresponding to each structural snapshot are obtained.

[0029] For example, taking a multi-scale structural snapshot of a perovskite photovoltaic material at a certain degradation time (e.g., after 120 hours of degradation) under the combined effects of light, humidity, and bias, feature extraction is performed. First, based on atomic-level structural information (e.g., the spatial coordinate distribution of Pb, I, and organic cations in the lattice), local neighborhoods of different radius scales (e.g., 0.5 nm, 1.0 nm, and 1.5 nm range) are constructed for each Pb atom, and the coordination number, elemental composition ratio, and electron density distribution within each neighborhood are statistically analyzed to form local chemical environment descriptors at different scales. Subsequently, the descriptors at different scales are weighted and fused to obtain atomic features characterizing the stability of the Pb atom and its surrounding local structure. Second, based on the interatomic interaction relationships in this structural snapshot, the interaction features of Pb-I bonds, weak II interactions, and the interaction between organic cations and the inorganic framework are extracted. The changes in bond length distribution (e.g., from 2.8 Å to 3.1 Å), bond angle distortion, and dynamic evolution trends at consecutive degradation times are statistically analyzed to form chemical bond structure features reflecting the stability of lattice connections. Furthermore, by comparing the current structural snapshot with the initial ideal perovskite crystal structure, lattice vacancies, interstitial atoms, and locally collapsed regions are identified. For example, a significant increase in I-vacancy concentration is observed, with clusters found in grain boundary regions. Based on this, defect density characteristics are calculated, and a spatial distribution map of defect clusters and the interaction network between defects are constructed, forming defect features. Finally, atomic features, chemical bond structure features, and defect features are unified and fused to obtain a comprehensive structural feature characterization result of the structural snapshot at this moment. This enables multi-scale structural characterization of perovskite materials under coupled degradation conditions, from local coordination changes and bond relaxation to defect accumulation evolution.

[0030] The multi-scale structural snapshot sequence feature extraction method described in this application comprehensively captures atomic-scale structural information through multi-scale local neighborhood construction and local chemical environment descriptor extraction. It deeply characterizes the structural connectivity and bonding state of photovoltaic materials by comprehensively extracting chemical bond types, bond length distributions, bond angle distributions, and bond dynamic evolution. By comparing deviations with a reference crystal structure, it identifies defects and extracts defect density, spatial distribution, and interaction network features, achieving a refined description of defect structures. Based on this, it correlates and fuses atomic features, chemical bond structure features, and defect features to form multi-level, high-information structural feature characterization results, providing rich and accurate structural information input for subsequent adaptive molecular graph construction.

[0031] In some embodiments, the method for obtaining the working condition adaptive molecular map sequence is as follows: Feature encoding is performed on the operating condition parameter information to obtain the corresponding operating condition features; Based on the operating condition characteristics, the atomic features are adaptively mapped and enhanced to obtain the operating condition adaptive atomic features; Based on the characteristics of operating conditions, the chemical bond structure characteristics and defect characteristics are dynamically adjusted and enhanced to obtain operating condition adaptive structural characteristics. By using the working condition adaptive atomic features as node attributes and the working condition adaptive structural features as edge attributes and graph-level attributes, a working condition adaptive molecular graph corresponding to each structural snapshot is constructed. The adaptive molecular diagrams for each operating condition are arranged according to the time sequence of the structural snapshots to form a sequence of adaptive molecular diagrams for each operating condition.

[0032] For example, taking the degradation stage of perovskite photovoltaic materials in a light-damp-heat-bias coupling degradation environment from 120h to 150h as an example, we construct and serialize molecular maps based on the operating conditions of multiple consecutive structural snapshots. First, we encode the operating conditions corresponding to this stage, for example, using a light intensity of 1000W / m². 2 Temperature (85℃), relative humidity (85%), and applied bias voltage (+0.8V) are uniformly mapped into a high-dimensional feature vector representing the current strong stress-coupled degradation state of the material. Subsequently, atomic features are adaptively enhanced based on these features. For example, under high temperature and humidity conditions, the weights on the migration activity of I atoms and the coordination instability around Pb are increased, making the nodal features of Pb, I, and organic cations exhibit stronger ion migration tendencies and local structure weakening effects in the graphical representation, thus forming condition-adaptive atomic features. Simultaneously, dynamic adjustments are made to chemical bond structure and defect features driven by the operating conditions. For example, under enhanced bias voltage conditions, the influence weights on Pb-I bond length stretching and bond angle distortion are strengthened, and the characterization intensity of I vacancy defect aggregation in grain boundary regions is increased, thereby obtaining condition-adaptive structural features reflecting electric field-driven ion migration and accelerated defect accumulation. Furthermore, by using the aforementioned adaptive atomic features as graph node attributes and the adaptive structural features as graph edge and global attributes, adaptive molecular graphs corresponding to multiple time points such as 120h, 130h, 140h, and 150h are constructed. Each graph simultaneously encodes the local atomic state, bond connections, and defect distribution state at the current moment. Finally, the adaptive molecular graphs are arranged in chronological order to form a sequence of adaptive molecular graphs, thereby achieving a graph-structured dynamic representation of the entire process of "condition-driven—structural response—defect evolution" of the perovskite material under coupled degradation conditions.

[0033] The above-mentioned method for obtaining adaptive molecular graph sequences under operating conditions in this application achieves deep coupling between structural features and operating condition information by encoding the operating condition parameters and dynamically adjusting the atomic and chemical bond structure features. This enables the node attributes, edge attributes, and graph-level attributes of the molecular graph to reflect the influence of actual operating conditions on the material structure. Based on this, an adaptive molecular graph sequence under operating conditions is constructed by arranging the sequences in chronological order. This allows the graph neural network to model the degraded structure under the premise of perceiving differences in operating conditions, thereby significantly improving the adaptability of structural characterization to different operating conditions and the pertinence of degradation behavior prediction.

[0034] In some embodiments, such as Figure 2 As shown, the defect-aware graph neural network includes a message passing subnetwork for normal regions, a reinforcement passing subnetwork for defective regions, and a cross-branch attention fusion module, wherein: The normal region message passing subnetwork is used to perform graph attention message passing and neighborhood information aggregation on non-defect nodes in the molecular graph to obtain a normal lattice region feature representation. The defect region reinforcement propagation subnetwork is used to introduce defect density information and defect type information during graph attention message propagation, and adjust the information propagation intensity between nodes based on the defect information to obtain defect region feature representation; The cross-branch attention fusion module is used to establish the interactive correlation between the feature representation of the normal lattice region and the feature representation of the defect region. It extracts the coupling relationship between the two types of features through the cross-attention mechanism and encodes the fused features to generate a defect-aware map representation. The defect-aware map representation is used to characterize the multi-scale structural evolution features of photovoltaic materials during the degradation process.

[0035] For example, taking the condition-adaptive molecular graph of a perovskite photovoltaic material at 150 hours under the coupled degradation conditions of light, humidity, and bias as an example, a defect-aware graph neural network modeling analysis is performed. First, in the normal region message passing subnetwork, nodes in regions with relatively intact lattice structures and no obvious defect aggregation (such as structurally stable Pb–I framework regions) are designated as non-defect nodes. Their neighborhood information is weighted and aggregated using a graph attention mechanism, considering factors such as the integrity of surrounding coordination and the stability of local bond lengths, to obtain a characteristic representation of the normal lattice region representing the overall lattice order. Second, in the defect region reinforcement and propagation subnetwork, regions with high-density I-vacancy aggregation, enhanced grain boundary dislocations, and local structural collapse are designated as defect-sensitive regions. Defect density features and defect type labels (such as vacancy defects, interstitial defects, and dislocation defects) are introduced. During message passing, the information propagation weights between defect region nodes are dynamically amplified or suppressed, for example, enhancing the correlation propagation within defect aggregation regions and weakening interference propagation across stable lattice regions, thereby obtaining a characteristic representation of the defect region that reflects the defect expansion trend. Furthermore, in the cross-branch attention fusion module, cross-attention calculations are performed on the feature representations of normal lattice regions and defect regions. For example, defect region features are used as query vectors and normal region features as key vectors to characterize the mutual influence relationship between the "lattice-ordered structure and defect-perturbed structure," thereby extracting the evolutionary pattern of mutual coupling between the two during degradation. The fused high-dimensional features are then uniformly encoded to generate a defect-aware map representation. Ultimately, this defect-aware map representation can simultaneously characterize the multi-scale structural evolution characteristics of the perovskite material under coupled degradation conditions, including the gradual instability of the lattice structure, the continuous accumulation and expansion of defects, and the mutual feedback influence between the two.

[0036] The defect-aware graph neural network described in this application employs a dual-branch parallel structure consisting of a message passing subnetwork for normal regions and a reinforcement passing subnetwork for defect regions. This allows for differentiated feature extraction for normal lattice regions and defect regions, avoiding the dilution of defect information by normal region features in conventional graph neural networks. Furthermore, a cross-branch attention fusion module utilizes a cross-attention mechanism to establish interactive relationships between the two types of region features, effectively capturing the coupling effect between normal lattice and defect regions. Based on this, a defect-aware graph representation is generated, thereby comprehensively and accurately characterizing the multi-scale structural evolution features of photovoltaic materials during degradation, significantly improving the graph neural network's ability to perceive defect-related degradation behavior and the quality of feature representation.

[0037] In some embodiments, the method for extracting multi-scale structural features is as follows: Based on the node embedding vectors output by each layer of the defect-aware graph neural network, the atomic local scale features are obtained by reading out and splicing them layer by layer. Hierarchical pooling is performed based on node embedding vectors, and nodes are clustered and aggregated according to chemical environment similarity to obtain grain-scale features; Based on the set of interface nodes in the adaptive molecular diagram, feature difference information of nodes on both sides of the interface is extracted to obtain the device interface scale features. A multi-scale structural feature representation is obtained by fusing and encoding atomic local scale features, grain scale features, and device interface scale features; By jointly encoding the multi-scale structural feature representation with the corresponding structural hierarchy label information, we obtain multi-scale structural features that characterize the degradation behavior of photovoltaic materials.

[0038] For example, taking the defect-aware map representation of perovskite photovoltaic materials at 150 hours under the coupled degradation conditions of light-humidity-bias voltage as an example, multi-scale structural feature extraction is performed. First, node embedding vectors are obtained from the outputs of different layers of the defect-aware map neural network. For example, the first layer mainly represents the local coordination information of atoms, the second layer captures the mid-range bond connection relationship, and the third layer reflects the defect propagation and structural perturbation propagation. The node embedding vectors of each layer are read out layer by layer and concatenated according to the node order to form atomic local-scale features that can characterize the small changes in the local environment of Pb, I, and organic cations. Second, hierarchical pooling is performed on the node embedding vectors, and the nodes are clustered according to the similarity of chemical environment (such as coordination number, local bond length distribution, and electron density similarity). For example, stable regions inside the lattice are clustered into one category, and defect-rich regions are clustered into another category, thereby forming grain-scale features that reflect the consistency of the internal structure and the differences in degradation. Furthermore, based on the device interface node set (such as the interface region between the perovskite layer and the electrode) in the adaptive molecular diagram, feature difference information of the nodes on both sides of the interface is extracted, such as carrier distribution inhomogeneity, bond structure distortion degree, and defect density gradient, thereby obtaining device interface-scale features that can characterize interface degradation and charge transport deterioration. Subsequently, atomic local scale features, grain scale features, and device interface-scale features are fused and encoded to form a unified multi-scale structural feature representation, which is used to comprehensively characterize the overall evolution state of the material from atomic-level mismatch to grain-level reconstruction and then to interface-level degradation. Finally, this multi-scale structural feature representation is jointly encoded with the corresponding structural hierarchy label information (such as "atomic-level defect dominance", "grain boundary expansion stage", "interface degradation acceleration stage", etc.) to obtain a multi-scale structural feature representation that can comprehensively characterize the degradation behavior of perovskite photovoltaic materials.

[0039] The multi-scale structural feature extraction method described in this application preserves rich local chemical environment information at the atomic scale by reading and concatenating the embedded vectors of each layer of the defect-aware graph neural network. Through hierarchical pooling and chemical environment similarity clustering, grain-scale features are obtained from atomic-scale features, achieving multi-scale upsampling of structural information. By extracting the feature differences between nodes on both sides of the interface, structural discontinuities and interface degradation information at the device interface are specifically captured. Based on this, the structural features at the three scales are fused and encoded, and jointly encoded with structural hierarchy labels, thereby constructing a multi-scale structural feature system covering the atomic local scale, grain scale, and device interface scale, significantly enhancing the ability to characterize the multi-level structure of photovoltaic material degradation behavior.

[0040] In some embodiments, such as Figure 3 As shown, the method for multi-timescale temporal coding and correlation modeling of temporal feature sequences is as follows: Parallel convolutional encoding of temporal feature sequences is performed based on causal temporal convolutional channels at multiple different time scales to extract temporal evolution features at the corresponding time scales, thereby obtaining temporal encoded features at multiple time scales. The time-series coding features corresponding to each time scale are normalized to obtain normalized time-series coding features; The normalized temporal coding features are weighted and fused based on the channel attention mechanism to obtain multi-timescale temporal coding vectors. Global temporal dependencies are established based on multi-timescale temporal coding vectors to obtain global temporal association features; Temporal state features are generated based on global temporal correlation features to characterize local time-varying patterns and global evolution trends during the degradation process.

[0041] For example, taking the multi-scale structural feature time series of perovskite photovoltaic materials under the conditions of light-damp-heat-bias coupling degradation from 0h to 150h as an example, we perform multi-timescale temporal coding and correlation modeling. First, for this temporal feature series, the time scale is divided into three levels: short-term (e.g., 1h window), medium-term (e.g., 6h window), and long-term (e.g., 24h window), and corresponding causal temporal convolution channels are constructed for each level. In the short-term channel, we focus on capturing rapidly changing features, such as the instantaneous transition of defect density in local grain boundary regions; in the medium-term channel, we characterize the gradual rearrangement process of grain-scale structure; and in the long-term channel, we learn the overall device performance degradation trend, thus obtaining the temporal coding features corresponding to multiple time scales. Second, the temporal coding features obtained at different time scales are normalized to eliminate the differences in numerical amplitude and frequency of change between different time scales, making them comparable in a unified feature space, thus forming normalized temporal coding features. Furthermore, based on a channel attention mechanism, the normalized multi-scale temporal coding features are adaptively weighted. For example, the weight of short-term channels is increased during the accelerated degradation phase, and the contribution of long-term channels is enhanced during the stable decay phase, thus fusing them to obtain a multi-timescale temporal coding vector that can simultaneously reflect both fast and slow-changing characteristics. Subsequently, a global temporal dependency relationship is established based on this multi-timescale temporal coding vector. For example, long-range dependency modeling is used to capture the evolutionary chain of "early defect initiation—mid-term expansion—late-term failure," thereby obtaining global temporal correlation features. Finally, temporal state features are generated from the global temporal correlation features to comprehensively characterize the unified temporal evolution law of the local rapid fluctuation behavior and the overall gradual performance degradation trend of the perovskite photovoltaic material during the coupled degradation process.

[0042] The multi-timescale temporal coding and correlation modeling method described in this application captures both the short-range local time-varying patterns and long-range slow evolution trends of the degradation process at different time scales through parallel convolutional coding of multiple causal temporal convolutional channels. This solves the problem that single-timescale modeling cannot simultaneously take into account both fast and slow degradation behaviors. Furthermore, an adaptive weighted fusion of multi-timescale temporal coding features is performed using a channel attention mechanism to highlight the temporal scale information that contributes most to degradation prediction. Based on this, a global temporal dependency relationship is established and temporal state features are generated, thereby comprehensively characterizing the temporal dynamics of photovoltaic material degradation evolution and significantly improving the temporal modeling accuracy and generalization ability for degradation trend prediction.

[0043] In some embodiments, the method for constructing the photovoltaic material structure-performance relationship model is as follows: Based on multi-scale structural features, feature decoupling is performed to obtain static structural components that reflect the inherent structural properties of the material and dynamic response components that reflect the response characteristics of the degradation process. A steady-state optoelectronic performance prediction branch is constructed based on the static structural components to extract the influence characteristics of material structure on steady-state optoelectronic performance and obtain the static structural performance characterization results. A transient performance fluctuation prediction branch is constructed based on the dynamic response component, and a cumulative degradation prediction branch is established in combination with the time-series state characteristics to obtain the corresponding dynamic performance characterization results and degradation state characterization results. By jointly integrating and correlating the static structural performance characterization results, dynamic performance characterization results, and degradation state characterization results, a structural performance relationship model of photovoltaic materials is constructed.

[0044] For example, taking the multi-scale structural characteristics and corresponding photoelectric performance data of perovskite photovoltaic materials under photo-humid heat-bias coupling degradation conditions from 0h to 150h as an example, a model of the relationship between the structure and performance of photovoltaic materials is constructed. First, the input multi-scale structural features are decoupled. For example, long-term stability-related features such as lattice defect density, grain size uniformity, and interface structure differences are classified as static structural components to characterize the inherent structural properties of the material; at the same time, features that change significantly over time, such as defect propagation rate, bond length dynamic fluctuations, and interface charge transport instability, are classified as dynamic response components to characterize the structural response behavior during degradation. Second, a steady-state photoelectric performance prediction branch is constructed based on the static structural components. For example, the open-circuit voltage (Voc), short-circuit current density (Jsc), and fill factor (FF) in the initial or steady-state stages are used as target outputs to learn the mapping relationship between lattice integrity and carrier transport capability, thereby obtaining the static structural performance characterization results, which are used to characterize the fundamental influence of material structure on steady-state photoelectric performance. Furthermore, a transient performance fluctuation prediction branch is constructed based on the dynamic response component. For example, efficiency fluctuations, output current jitter, and local response delays are modeled during continuous illumination loading. A cumulative degradation prediction branch is introduced, combining the aforementioned temporal state characteristics, to estimate factors such as the degree of interface damage accumulation, defect density growth, and overall performance degradation index, thereby obtaining dynamic performance characterization results and degradation state characterization results. Finally, the static structural performance characterization results, dynamic performance characterization results, and degradation state characterization results are jointly fused and modeled. For example, a coupling relationship between the three is established through multi-task learning or attention fusion mechanisms to achieve a unified mapping between "structural stability—transient response—long-term degradation," thus constructing a photovoltaic material structure-performance relationship model to comprehensively characterize the intrinsic correlation mechanism between the structural evolution and photoelectric performance degradation of this perovskite material under coupled degradation conditions.

[0045] The method for constructing the photovoltaic material structure-performance relationship model described in this application decouples the static structural components and dynamic response components of the multi-scale structural features, separating the inherent properties of the material from the response characteristics of the degradation process, thus avoiding mutual interference between the two. By constructing dedicated prediction branches for the static and dynamic components respectively, and establishing a cumulative degradation prediction branch in combination with time-series state characteristics, refined and independent modeling of steady-state photoelectric performance, transient performance fluctuations, and cumulative degradation state is achieved. On this basis, the three types of characterization results are jointly fused and correlated to construct a coupled mapping relationship between structure and performance, thereby significantly improving the comprehensive prediction accuracy of the photovoltaic material structure-performance relationship model for multi-dimensional performance degradation behavior.

[0046] In some embodiments, the method for obtaining the performance degradation result is as follows: Obtain the performance characterization features output by the photovoltaic material structure-performance relationship model; Based on performance characterization features, we establish branches for predicting photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation. Based on the photoelectric conversion efficiency attenuation rate prediction branch, a predicted value for the photoelectric conversion efficiency attenuation rate is generated. Based on the open-circuit voltage drift prediction branch and the short-circuit current degradation prediction branch, the predicted values ​​of open-circuit voltage drift and short-circuit current degradation are generated respectively. Consistency verification and joint constraint analysis are performed on the predicted values ​​of photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation to generate corresponding performance degradation results.

[0047] For example, taking the performance characterization characteristics of perovskite photovoltaic materials under irradiation-humidity-bias coupling degradation conditions for 150 hours as an example, the performance degradation results are predicted and verified. First, the performance characterization characteristics output by the structure-performance relationship model are obtained. These characteristics comprehensively reflect information such as the degree of lattice defect accumulation, the decrease in grain structure stability, and the decay of interface charge transport capability, and are used as a unified input for subsequent multi-branch predictions. Second, based on these performance characterization characteristics, three prediction branches are constructed: the photoelectric conversion efficiency degradation rate prediction branch is used to output the decrease in the overall energy conversion efficiency of the device; the open-circuit voltage drift prediction branch is used to characterize the Voc change trend caused by carrier recombination enhancement; and the short-circuit current degradation prediction branch is used to reflect the degree of Jsc decay caused by the decrease in charge collection efficiency. Furthermore, through the photoelectric conversion efficiency degradation prediction branch, a predicted value of approximately 23.2% is obtained, for example, that the efficiency decreases from the initial 18.5% to 14.2%. Through the open-circuit voltage drift prediction branch, a predicted drift value of Voc decreasing from 1.10V to 1.02V is obtained. Through the short-circuit current degradation prediction branch, a predicted value of Jsc decreasing from 22mA / cm² is obtained. 2Decreased to 18 mA / cm 2 The degradation prediction values ​​were then obtained. Subsequently, consistency verification and joint constraint analysis were performed on the three types of prediction results. For example, it was verified whether the efficiency degradation satisfies a product coupling relationship with changes in Voc and Jsc in terms of physical mechanism, and whether the three conform to the common degradation trend of enhanced carrier recombination and accelerated interface damage, thereby eliminating abnormal prediction results that do not meet physical consistency. Finally, based on satisfying the consistency constraints, the outputs were fused to generate the corresponding performance degradation results, which are used to comprehensively characterize the overall evolution of the perovskite photovoltaic material's photoelectric conversion capability decline and key electrical parameter degradation under coupled degradation conditions.

[0048] The method for obtaining the performance degradation results described in this application achieves refined and independent prediction of multi-dimensional performance degradation indicators of photovoltaic materials by constructing independent prediction branches for photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation. By performing consistency verification and joint constraint analysis on the three performance prediction values, it ensures that the physical constraint relationship between each performance degradation indicator is met, effectively filtering out unreasonable prediction results. On this basis, a comprehensive performance degradation result after constraint verification is generated, thereby significantly improving the physical consistency and reliability of performance degradation prediction, and providing accurate performance degradation basis for subsequent degradation mechanism analysis and material optimization direction generation.

[0049] In some embodiments, the method for identifying critical defect structures is as follows: Based on the performance degradation results and multi-scale structural features, the performance contribution of each structural feature node is determined. Candidate defect sites that are highly correlated with performance degradation are selected based on their performance contribution. Establish a causal relationship between candidate defect sites and performance degradation results to identify key defect sites that have a significant impact on performance degradation. Based on the defect attribute information, spatial distribution information and structural hierarchy information of key defect sites, the key defect sites are aggregated and characterized to generate key defect structure identification results.

[0050] For example, taking the multi-scale structural features and performance degradation results of perovskite photovoltaic materials under irradiation-humidity-bias coupling degradation conditions for 150 hours as an example, key defect structures are identified and analyzed. First, based on the multi-scale structural features and performance degradation results output by the structure-performance relationship model (e.g., efficiency degradation rate of approximately 23.2%, Voc drift, and Jsc degradation), the contribution of each structural feature node to performance degradation is calculated. For example, the influence intensity of different atomic nodes, grain regions, and interface nodes on the overall performance degradation is evaluated through gradient attribution or attention weight allocation. Second, candidate defect sites are screened according to the performance contribution. For example, regions with dense I-vacancies, concentrated grain boundary dislocations, and local collapse regions at electrode contact interfaces are considered as candidate defect regions highly correlated with performance degradation, while stable lattice regions with weak performance impact are eliminated. Furthermore, a causal relationship is established based on candidate defect sites and performance degradation results. For example, the synchronicity between defect propagation rate and efficiency degradation, and the coupling relationship between interface defect growth and Voc drift are analyzed to identify key defect sites that significantly drive performance degradation, such as grain boundary I-vacancy chain propagation regions and interface charge accumulation imbalance regions. Finally, by combining the defect attribute information (such as vacancy type and defect concentration), spatial distribution information (such as grain boundary aggregation or interface local enrichment), and structural hierarchy information (such as atomic level / grain level / interface level) of key defect sites, they are aggregated and characterized to form structured key defect structure identification results. This enables the localization and expression of the core defect mechanism that dominates performance degradation in perovskite photovoltaic materials during the coupling degradation process.

[0051] The critical defect structure identification method described in this application calculates the performance contribution of each structural feature node based on performance degradation results, quantitatively correlates structural features with performance degradation, and achieves accurate screening of candidate defect sites. By establishing a causal relationship between candidate defect sites and performance degradation results, it effectively distinguishes between critical defect sites that have a significant causal impact on performance degradation and ordinary defect sites that only have statistical correlation. On this basis, it combines defect attribute information, spatial distribution information, and structural hierarchy information to aggregate and characterize critical defect sites, generating structurally complete and information-rich critical defect structure identification results, providing clear structural targets for the analysis of dominant degradation mechanisms and the generation of material modification directions.

[0052] In some embodiments, the method for identifying critical degradation stages is as follows: The cosine distance between temporal features at adjacent time points is calculated based on the temporal feature sequence to obtain the degradation behavior change sequence. Statistical analysis of the degradation behavior change sequence is performed based on a sliding time window, and an adaptive change threshold is generated; The moment of abrupt change in the degradation state is identified by comparing the sequence of changes in degradation behavior with the adaptive change threshold, and the moment of abrupt change in the degradation state is used as the boundary of the degradation stage. The temporal feature sequence is segmented based on the degradation stage boundary to obtain multiple degradation stages, and the stage representative features corresponding to each degradation stage are extracted. The correlation strength between representative characteristics of each stage and photoelectric conversion efficiency attenuation rate, open-circuit voltage drift and short-circuit current degradation; Based on the correlation strength, each degradation stage is screened to determine the key degradation stage, and its corresponding start and end times and cumulative performance loss percentage are output.

[0053] For example, taking the temporal characteristic sequence and performance degradation results of perovskite photovoltaic materials under the coupled degradation conditions of light-dampness-bias voltage (LED-PDV) from 0h to 150h as an example, key degradation stages are identified and analyzed. First, similarity calculations are performed on the temporal characteristics of adjacent time points. For instance, based on multi-scale structural features and condition-adaptive molecular graph embedding representation, the cosine distance between time points t and t+1 is calculated to obtain a degradation behavior change sequence reflecting the severity of structural evolution. The cosine distance increases significantly during local structural abrupt changes (such as rapid propagation of grain boundary defects). Second, a sliding time window (e.g., a 10h window) is used to perform statistical analysis on the change sequence, calculating the local mean and variance. Based on this, a threshold that adapts to time is generated, allowing the threshold to gradually adapt from a slow change to a rapid change range as the degradation stage progresses. Furthermore, the degradation behavior change sequence is compared with an adaptive threshold. When the change value continuously exceeds the threshold, it is identified as a moment of abrupt change in the degradation state. For example, significant abrupt change points were detected around 30h, 85h, and 120h, and these were used as the boundaries of the degradation stages, thus dividing the entire 150h degradation process into multiple stages. Subsequently, representative features of each degradation stage are extracted. For example, the early stage is dominated by slight lattice distortion, the middle stage is dominated by rapid increase in defect density, and the late stage is dominated by interface failure and carrier transport degradation. The correlation strength between these features and the photoelectric conversion efficiency decay rate, Voc drift, and Jsc degradation is calculated to quantify the contribution of each stage to the overall performance loss. Finally, key degradation stages are screened based on the correlation strength. For example, the 85h to 120h stage is identified as the key interval for dominant performance decay, and its corresponding start and end times and cumulative performance loss percentage (e.g., this stage contributes more than 60% to the overall efficiency decay) are output, thereby achieving accurate location and quantitative characterization of the dominant failure stage in the degradation process of perovskite photovoltaic materials.

[0054] The key degradation stage identification method described in this application constructs a degradation behavior change sequence by calculating the cosine distance of temporal features at adjacent time points, transforming the abrupt change information of the degradation state into a quantifiable measure of change. By adaptively generating a change threshold based on a sliding time window, it overcomes the problem of insufficient adaptability of fixed thresholds in different degradation rate ranges. On this basis, it uses the abrupt change time of the degradation state to delineate stage boundaries, extracts representative features of each stage and evaluates their correlation strength with multidimensional performance degradation indicators, accurately screens out key degradation stages that significantly contribute to performance degradation, and outputs the start and end times and the cumulative performance loss percentage, providing a clear stage division basis for degradation mechanism analysis.

[0055] In some embodiments, the method for identifying the dominant degradation mechanism is as follows: The pre-built candidate degradation mechanism library is retrieved, and the key defect structure, key degradation stage and performance degradation result corresponding to the current degradation process are obtained; the candidate degradation mechanism library includes multiple degradation mechanism prototypes jointly characterized by structural evolution mode, key defect characteristics and performance degradation behavior; The key defect structure, key degradation stage and performance degradation results are matched with each degradation mechanism prototype in the candidate degradation mechanism library in a multimodal manner to obtain the comprehensive matching score of each degradation mechanism prototype. The prototypes of each degradation mechanism are ranked based on the comprehensive matching score, and candidate dominant degradation mechanisms that meet the preset conditions are selected based on the comprehensive matching score. Contribution analysis and synergy analysis are performed on the candidate dominant degradation mechanisms to determine the dominant degradation mechanism and output the corresponding mechanism contribution information.

[0056] For example, taking the complete 150-hour degradation process of perovskite photovoltaic materials under the coupled degradation conditions of light-humid heat-bias voltage as an example, the dominant degradation mechanism is identified and analyzed. First, various degradation mechanism prototypes are retrieved from a pre-constructed candidate degradation mechanism library, such as "ion migration-dominated mechanism", "humid heat-induced grain boundary corrosion mechanism", "electric field-driven defect aggregation mechanism" and "interfacial carrier recombination enhancement mechanism". Each mechanism prototype is defined by the structural evolution mode (such as lattice distortion path), key defect characteristics (such as I-vacancy diffusion or increase in interfacial state density) and performance degradation behavior (such as the shape of the efficiency decline curve). Secondly, for the current degradation process, a multimodal matching is performed with the key defect structures (such as the chain expansion of grain boundary I vacancies and enrichment of interface defects), key degradation stages (such as the rapid decay stage from 85 to 120 hours), and performance degradation results (such as efficiency degradation of approximately 23%, Voc drift, and Jsc decrease characteristics) and various degradation mechanism prototypes. For example, the matching degree is calculated in three dimensions: structural similarity, defect evolution consistency, and performance curve fitting degree, thereby obtaining a comprehensive matching score for each degradation mechanism prototype. Furthermore, the degradation mechanism prototypes are ranked according to their comprehensive matching scores. For example, it was found that the scores of "ion migration-dominated mechanism" and "electric field-driven defect aggregation mechanism" are significantly higher than other mechanisms, and a set of candidate dominant degradation mechanisms that meet the preset threshold is selected. Finally, the contribution and synergistic relationship of candidate dominant degradation mechanisms are analyzed. For example, the direct contribution ratio of ion migration to defect propagation and the synergistic enhancement effect of electric field and humid heat environment on grain boundary degradation are evaluated to determine the dominant degradation mechanism and its mechanism contribution information, such as "electric field driven ion migration-defect aggregation synergistic dominant degradation mechanism", and its dominant contribution to the overall performance degradation is output.

[0057] For example, taking the mechanism analysis results of a complete 150-hour degradation process of perovskite photovoltaic materials under the coupled degradation conditions of light-humid heat-bias voltage as an example, the pre-set integrity condition of mechanism analysis is verified. First, in the identification results of the dominant degradation mechanism, for example, the "electric field-driven ion migration-defect aggregation synergistic dominant mechanism" is obtained. Its corresponding structural evolution characteristics can explain about 90% of the sources of photoelectric conversion efficiency degradation, such as the carrier recombination enhancement and interface state density increase caused by the migration of grain boundary I vacancies. Therefore, its cumulative performance degradation explanation rate is 90%, satisfying the first sub-condition of not less than 85%. Second, in the identification results of the key degradation stage, for example, 85h to 120h is identified as the dominant performance degradation stage. This stage accounts for about 40% of the entire 150-hour degradation cycle. Combined with other secondary abrupt change stages (such as the early defect initiation stage around 30h), the cumulative proportion of time covering the key degradation behavior reaches about 92%, satisfying the second sub-condition of not less than 90%. Furthermore, during the multimodal matching process, the confidence level of mechanism identification is obtained by comprehensively calibrating the structural matching similarity, defect evolution consistency, and performance curve fitting degree. For example, the confidence level of the dominant mechanism reaches 0.87, and all candidate mechanisms are higher than the 0.80 threshold, thus satisfying the third sub-condition. Subsequently, during three consecutive supplementary acquisitions and re-analysis processes, since the newly added observation data only caused minor adjustments to local parameters, the dominant mechanism remained stable as the "electric field-driven ion migration-defect aggregation synergistic mechanism," and the change in the contribution scores of each mechanism was less than the set tolerance range, for example, the change did not exceed 3%, thus satisfying the fourth stability sub-condition. Finally, when all four sub-conditions are satisfied, the mechanism analysis integrity condition is established, thus confirming that the current dominant degradation mechanism identification result has high interpretability, high coverage, and high stability, and can be used as a reliable output result for the degradation mechanism analysis of this perovskite photovoltaic material.

[0058] The dominant degradation mechanism identification method described in this application constructs a candidate degradation mechanism library covering structural evolution patterns, key defect characteristics, and performance degradation behavior. This library structurally encodes prior knowledge of known degradation mechanisms, providing high-quality reference prototypes for mechanism matching. By performing multimodal joint matching of key defect structures, key degradation stages, and performance degradation results, the matching degree of the mechanism prototype is comprehensively evaluated using multidimensional degradation information, avoiding misidentification caused by single-dimensional matching. Furthermore, contribution analysis and synergy analysis are used to further determine the dominant degradation mechanism and its synergistic relationships, outputting mechanism contribution information. This achieves accurate and interpretable identification of photovoltaic material degradation mechanisms.

[0059] In some embodiments, the preset mechanism resolution integrity conditions include the following four sub-conditions, and all sub-conditions must be met simultaneously for the integrity conditions to be determined: The first sub-condition is that the cumulative performance degradation explanation rate corresponding to the identified dominant degradation mechanism is not less than 85%, and the explanation rate is defined as the ratio of the performance degradation that can be explained by the structural evolution characteristics of the identified mechanism to the total performance degradation; the second sub-condition is that the coverage rate of key degradation stages is not less than 90%, that is, the ratio of the sum of the durations of key degradation stages on the entire life cycle time axis to the total degradation duration is not less than the threshold; the third sub-condition is that the identification confidence of the dominant degradation mechanism is not less than 0.80, and the confidence is obtained by calibrating the multimodal similarity matching score; the fourth sub-condition is that the identification result of the dominant degradation mechanism remains stable during the three consecutive iterations of supplementary acquisition and re-resolution, and the stability is determined by the fact that the change in the mechanism type set and the contribution score of each mechanism in the two adjacent identification results does not exceed the set tolerance.

[0060] For example, taking the complete degradation analysis process of perovskite photovoltaic materials under the coupled degradation conditions of light-humid heat-bias voltage for 150 hours as an example, the completeness condition of mechanism analysis is verified. First, in the results of identifying the dominant degradation mechanism, for example, the "electric field-driven ion migration-defect aggregation synergistic mechanism" is identified as the main degradation mechanism, and its structural evolution characteristics (such as I...) are analyzed. -Migration path formation, grain boundary vacancy chain expansion, and interface defect enrichment can explain approximately 88%–92% of the photoelectric conversion efficiency degradation. Therefore, the cumulative performance degradation explanation rate is approximately 0.90, which is higher than the 85% threshold, satisfying the first sub-condition. Secondly, in the analysis results of the critical degradation stages, for example, 30–40h was identified as the early defect initiation stage, 85–120h as the rapid degradation stage, and 120–150h as the stable degradation stage. Among them, the critical degradation stage (the 85–120h stage, which mainly contributes to the degradation) accounts for approximately 40h of the total time. After combining multi-stage weighted statistics, it cumulatively covers approximately 92% of the effective degradation behavior time range. Therefore, the critical degradation stage coverage rate meets the requirement of not less than 90%, satisfying the second sub-condition. Furthermore, during the multimodal mechanism matching process, the structural evolution characteristics, defect propagation modes, and performance degradation curves are jointly similarized. For example, the confidence level for identifying the dominant mechanism is 0.86, and the confidence levels for each secondary candidate mechanism are not lower than the calibration threshold of 0.80, thus satisfying the third sub-condition. Subsequently, in three consecutive supplementary acquisitions and re-analysis iterations, since the new data only caused minor adjustments to local parameters, the dominant degradation mechanism remained stable as the "electric field-driven ion migration-defect aggregation synergistic mechanism," and the mechanism set did not change in adjacent results. The change in the contribution score of each mechanism was less than the set tolerance (e.g., ≤3%), thus satisfying the fourth stability sub-condition. Finally, when all four sub-conditions are satisfied simultaneously, the mechanism analysis integrity condition is determined to be met, thus confirming that the current dominant degradation mechanism analysis results meet the consistency requirements in terms of interpretability, coverage, confidence, and stability.

[0061] The aforementioned pre-defined mechanism parsing integrity conditions in this application comprehensively evaluate the quality of mechanism parsing from multiple perspectives, including comprehensiveness of explanation, temporal coverage, identification reliability, and result stability, by simultaneously setting four quantitative sub-conditions: performance degradation explanation rate, key degradation stage coverage, identification confidence, and iterative stability. By requiring all four sub-conditions to be met simultaneously, it ensures that the mechanism parsing results are free from quality defects such as obvious omissions, temporal blind spots, low-confidence identification, or unstable results. Furthermore, when any sub-condition is not met, a data supplementation and iterative re-parsing process is triggered, thereby effectively guaranteeing the completeness, reliability, and accuracy of the final degradation mechanism parsing results.

[0062] In some embodiments, the method for supplementing the collection of degradation process data under corresponding operating conditions is as follows: If the integrity condition of mechanism parsing is not met, the data supplementation and collection process for the degradation process will be initiated. Kernel density estimation is performed on the distribution density of existing degradation process data in the operating condition parameter space to identify low-density operating condition areas with insufficient data coverage. By combining the prediction uncertainty of the photovoltaic material structure-performance relationship model in each operating condition region, the low-density operating condition region is evaluated to determine the data gap operating condition region; A supplementary data acquisition plan is generated based on the data gap operating condition area. The supplementary data acquisition plan includes the target operating condition parameter range, sampling time step and minimum sampling quantity. The collection priority is determined based on the prediction uncertainty and data scarcity corresponding to the data gap operating condition area, and supplementary collection is performed according to the collection priority to obtain the newly added degradation process data.

[0063] For example, taking the initial 150-hour data collection of perovskite photovoltaic materials under coupled light-humidity-bias degradation conditions as an example, the supplementary data collection process is explained. First, during the mechanism analysis, it was found that the integrity conditions of the current "electric field-driven ion migration-defect aggregation synergistic mechanism" were not fully met. For example, the boundary of the key degradation stage still showed unstable identification in some high-humidity, high-bias conditions. Therefore, the supplementary data collection process for the degradation process was initiated. Second, kernel density estimation analysis was performed on the existing 150-hour degradation data in the operating parameter space (light intensity, temperature, humidity, and bias combination space). It was found that the data point density in the medium-high humidity (>90%RH) and high bias (>0.9V) regions was significantly low, forming low-density coverage areas, indicating insufficient samples of degradation behavior under these conditions. Furthermore, combined with the prediction uncertainty assessment results of the structural performance relationship model in this operating condition region, such as the significantly increased variance in performance degradation prediction and unstable defect density growth prediction, this region was identified as a data gap operating condition region, i.e., a key region simultaneously exhibiting "data scarcity + high model uncertainty". Subsequently, a supplementary data acquisition scheme was generated based on the data gap operating condition range. For example, the target operating condition range was set as 85–95℃, 90–95%RH, and 0.9–1.0V bias voltage range. The sampling time step was shortened to less than 1 hour, and the minimum number of samples was required to cover at least 30 time points to enhance the continuity and observability of the degradation process in this region. Finally, the acquisition priority was determined according to the uncertainty level and data scarcity of this operating condition range. For example, the high bias voltage + high humidity region was set as the highest priority, and supplementary acquisition experiments were performed first to obtain new degradation process data for subsequent retraining of the mechanism analysis model and verification of the integrity conditions.

[0064] The degradation process data supplementation method described in this application accurately locates low-density operating condition areas with insufficient data coverage by estimating the kernel density of existing data distribution density in the operating condition parameter space, thus avoiding resource waste caused by blind supplementation. By combining the prediction uncertainty of the photovoltaic material structure-performance relationship model to comprehensively evaluate the low-density operating condition areas, both data scarcity and model uncertainty are incorporated into the data gap judgment, improving the targeting of data supplementation. On this basis, the collection priority is determined according to the prediction uncertainty and the degree of data scarcity, and supplementation is performed according to the priority, thereby rapidly improving the satisfaction of the mechanism analysis integrity conditions with minimal supplementation cost and improving the overall analysis efficiency.

[0065] In some embodiments, the method for generating the material modification direction is as follows: The material degradation suppression targets are determined based on the key defect structure, key degradation stage, and dominant degradation mechanism; Inverse optimization is performed based on a multi-scale structural feature and photovoltaic material structure-performance relationship model to obtain the target structural feature configuration. Based on the target structural characteristics, corresponding material composition adjustment schemes, defect control schemes, and interface optimization schemes are generated. Based on the photovoltaic material structure-performance relationship model, the performance improvement effect of each scheme is predicted, and the performance improvement effect is obtained. The feasibility, process compatibility, and implementation cost of each scheme are evaluated by combining the pre-set process knowledge base, and the process feasibility evaluation results are obtained. Based on the performance improvement effect and process feasibility assessment results, the various schemes are comprehensively ranked to obtain the recommended results for material modification directions.

[0066] For example, taking the degradation analysis results of perovskite photovoltaic materials under irradiation-humid heat-bias coupling degradation conditions for 150 hours as an example, the generation process of material modification direction is explained. First, based on the identified key defect structures (such as the chain expansion of grain boundary I⁻ vacancies and enrichment of interface defects), key degradation stages (such as the rapid decay stage from 85 to 120 hours), and dominant degradation mechanisms (such as the electric field-driven ion migration-defect aggregation synergistic mechanism), the material degradation suppression targets are determined, such as suppressing the I⁻ migration rate, reducing the probability of grain boundary defect formation, and enhancing the interface charge passivation capability. Second, based on the multi-scale structural characteristics (atomic-scale coordination stability, grain-scale uniformity, and interface-scale carrier transport characteristics) and the structural performance relationship model, inverse optimization is performed to obtain the configuration of target structural characteristics, such as reducing the I⁻ vacancy formation energy, improving the average grain size uniformity, and enhancing the interface energy level matching degree, through gradient inversion or constraint optimization. Furthermore, multiple candidate material modification schemes are generated based on the target structural characteristics. For example, in terms of material composition, a small amount of Cs / FA ratio is introduced to improve lattice stability; in terms of defect control, halogen enrichment inhibitors are introduced to reduce I-vacancy concentration; and in terms of interface optimization, a 2D passivation layer is introduced to reduce interfacial recombination centers, thus forming a combination of material composition adjustment schemes, defect control schemes, and interface optimization schemes. Subsequently, the performance of each scheme is predicted based on a structure-performance relationship model. For example, it is predicted that after introducing an interface passivation layer, the photoelectric conversion efficiency can be increased from 18.5% to over 20%, the defect density can be reduced by about 30%, and the Voc and Jsc decay trends can be improved simultaneously, thus obtaining the corresponding performance improvement effects of each scheme. Then, the feasibility of each scheme is evaluated in conjunction with a pre-set process knowledge base. For example, the 2D passivation layer scheme is evaluated to have high process compatibility in low-temperature solution preparation, while some highly doped composition schemes may introduce phase separation risks and increase preparation costs, thus obtaining the process feasibility evaluation results of each scheme. Finally, the performance improvement effect and the process feasibility assessment results are comprehensively ranked. For example, the combination of "interface passivation + appropriate component regulation" is given priority to achieve the optimal balance between performance improvement and preparation feasibility, thereby generating the recommended results for material modification direction.

[0067] The method for generating the aforementioned material modification direction in this application determines the degradation suppression target based on the key defect structure, key degradation stage, and dominant degradation mechanism, thus making the generation of modification direction specifically targeted at the degradation mechanism. By performing reverse optimization on the multi-scale structural features and structural performance relationship model, the configuration of target structural features is derived in reverse from the performance improvement target, directly coupling structural design and performance improvement. By combining a preset process knowledge base, the preparation feasibility, process compatibility, and implementation cost of material composition adjustment schemes, defect control schemes, and interface optimization schemes are comprehensively evaluated, and the schemes are ranked comprehensively according to the performance improvement effect and process feasibility, thereby providing a high-quality recommended scheme for photovoltaic material modification that takes into account both performance improvement effect and engineering feasibility.

[0068] In some embodiments, the method for generating structural optimization suggestions is as follows: When the dominant degradation mechanism is thermal stress-induced degradation, optimization suggestions for grain interface stress buffer structure are generated, and the corresponding grain size distribution control parameters are determined. When the dominant degradation mechanism is the hydrothermal-induced ion migration degradation mechanism, optimization suggestions for ion migration inhibition structures are generated, and the corresponding ion barrier layer design parameters are determined. When the dominant degradation mechanism is the illumination-induced carrier recombination degradation mechanism, optimization suggestions for defect passivation structures are generated, and the corresponding defect energy level matching parameters are determined. The performance improvement effect of each structural optimization suggestion is evaluated based on the photovoltaic material structure-performance relationship model.

[0069] For example, taking the dominant degradation mechanism and structure-performance relationship model identified under the coupled degradation conditions of light-humidity-temperature-bias voltage in perovskite photovoltaic materials as an example, the process of generating structural optimization suggestions is explained. First, when the dominant degradation mechanism is determined to be "thermal stress-induced lattice distortion and grain boundary mismatch mechanism," based on the characteristics of thermal stress concentration and grain boundary crack propagation during the degradation process, optimization suggestions for grain interface stress buffer structures are generated. For example, by controlling the grain size distribution (increasing the average grain size from 200 nm to the 300–400 nm range and reducing the size variance), stress dispersion ability is enhanced and the risk of grain boundary cracking is suppressed. Second, when the dominant degradation mechanism is "humidity-temperature-induced ion migration mechanism," optimization suggestions are generated for I under high humidity and high temperature conditions. -Accelerated migration and grain boundary enrichment phenomena lead to structural optimization suggestions for ion migration suppression, such as introducing ion blocking layers or interface densification structures, and determining corresponding design parameters, such as blocking layer thickness, densification layer coverage, and diffusion barrier height, to reduce ion migration rates and defect accumulation. Furthermore, when the dominant degradation mechanism is "photoinduced carrier recombination," defect passivation structural optimization suggestions are generated to address the increased interface state density and enhanced non-radiative recombination under photoexcitation. For example, introducing surface passivation molecules or low-dimensional passivation layers to regulate defect energy level matching relationships can distance defect energy levels from carrier energy levels, thereby reducing recombination probability and improving carrier lifetime. Finally, a unified evaluation of the above different structural optimization suggestions is conducted based on a photovoltaic material structure-performance relationship model. For example, it is predicted that grain size optimization can improve efficiency by approximately 5%, ion blocking layer optimization can reduce efficiency decay by approximately 30%, and defect passivation can significantly improve Voc loss. This quantifies the performance improvement effects of each structural optimization path, providing a basis for the final optimization scheme selection.

[0070] The structural optimization suggestion generation method described in this application generates differentiated structural optimization suggestions for three dominant degradation mechanisms: thermal stress-induced degradation, damp heat-induced ion migration degradation, and light-induced carrier recombination degradation. This achieves a precise correspondence between degradation mechanisms and optimization strategies, avoiding the problem of insufficient specificity in generalized optimization suggestions. By separately determining the grain size distribution control parameters, ion barrier layer design parameters, and defect energy level matching parameters, the structural optimization suggestions are concretized into quantitative design parameters that can directly guide material preparation. Based on this, a photovoltaic material structure-performance relationship model is used to predict and evaluate the performance improvement effect corresponding to each optimization suggestion, providing quantitative support for the selection of optimization suggestions, thereby significantly improving the specificity, operability, and engineering guidance value of the structural optimization suggestions.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing the structural-performance relationship of photovoltaic materials based on graph neural networks, characterized in that, Includes the following steps: Data on the degradation process of photovoltaic materials under multi-condition coupling are obtained, including multi-scale structural snapshot sequences and corresponding condition parameter information. Feature extraction is performed on the multi-scale structural snapshot sequence to obtain the structural feature representation results corresponding to each structural snapshot. The structural feature representation results are then fused with the corresponding operating condition parameter information to obtain the operating condition adaptive molecular diagram sequence. The working condition adaptive molecular graph sequence is input into the defect perception graph neural network for structural characterization, extracting multi-scale structural features that characterize the degradation behavior of photovoltaic materials, and organizing the multi-scale structural features in chronological order to form the corresponding temporal feature sequence. Multi-timescale temporal coding and correlation modeling are performed on the temporal feature sequences to obtain temporal state features characterizing the degradation and evolution process of photovoltaic materials; Based on multi-scale structural features and temporal state features, a structural performance relationship model for photovoltaic materials is constructed. Based on the photovoltaic material structure-performance relationship model, the photoelectric conversion efficiency degradation rate, open-circuit voltage drift and short-circuit current degradation of photovoltaic materials are predicted, and the corresponding performance degradation results are obtained. Based on performance degradation results and multi-scale structural characteristics, we identify the key defect structures, key degradation stages, and dominant degradation mechanisms that affect performance degradation during the degradation process of photovoltaic materials.

2. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 1, characterized in that, The method for feature extraction from multi-scale structural snapshot sequences is as follows: Based on the spatial distribution information of atoms in the multi-scale structural snapshot sequence, a multi-scale local neighborhood is constructed for each atom, and local chemical environment descriptors at different neighborhood scales are extracted. Scale fusion is performed on local chemical environment descriptors to obtain atomic features that characterize the local environment of atoms; Based on the interatomic interaction relationships in the structural snapshot, chemical bond type features, bond length distribution features, bond angle distribution features, and bond dynamic evolution features are extracted to obtain chemical bond structural features that characterize structural connection relationships. Defects are identified based on the structural deviation between the structural snapshot and the reference crystal structure, and defect density features, defect spatial distribution features, and defect interaction network features are extracted to obtain defect features. By correlating and fusing atomic features, chemical bond structure features, and defect features, structural feature characterization results corresponding to each structural snapshot are obtained.

3. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 1, characterized in that, The method for obtaining the adaptive molecular map sequence is as follows: Feature encoding is performed on the operating condition parameter information to obtain the corresponding operating condition features; Based on the operating condition characteristics, the atomic features are adaptively mapped and enhanced to obtain the operating condition adaptive atomic features; Based on the characteristics of operating conditions, the chemical bond structure characteristics and defect characteristics are dynamically adjusted and enhanced to obtain operating condition adaptive structural characteristics. By using the working condition adaptive atomic features as node attributes and the working condition adaptive structural features as edge attributes and graph-level attributes, a working condition adaptive molecular graph corresponding to each structural snapshot is constructed. The adaptive molecular diagrams for each operating condition are arranged according to the time sequence of the structural snapshots to form a sequence of adaptive molecular diagrams for each operating condition.

4. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 1, characterized in that, The method for extracting multi-scale structural features is as follows: Based on the node embedding vectors output by each layer of the defect-aware graph neural network, the atomic local scale features are obtained by reading out and splicing them layer by layer. Hierarchical pooling is performed based on node embedding vectors, and nodes are clustered and aggregated according to chemical environment similarity to obtain grain-scale features; Based on the set of interface nodes in the adaptive molecular diagram, feature difference information of nodes on both sides of the interface is extracted to obtain the device interface scale features. A multi-scale structural feature representation is obtained by fusing and encoding atomic local scale features, grain scale features, and device interface scale features; By jointly encoding the multi-scale structural feature representation with the corresponding structural hierarchy label information, we obtain multi-scale structural features that characterize the degradation behavior of photovoltaic materials.

5. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 1, characterized in that, The method for multi-timescale temporal coding and correlation modeling of temporal feature sequences is as follows: Parallel convolutional encoding of temporal feature sequences is performed based on causal temporal convolutional channels at multiple different time scales to extract temporal evolution features at the corresponding time scales, thereby obtaining temporal encoded features at multiple time scales. The time-series coding features corresponding to each time scale are normalized to obtain normalized time-series coding features; The normalized temporal coding features are weighted and fused based on the channel attention mechanism to obtain multi-timescale temporal coding vectors. Global temporal dependencies are established based on multi-timescale temporal coding vectors to obtain global temporal association features; Temporal state features are generated based on global temporal correlation features to characterize local time-varying patterns and global evolution trends during the degradation process.

6. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 1, characterized in that, The method for obtaining the performance degradation results is as follows: Obtain the performance characterization features output by the photovoltaic material structure-performance relationship model; Based on performance characterization features, we establish branches for predicting photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation. Based on the photoelectric conversion efficiency attenuation rate prediction branch, a predicted value for the photoelectric conversion efficiency attenuation rate is generated. Based on the open-circuit voltage drift prediction branch and the short-circuit current degradation prediction branch, the predicted values ​​of open-circuit voltage drift and short-circuit current degradation are generated respectively. Consistency verification and joint constraint analysis are performed on the predicted values ​​of photoelectric conversion efficiency degradation rate, open-circuit voltage drift, and short-circuit current degradation to generate corresponding performance degradation results.

7. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 1, characterized in that, The method for identifying critical defect structures is as follows: Based on the performance degradation results and multi-scale structural features, the performance contribution of each structural feature node is determined. Candidate defect sites that are highly correlated with performance degradation are selected based on their performance contribution. Establish a causal relationship between candidate defect sites and performance degradation results to identify key defect sites that have a significant impact on performance degradation. Based on the defect attribute information, spatial distribution information and structural hierarchy information of key defect sites, the key defect sites are aggregated and characterized to generate key defect structure identification results.

8. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 7, characterized in that, The method for identifying critical degradation stages is as follows: The cosine distance between temporal features at adjacent time points is calculated based on the temporal feature sequence to obtain the degradation behavior change sequence. Statistical analysis of the degradation behavior change sequence is performed based on a sliding time window, and an adaptive change threshold is generated; The moment of abrupt change in the degradation state is identified by comparing the sequence of changes in degradation behavior with the adaptive change threshold, and the moment of abrupt change in the degradation state is used as the boundary of the degradation stage. The temporal feature sequence is segmented based on the degradation stage boundary to obtain multiple degradation stages, and the stage representative features corresponding to each degradation stage are extracted. The correlation strength between the representative characteristics of each stage and the photoelectric conversion efficiency attenuation rate, open-circuit voltage drift, and short-circuit current degradation was calculated respectively. Based on the correlation strength, each degradation stage is screened to determine the key degradation stage, and its corresponding start and end times and cumulative performance loss percentage are output.

9. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to claim 8, characterized in that, The method for identifying the dominant degradation mechanism is as follows: The pre-built candidate degradation mechanism library is retrieved, and the key defect structure, key degradation stage and performance degradation result corresponding to the current degradation process are obtained; the candidate degradation mechanism library includes multiple degradation mechanism prototypes jointly characterized by structural evolution mode, key defect characteristics and performance degradation behavior; The key defect structure, key degradation stage and performance degradation results are matched with each degradation mechanism prototype in the candidate degradation mechanism library in a multimodal manner to obtain the comprehensive matching score of each degradation mechanism prototype. Based on the comprehensive matching score, the prototypes of each degradation mechanism are ranked, and candidate dominant degradation mechanisms that meet the preset conditions are selected. Contribution analysis and synergy analysis are performed on the candidate dominant degradation mechanisms to determine the dominant degradation mechanism and output the corresponding mechanism contribution information.

10. The method for analyzing the relationship between the structure and performance of photovoltaic materials according to any one of claims 1-9, characterized in that, Also includes: Determine whether the dominant degradation mechanism meets the preset mechanism parsing integrity condition; If not, supplementary degradation process data under the corresponding working conditions will be collected, and the structural characterization, temporal modeling and degradation analysis processes will be re-executed until the integrity condition of mechanism analysis is met. If the conditions are met, then based on the identified key defect structures, key degradation stages, and dominant degradation mechanisms, corresponding material modification directions and structural optimization suggestions will be generated.

Images