An artificial intelligence core self-assembled monolayer interface passivation layer closed-loop development system

Abstract

The application discloses a self-assembled monomolecular interface passivation layer closed-loop development system taking artificial intelligence as a core, and the system comprises a feature engineering and data set construction module, which is used for constructing a feature vector data set containing molecular electronic structure, physical and chemical properties and field knowledge based on an interface action mechanism; an artificial intelligence prediction model module, which is used for training a machine learning model to predict photovoltaic device performance; a virtual screening and molecular design module, which is used for generating candidate molecules based on a fragment library and predicting and sorting performance by using the model; a theoretical calculation verification module, which is used for performing first-principle calculation verification on the interface performance feasibility of screened molecules; and an experimental verification and feedback iteration module, which is used for preparing and testing the verified molecules, and feeding back the results to a data set iteration optimization model. The system constructed by the application realizes efficient, accurate and interpretable design and verification of the interface material, and significantly improves the research and development efficiency and device performance.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic materials and devices technology, and in particular relates to a closed-loop development system for self-assembled monomolecular interface passivation layers based on artificial intelligence. Background Technology

[0002] Perovskite solar cells, with their excellent solution processability, low-temperature fabrication technology, and outstanding photoelectric performance, have become one of the most competitive candidates for next-generation photovoltaic technology. The laboratory-certified efficiency of their single-junction devices has exceeded 27%, demonstrating enormous industrialization potential. The key to achieving this high performance lies in precise interface engineering. Among these strategies, employing self-assembled monolayers to modulate energy levels and passivate defects at the device interface is a core approach to simultaneously improve efficiency and stability.

[0003] Self-assembled monomolecules typically consist of three parts: a functional head group that interacts with the perovskite active layer, an inorganic transport layer (such as NiO), and an inorganic transport layer (such as NiO). x The structure-matrix interface (SAM) consists of an anchoring group (or ITO) bonded to the surface, and a bridging group connecting the two. Currently, molecular design for these three parts has developed certain standardized strategies: for example, introducing Lewis acid / base, ammonium salt, and other functional groups into the functional head group to enhance passivation; adjusting the flexibility and conjugation length of the bridging group to optimize molecular arrangement and charge transport; and selecting phosphonic acid groups, silane groups, etc., as anchoring groups to improve interfacial bonding strength. However, facing the vast chemical space composed of head groups, bridging groups, and anchoring groups, traditional research paradigms heavily rely on researchers' experience and chemical intuition, requiring numerous and repetitive trial-and-error experiments to screen candidate molecules. This method is not only time-consuming and costly, but also makes it difficult to systematically explore and verify deep-seated structure-property relationships, becoming a major bottleneck hindering the large-scale discovery and application of high-performance SAM materials.

[0004] In recent years, the development of machine learning technology has provided a new paradigm for efficiently exploring the space of materials chemistry. However, most existing research focuses on building and optimizing single predictive models, often establishing a "black box" relationship between molecular structure and final performance, lacking a complete R&D system that deeply integrates underlying physicochemical mechanisms, efficient intelligent computing, and reliable experimental verification. How to build a systematic intelligent platform that can integrate mechanism guidance, autonomously conduct efficient virtual screening, and rapidly verify and self-optimize through experimental closed-loop is a core challenge that urgently needs to be addressed to achieve rational design and rapid development of SAM materials. Summary of the Invention

[0005] To address the aforementioned technical challenges, this invention provides a closed-loop development system for self-assembled single-molecule interface passivation layers based on artificial intelligence. This system constructs a complete workflow of "mechanism-guided modeling—intelligent virtual screening—first-principles verification—experimental synthesis and testing," enabling the efficient and automated discovery and verification of optimal SAM molecules from a vast chemical space. Ultimately, this system is applied to the fabrication of high-performance, high-stability perovskite solar cells.

[0006] This invention provides a closed-loop development system for self-assembled single-molecule interface passivation layers based on artificial intelligence, comprising the following steps: The feature engineering and dataset construction module is used to extract and construct feature vectors containing molecular electronic structure, physicochemical properties and domain knowledge rules from historical data based on the interfacial interaction mechanism of self-assembled single molecules, forming a standardized dataset for model training. The artificial intelligence prediction model module is connected to the feature engineering and dataset construction module. It is used to train a machine learning model based on the standardized dataset and predict the corresponding photovoltaic device performance indicators based on the input molecular feature vector after training the machine learning model. The virtual screening and molecular design module is connected to the artificial intelligence prediction model module. It is used to generate candidate molecules based on the modular fragment library, call the trained machine learning model to predict and rank the performance of the candidate molecules, and output a list of screened candidate molecules. The theoretical calculation and verification module is connected to the virtual screening and molecular design module and is used to perform first-principles calculations or simulations on the screened candidate molecule list to verify its theoretical feasibility in forming a dense and ordered monolayer and optimizing interface performance. The experimental verification and feedback iteration module is connected to the theoretical calculation verification module and the feature engineering and dataset construction module, respectively. It is used to perform experimental preparation and performance testing on the theoretically verified candidate molecules, and feed the obtained experimental results as new data back to the feature engineering and dataset construction module to update the standardized dataset and iteratively optimize the machine learning model.

[0007] Optionally, the feature engineering and dataset construction module includes: The data collection unit is used to collect historical data including the molecular structure and electronic properties of SAM and the corresponding photovoltaic device performance parameters. First-principles calculation unit, used to obtain at least one electronic structure descriptor of frontier orbital energy level, dipole moment and polarizability of a molecule through density functional theory calculation. The chemical descriptor extraction unit is used to extract physicochemical descriptors from molecular structures using cheminformatics tools. The rule feature encoding unit is used to construct and encode rule features related to interface passivation and energy level matching based on functional group type and substrate material knowledge.

[0008] Optionally, the artificial intelligence prediction model module includes: The model training unit is used to train a machine learning model with the feature vectors in the standardized dataset as input and the photoelectric conversion efficiency of the photovoltaic device as the output target. The model application unit is used to input the feature vectors of the candidate molecules to be predicted into the trained machine learning model and output the performance prediction values.

[0009] Optionally, the machine learning model is a model built based on a predictive neural network for tabular data.

[0010] Optionally, the virtual screening and molecular design module includes: The fragment library management unit is used to manage modular fragment libraries containing head functional groups, bridging groups, and anchoring groups; Molecular splicing units are used to splice fragments selected from different fragment libraries according to chemical rules to generate complete candidate molecular structures; The feature mapping and prediction unit is used to calculate feature vectors for the complete candidate molecular structures and call the machine learning model to perform batch prediction and sorting.

[0011] Optionally, the theoretical calculation verification module is configured to perform at least one of the following calculations: The adsorption configuration, adsorption energy, and interfacial electronic structure of candidate molecules on the target substrate were calculated using density functional theory. Molecular dynamics simulations were used to model the self-assembly behavior and monolayer order of candidate molecules at interfaces.

[0012] Optionally, the experimental verification and feedback iteration module includes: The synthesis and characterization unit is used to synthesize target SAM molecules and characterize their film-forming properties and their effects on the morphology and defects of perovskite films. The device fabrication and testing unit is used to fabricate photovoltaic devices containing SAM layers and test their photoelectric conversion efficiency and stability parameters. The data feedback unit is used to add the test results and their corresponding molecular characteristic data to the standardized dataset after standardization processing.

[0013] On the other hand, this invention also proposes a closed-loop development method for self-assembled monomolecular interface passivation layers based on artificial intelligence, executed according to the aforementioned closed-loop development system, the method comprising: Based on the interfacial interaction mechanism of self-assembled single molecules, a feature vector integrating molecular electronic structure, physicochemical properties and domain rules is constructed to form a standardized dataset. A machine learning model is trained using the standardized dataset to obtain a machine learning model that can predict the performance indicators of corresponding photovoltaic devices based on molecular features. Based on the trained machine learning model, candidate molecules generated from the modular fragment library are virtually screened and sorted. First-principles calculations and / or simulations were performed on the selected candidate molecules to theoretically verify their feasibility in forming a dense, ordered monolayer and optimizing interface properties. Candidate molecules that have passed theoretical verification are experimentally prepared and their performance is tested to obtain actual photovoltaic device performance data; The actual performance data and its corresponding molecular features are fed back as new data to the standardized dataset and used to update the machine learning model, thereby achieving closed-loop iterative optimization.

[0014] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0015] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention achieves a deep integration of data-driven approaches and physical mechanisms by constructing a "mechanism-guided feature engineering and dataset construction module." This module actively constructs a feature set based on the interfacial interaction mechanism of self-assembled single molecules, replacing traditional feature engineering methods that rely on purely statistical correlations. The resulting standardized dataset not only provides high-quality input to the model but also ensures the consistency between the model's learning process and the underlying physicochemical principles, fundamentally enhancing the model's interpretability and generalization ability, and giving the prediction results a solid scientific basis.

[0017] This invention achieves efficient and precise exploration of a vast chemical space by deploying an "artificial intelligence prediction model module." This module, trained using the aforementioned mechanistic dataset, generates a high-performance prediction model capable of rapidly and accurately predicting device performance based on molecular characteristics. This technique makes it possible to identify a very small number of high-potential targets from tens of thousands of candidate molecules, significantly improving the efficiency and accuracy of molecular screening and shifting materials discovery from "trial and error" to "prediction."

[0018] This invention constructs a computationally-driven rational design verification chain by integrating a "virtual screening and molecular design module" and a "theoretical calculation and verification module." The system first uses a predictive model for large-scale virtual screening, followed by first-principles calculations to verify the selected molecules. This coherent technical approach enables the theoretical evaluation and verification of molecular interfacial properties (such as adsorption capacity and energy level matching) before investing expensive experimental resources, significantly reducing the blind spots and risks of failure in research and development, and providing in-depth mechanistic insights.

[0019] This invention achieves self-evolution and continuous optimization of the R&D process by establishing an "experimental verification and feedback closed-loop module." This module automatically feeds back and updates the system's dataset with actual performance data obtained from experimental synthesis and device testing, for iterative optimization of the prediction model. This technique transforms the system from a one-time screening tool into an intelligent R&D platform capable of continuously learning and evolving from the real world, gradually accumulating domain knowledge and improving long-term performance.

[0020] This invention organically connects the five modules mentioned above to form a complete automated closed-loop workflow, achieving a revolution in the research and development paradigm. The collaborative operation between the modules covers the entire chain from mechanism analysis, intelligent design, theoretical verification to experimental feedback. This system-level integration completely changes the fragmented, manual, and slow-iterative situation in traditional materials research and development, providing a highly automated, intelligent, and scalable new paradigm that is not only applicable to the development of interface materials but can also be extended to the design and discovery of other functional molecules. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the artificial intelligence-driven closed-loop molecular design according to an embodiment of the present invention, wherein (a) is an overview of the overall system workflow; and (b) is a schematic diagram of the feature extraction process and machine learning principle. Figure 2The figures shown are the dataset features, model development and validation results of the AI-assisted screening framework in this embodiment of the invention. Among them, (a) is a comparison of the efficiency distribution of the control group and SAM-modified devices matched with literature; (b) is a distribution of device efficiency gain (ΔPCE); (c) is a heatmap of the correlation between key features calculated based on density functional theory (DFT); (d) is a comparison of the prediction performance of six different regression models; (e) is a calibrated scatter plot of the prediction efficiency and measured efficiency of the TabPFN model, showing its unbiased prediction characteristics; (g) is a schematic diagram of the construction of the virtual fragment library and molecular feature mapping, presenting the combination of head base-bridging-anchoring base and unified feature space; and (h) is a schematic diagram of the two-dimensional chemical structure of MTP-S, the best candidate molecule screened by the AI ​​model. Figure 3 This is a comparative analysis of first-principles calculations (DFT) performed on the preferred molecule (MTP-S) and the conventional molecule (MeO-4PACz) in this embodiment of the invention. (a) is a comparison of the ground-state geometry of the two molecules, showing their differences in planarity; (b) isosurface plots of non-covalent interactions (NCI) between molecular dimers and a comparison of binding energies; (c) isosurface plots of the highest occupied molecular orbitals (HOMO) of the two molecules, showing the differences in the degree of electron cloud delocalization; (d) isosurface plots of the electrostatic potential (ESP) and electronic localization function (ELF) of the molecules; and (e) isosurface plots of the molecules in NiO. x A schematic diagram of the theoretical adsorption configuration at the substrate surface and the interface of the buried perovskite. Figure 4 These are multi-scale characterization images of perovskite (PVK) thin films based on different SAMs (MeO-4PACz and MTP-S) in embodiments of the present invention. Among them, (a) is a pseudo-color spectrum of in-situ photoluminescence (PL) intensity changing with time during the one-step spin-coating process of the perovskite precursor solution; (b) is a two-dimensional grazing-incidence wide-angle X-ray scattering (GIWAXS) pattern of the perovskite thin film after film formation; (c) and (d) are top-view and cross-sectional scanning electron microscope (SEM) images of the perovskite thin film, respectively; (e) is a comparison of time-resolved photoluminescence (TRPL) decay curves of perovskite thin films based on different SAMs; (f) is a steady-state photoluminescence (PL) spectrum of the perovskite thin film; and (g) is an X-ray diffraction (XRD) spectrum of the perovskite thin film based on MeO-4PACz at different tilt angles (Ψ). Figure 5 NiO in the embodiments of the present invention x The interface morphology, surface potential, and energy level arrangement of the / SAM / perovskite (PVK) stacked structure are shown in the figure. Among them, (a) shows NiO modified with different SAMs. x (a) and (c) are top-view SEM images of NiO with different SAM modifications, respectively. x(d) Kelvin probe force microscopy (KPFM) distribution of contact potential difference (CPD) on the surface and the perovskite buried interface above it; (e) Secondary electron cutoff edge and valence band spectra measured by ultraviolet photoelectron spectroscopy (UPS), used to extract work function (WF) and valence band peak (VBM) information; (f) UPS data plotted along NiO in the MeO-4PACz and MTP-S systems. x →SAM→PVK energy level arrangement diagram; Figure 6 The figures show the device performance, charge transport kinetics, and stability test results of inverted perovskite solar cells (PSCs) using different SAMs in the embodiments of the present invention. Among them, (a) is a schematic diagram of the inverted device structure; (b) is the current density-voltage (JV) characteristic curve of the champion devices based on MTP-S and MeO-4PACz with small area (0.043 cm²) and large area (1.05 cm²) under standard illumination (AM1.5G); (c) and (d) are the transient photocurrent (TPC) decay curves of the device under short-circuit conditions and the transient photovoltage (TPV) decay curves under open-circuit conditions, respectively; (e) is the Nyquist plot of the electrochemical impedance spectroscopy (EIS) of the device under illumination and bias voltage; (f) is the long-term stability test curve of the device under continuous operation of maximum power point tracking (MPPT) (nitrogen environment) and dark storage (high humidity air environment), showing the change of normalized efficiency over time. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0024] Example 1 like Figure 1 As shown, this embodiment provides a closed-loop development system for self-assembled single-molecule interface passivation layers based on artificial intelligence, including the following steps: The feature engineering and dataset construction module is used to extract and construct feature vectors containing molecular electronic structure, physicochemical properties and domain knowledge rules from historical data based on the interfacial interaction mechanism of self-assembled single molecules, forming a standardized dataset for model training. The artificial intelligence prediction model module is connected to the feature engineering and dataset construction module. It is used to train a machine learning model based on the standardized dataset and predict the corresponding photovoltaic device performance indicators based on the input molecular feature vector after training the machine learning model. The virtual screening and molecular design module is connected to the artificial intelligence prediction model module. It is used to generate candidate molecules based on the modular fragment library, call the trained machine learning model to predict and rank the performance of the candidate molecules, and output a list of screened candidate molecules. The theoretical calculation and verification module is connected to the virtual screening and molecular design module and is used to perform first-principles calculations or simulations on the screened candidate molecule list to verify its theoretical feasibility in forming a dense and ordered monolayer and optimizing interface performance. The experimental verification and feedback iteration module is connected to the theoretical calculation verification module and the feature engineering and dataset construction module, respectively. It is used to perform experimental preparation and performance testing on the theoretically verified candidate molecules, and feed the obtained experimental results as new data back to the feature engineering and dataset construction module to update the standardized dataset and iteratively optimize the machine learning model.

[0025] The feature engineering and dataset construction module, which is feasible, includes: The data collection unit is used to collect historical data including the molecular structure, electronic properties, and corresponding photovoltaic device performance parameters of SAM; the first-principles calculation unit is used to obtain at least one electronic structure descriptor among the frontier orbital energy levels, dipole moments, and polarizabilities of the molecule through density functional theory calculations; the chemical descriptor extraction unit is used to extract physicochemical descriptors from the molecular structure using cheminformatics tools; and the rule feature encoding unit is used to construct and encode rule features related to interface passivation and energy level matching based on functional group types and substrate material knowledge.

[0026] As a specific implementation method, the implementation process of the feature engineering and dataset construction module includes: The feature engineering implemented in this embodiment is based on a mechanism-guided active design paradigm, which completely changes the traditional passive analysis mode that relies on post-hoc data statistical correlation. Specifically, when constructing the input features for the machine learning model, this embodiment uses the physicochemical interaction mechanism of self-assembled single molecules (SAMs) in photovoltaic interfaces as a clear theoretical framework to pre-generate an initial feature set. This framework systematically integrates three dimensions: based on molecular orbital theory, it calculates frontier orbital energy levels, dipole moments, and other electronic descriptors through density functional theory (DFT) to directly quantify key physical processes such as "energy level alignment" and "interface dipoles"; based on surface and interface interaction theory, it uses DFT calculations and rule encoding to describe the adsorption, anchoring strength, and defect passivation potential of molecules on the substrate; and simultaneously, it combines molecular structure analysis to extract features such as bridge segment rigidity and conjugation length to characterize monolayer order and charge transport continuity. In subsequent feature selection and optimization, this embodiment does not simply rely on statistical indicators, but prioritizes retaining those features with clearer physical meaning and stronger correlation with device efficiency (PCE), thereby ensuring that the final feature space is highly consistent with the deep mechanism assumptions, significantly improving the interpretability and generalization ability of the model.

[0027] Specifically: The system collects historical literature and experimental data from 2016 to 2025 to construct a high-quality standardized dataset on the molecular structure, electronic properties, and final photovoltaic device performance (such as photoelectric conversion efficiency) of SAM.

[0028] Key electronic characteristics of molecules (such as HOMO / LUMO levels, dipole moments, and polarizability) are obtained through density functional theory calculations. To achieve interpretable prediction of molecular device performance, this embodiment selects frontier orbital levels (HOMO / LUMO), dipole moments (μ), and polarizability (α) obtained from DFT calculations as key electronic structure descriptors. These characteristics correspond to the molecule's electron-donating / accepting ability, interface potential modulation ability, and dielectric shielding response ability, respectively, covering the main physical processes in the device, such as "level alignment—interface charge transfer—Coulomb interaction shielding." Specifically, HOMO / LUMO directly determines the level matching between the electrode or adjacent functional layers, thereby affecting the carrier injection / extraction barrier and the thermodynamic driving force of interface charge transfer, and thus regulating the interface charge transport efficiency and device output indicators (such as mobility / turn-on voltage / current density / efficiency, etc.). The dipole moment may form a dipole layer at the interface and cause vacuum level shift, effectively changing the effective work function of the electrode and the injection barrier, while also affecting the local electric field and electrostatic disorder. Polarizability reflects a molecule's response to an external electric field and its local dielectric shielding ability. It can affect the Coulomb interaction between charges and the CT state / exciton binding energy, thereby influencing charge separation, recombination loss, and transport processes. Based on these physical relationships, these DFT features provide low-dimensional, transferable, and interpretable inputs for machine learning models, helping to improve the reliability and interpretability of device performance predictions.

[0029] Use cheminformatics tools (such as RDKit) to extract physicochemical descriptors of molecules.

[0030] By incorporating domain knowledge, key descriptors based on functional groups and substrate materials are introduced to form a comprehensive feature vector, used to characterize the intrinsic relationship between molecular structure, properties, and performance. The "incorporation of domain knowledge" described in this embodiment employs a reproducible feature engineering process: first, a candidate descriptor pool is established based on device physical mechanisms (energy level alignment, interface charge transfer / recombination, interface chemical passivation, and film morphology, etc.); second, functional groups and substrate-related factors that make stable contributions to the target performance are identified through literature evidence and data-driven screening (correlation and interpretability analysis). Functional group descriptors are automatically identified by molecular SMILES using substructure rules (SMARTS / fragment dictionary) and quantified using methods such as existence / counting / normalized counting; substrate factors are numerically characterized by one-hot encoding of substrate categories. Furthermore, this embodiment constructs derived features directly corresponding to the mechanisms to explicitly characterize the impact of energy level alignment and injection barriers on interface charge transport efficiency. The final feature vector is constructed by concatenating DFT electronic properties, functional groups, substrate features, and derived mechanism features.

[0031] Implementable, the artificial intelligence prediction model module includes: The model training unit is used to train a machine learning model with the feature vectors in the standardized dataset as input and the photoelectric conversion efficiency of the photovoltaic device as the output target; the model application unit is used to input the feature vectors of the candidate molecules to be predicted into the trained machine learning model and output the performance prediction value.

[0032] Furthermore, the machine learning model is a model built based on a predictive neural network for tabular data.

[0033] As one specific implementation method, the implementation process of the artificial intelligence prediction model module includes: The feature engineering and dataset construction module is responsible for constructing and summarizing repeatable and physically interpretable features. Based on the DFT, it obtains DM / Gap / Polar / HOMO / LUMO features, and combines them with RDKit structure / polarity descriptors and physicochemical properties to form a "feature template" with fixed field names and order, outputting a structured feature vector / table. The AI ​​prediction model module strictly follows this template during input, performing column name matching and order rearrangement, and checking dimensional consistency (if a field is missing, it is back-calculated or the sample is removed), thus ensuring complete consistency between the output of the feature engineering and dataset construction module and the model input. To address dimensional differences, the AI ​​prediction model module uses standardization / normalization based on training set statistics for continuous features before training, and employs consistent encoding for categorical features (such as BA, CA). To address multicollinearity, the AI ​​prediction model module performs Pearson correlation screening before training, removing redundant features with correlation coefficients >0.85 and retaining only low-correlation features for modeling, thereby reducing the risk of multicollinearity and overfitting. Therefore, the preprocessing (format validation / encoding / standardization / relevance filtering) that is strongly correlated with the model input is uniformly categorized into the artificial intelligence prediction model module, while the feature engineering and dataset construction module only outputs "raw and comprehensive" feature templates and values.

[0034] Specifically, a regression prediction model is constructed using the aforementioned features as input and device performance parameters (such as PCE) as output targets.

[0035] This paper employs an advanced tabular data prediction neural network (such as TabPFN) as its core algorithm. This model demonstrates excellent prediction accuracy and generalization ability in small-sample, high-dimensional tabular data, and requires no cumbersome hyperparameter tuning, achieving "plug and play". This embodiment chooses TabPFN instead of MLP or other neural networks primarily based on the data characteristics of this problem—"medium-sized, structured molecular / DFT descriptors + significant cross-document noise"—and the need for interpretability. On the same dataset, this embodiment uses Grid Search + 10-fold CV to uniformly compare KNN / SVM / MLP / RF / XGBoost / TabPFN. TabPFN achieves higher and more consistent accuracy on both the training and test sets (94.2% on the test set), demonstrating better robustness and generalization ability. Furthermore, as a gradient boosting tree ensemble model, TabPFN can automatically learn higher-order nonlinear and interactive relationships between descriptors through tree splitting (without relying on larger data volumes and more complex structures like MLP).

[0036] A pre-trained artificial intelligence prediction model is used as the core screening tool: the standardized feature vectors of candidate SAM molecules are input into the prediction model, which outputs the predicted values ​​of the target performance parameters and performs rapid screening of candidate molecules accordingly. The prediction model is a model that has been verified to have superior prediction accuracy, calibration, and generalization ability compared to the comparison model on the regression prediction task of the standardized dataset. In the preliminary preparation for model selection, this embodiment divides the training and test sets into 90% and 10% sets, respectively. Models such as KNN, SVM (SVR), MLP, RF, XGBoost, and TabPFN are optimized and ranked using GridSearch combined with 10-fold cross-validation. The test set R² and the average R² of 10-fold cross-validation are used as the main indicators, and the difference between training and test performance is compared to assess the risk of overfitting and model calibration. Finally, the model that is optimal in all the above indicators and has the closest training / test accuracy is selected as the core screening model for unified prediction and ranking in the subsequent combined library.

[0037] The model can autonomously identify key molecular descriptors (such as rigidity, dipole orientation, and conjugation length) from data, training the model's "chemical intuition" to guide molecular design. This "autonomous identification" does not involve manually writing chemical rules or relying on specific prior structures. Instead, it trains the TabPFN regression model on structure-property input features constructed using DFT+RDKit+functional group / base information (retaining 32 features after correlation deredundancy removal). Since TabPFN is a neural network model, it cannot perform SHAP analysis. Instead, the model learns the electronic / structural characteristics and physicochemical properties of molecules, such as the relationship between "molecular rigidity / conformational freedom" (e.g., number of rotatable bonds, aromatic / conjugation correlation indices), "dipole orientation" (e.g., dipole moment calculated by DFT and its component along the interface normal), and "conjugation length / charge delocalization" (e.g., π-conjugation fragment length, HOMO delocalization correlation descriptor) and PCE, thus forming the model's "chemical intuition." These key descriptors are then used to constrain and guide the combination and prioritization of molecular fragments: for example, preferentially selecting more rigid π-bridged frameworks to reduce bending, introducing polar end groups / heterogeneous atoms to regulate dipole orientation and improve interfacial work function matching, and moderately extending conjugation to enhance HOMO delocalization and lateral transport. Based on this, candidates that meet the above trends are screened in the head-bridged-anchor fragment library (ultimately corresponding to the rigid conjugated framework and favorable dipole component of MTP-S, and further verified by DFT and experiments for its advantages in dense assembly, energy level alignment and defect passivation).

[0038] Implementable, the virtual screening and molecular design module includes: The fragment library management unit manages modular fragment libraries containing head functional groups, bridging groups, and anchoring groups; the molecular splicing unit splices fragments selected from different fragment libraries according to chemical rules to generate complete candidate molecular structures; and the feature mapping and prediction unit calculates feature vectors for the complete candidate molecular structures and calls the machine learning model for batch prediction and sorting.

[0039] As a specific implementation, the virtual screening and molecular design module is a downstream step after the "model finalization" of the artificial intelligence prediction model module: when the artificial intelligence prediction model module completes training / cross-validation and reaches a preset accuracy on the validation (or test) set (in this embodiment, TFPN is ultimately determined as the main model for subsequent screening, and the test set R²=0.942), the pipeline script can automatically call this module to carry out high-throughput virtual screening; the human operator is mainly responsible for setting the threshold and reviewing the top candidates before they enter DFT / experimental validation. Regarding the fragment library, this embodiment organizes head-bridge-anchor fragments based on literature evidence and syntheticity constraints and performs combined screening. The system integrates computable constraints on structural connectivity and chemical legitimacy (such as consistency of connection sites, valence / structure sanitization, etc.); the "synthetic rules" are not hard-coded into the model or screener in the form of reaction templates such as SMARTS / Reaction SMARTS, so as to maintain the data-driven characteristic of "no manually encoded chemical rules".

[0040] Specifically, this embodiment constructs a modular fragment library of head functional groups, bridging groups, and anchoring groups based on chemical feasibility and synthesis rules.

[0041] The fragment combinations are mapped to a unified feature space, and the performance of all candidate molecules in the combination library is predicted using a trained prediction model. "Mapping fragment combinations to a unified feature space" is not simply adding up the fragment descriptors; in this embodiment, the three structures are first spliced ​​into complete candidate molecules (SMILES / structures) according to predefined connection sites based on the "head-bridge-anchor" fragment library. Then, feature vectors are calculated for each complete molecule using a feature engineering process identical to that of the training set: including 8 key electronic / structural descriptors obtained from DFT, 14 chemical and physical descriptors extracted from RDKit, and 14 regular features based on functional group and substrate type (consistent with the redundancy removal / screening process used in the model training phase). Fragment identity information is incorporated into the feature vector using one-hot encoding, rather than numerical summation; the nonlinear coupling between fragments is naturally reflected by DFT / RDKit descriptors at the whole-molecule level (such as dipole moments, frontier orbitals, etc.). Thus, 64 4×4×4 combination molecules can be mapped to a unified feature space consistent with the training data and directly input into the trained model for prediction.

[0042] Based on the prediction results, the molecules are ranked to quickly identify several candidate molecules with the highest potential performance, achieving efficient trimming and convergence of the chemical space. This embodiment adopts a "two-step screening + ranking and locking" strategy: First, a head-base-bridging-anchoring fragment library is constructed based on literature evidence and syntheticity, while deliberately preserving chemical diversity to avoid premature convergence; after mapping this combined library to a unified feature space, a total of 64 realizable candidate molecules (4×4×4) are obtained. Subsequently, the PCE of all candidates is predicted and ranked one by one. The "locking in several" of the shortlist adopts a hard constraint (syntheticity / adsorption / substrate adaptation, etc.) + Top-N approach: under the premise of satisfying the hard constraints, the top few molecules in the ranking are preferentially selected for expensive DFT and experimental verification; within the resource-allowed range, this embodiment further focuses on the Top1 (MTP-S, predicted PCE 26.12%) for DFT and experimental closed-loop verification. The balance between exploration and utilization is reflected in the fact that the front end achieves "exploration" through diverse fragment libraries and combination design, while the back end achieves "utilization" through sorting and optimization. If it is extended to multiple rounds of iteration, a small number of candidates with "high uncertainty / farther feature space" can be added in addition to the Top-N in each round for verification and backflow training, thereby reducing the risk of getting trapped in local optima from a mechanism perspective.

[0043] Implementably, the theoretical calculation verification module is configured to perform at least one of the following calculations: The adsorption configuration, adsorption energy, and interfacial electronic structure of candidate molecules on the target substrate were calculated using density functional theory; the self-assembly behavior and monolayer order of candidate molecules on the interface were simulated using molecular dynamics.

[0044] As a specific implementation method, the top candidate molecules screened by AI are subjected to systematic first-principles calculations (such as DFT) and molecular dynamics simulations.

[0045] The calculations cover: molecular geometry, electronic structure (HOMO / LUMO distribution, electrostatic potential), intermolecular interaction energy, and on the target substrate (such as NiO). x Adsorption configuration and binding energy, self-assembly behavior simulation, etc. on ).

[0046] The predictive model primarily uses PCE as the evaluation target, and its judgments on interface stability, self-assembly density, and energy level matching are mostly implicit inferences. Therefore, DFT / MD calculations are introduced to use physical quantities such as adsorption binding energy (E_ads), multidentate coordination configuration, intermolecular binding energy, and monolayer coverage / order as external interpretable evidence to verify the mechanism and correct the risks of the AI ​​screening results. When necessary, these physical quantities are fed back as features or auxiliary tasks to achieve model recalibration. This embodiment theoretically verifies the advantages of candidate molecules in forming dense, ordered monolayers, achieving energy level matching, enhancing interfacial bonding, and passivation effects, providing solid mechanistic support for subsequent experiments. Before entering the experiment, this embodiment employs a two-tiered theoretical verification standard of "preset hard threshold + dynamic sorting": candidate molecules must first meet the following conditions: (i) possessing the structural and stacking conditions for forming a dense, ordered monolayer (rigid / near-planar main framework, stronger intermolecular stacking, and significantly higher coverage density than the baseline under the same surface area); (ii) having stable multi-point anchoring / coordination with the substrate and defect sites, selecting molecules with stronger adsorption energy or defect binding ability as experimental targets. Molecules screened by the machine learning model in this embodiment exhibit stronger adsorption and self-assembly capabilities than traditional MeO-4PACz molecules in the industry. Those meeting the above hard thresholds are considered "verified" and proceed to the experiment.

[0047] Implementable, the experimental verification and feedback iteration module includes: The synthesis and characterization unit is used to synthesize the target SAM molecule and characterize its film-forming properties and its influence on the morphology and defects of perovskite films; the device fabrication and testing unit is used to fabricate photovoltaic devices containing SAM layers and test their photoelectric conversion efficiency and stability parameters; the data feedback unit is used to add the test results and their corresponding molecular characteristic data to the standardized dataset after standardization processing.

[0048] As a specific implementation method, to ensure the stability and comparability of closed-loop learning, this embodiment only includes results obtained under completely identical experimental conditions in the feedback phase: the same molecule is repeatedly prepared and tested using the same process flow and testing protocol (such as substrate / treatment, solution concentration and spin coating parameters, annealing conditions, electrodes, device area and test light source calibration, etc.), and the obtained device performance is used as the training label by the statistical summary value of the repeated samples (such as the average value, and the corresponding dispersion is recorded); then the standardized experimental results are written back to the dataset for updating / retraining the model, thereby realizing the experimental verification-data feedback-model iteration closed loop of "same conditions, reproducible".

[0049] Specifically, the optimal candidate molecules are synthesized based on AI predictions and computational verification.

[0050] The system characterizes the film-forming properties, optical properties, wettability of the SAM layer and its effects on the crystallization kinetics, morphology, crystal structure, defect states, and stress of perovskite thin films.

[0051] Complete perovskite solar cell devices were fabricated, and their photovoltaic performance parameters (JV curve, EQE, etc.) and operational stability under different stress conditions were tested.

[0052] The new performance data obtained from the experiments are fed back into the dataset for iterative optimization of the AI ​​prediction model, forming a self-evolving and continuously improving closed-loop system. This embodiment adopts batch closed-loop iteration instead of online learning: the new PCE / ΔPCE data obtained in each round of experiments are first subjected to consistency quality control (condition / device structure alignment, outlier removal), and then the same set of DFT+RDKit+structure descriptor features are regenerated according to the established process and added to the dataset; when the new samples accumulate to a certain scale or cover a new chemical subspace, a full-process refit / retraining is triggered (the TabFPN is remodeled using the updated training set in a "no parameter tuning" manner), and 85 / 15 independent test sets + training set 10-fold cross-validation is used to evaluate whether there is true gain and whether there is bias drift. To avoid the addition of small samples from damaging generalization ability, this embodiment maintains the same feature cleaning and redundancy removal strategy (removing highly relevant features to suppress overfitting and improve generalization), while performing stratification / weight control on the new samples and continuously retaining low-performance "hard controls" to reduce selection bias and stabilize the model's discrimination boundary; if the performance degrades on the frozen test set after the update, it will be rolled back to the previous data / model version.

[0053] Based on the complete closed-loop system constructed in this embodiment, which consists of "mechanism-guided modeling, intelligent virtual screening, theoretical calculation verification, experimental synthesis and testing, and data feedback iteration," the following significant technical effects have been achieved: 1. Achieving high-efficiency and low-cost molecular discovery: By employing artificial intelligence prediction models to replace traditional trial-and-error experiments, this system can perform high-throughput virtual screening of thousands or even tens of thousands of candidate molecules generated from modular fragment libraries. This process can be completed within days, greatly shortening the development cycle of new materials and significantly reducing the high human and material costs associated with relying on numerous synthesis and characterization experiments.

[0054] 2. Achieve rational design with high precision and predictability: Based on a high-quality feature dataset constructed using mechanism guidance and an advanced tabular data prediction neural network model (such as TabPFN), this system achieves accurate prediction of the performance of SAM molecules. The model achieves a prediction accuracy of R²>0.94 on the independent test set and is unbiased in the high-performance range. This transforms molecular design from an experience-dependent process into a quantifiable and predictable scientific process, as evidenced by the high agreement between the prediction efficiency (26.12%) and the efficiency of the experimental champion (26.5%) in the examples.

[0055] 3. Provides in-depth chemical insights with transparent mechanisms: The mechanism-guided feature engineering and model interpretability analysis employed in this system can surpass traditional "black box" predictions, proactively identifying and revealing key molecular structural factors (such as rigid π frameworks and specific functional groups) that affect device performance. This not only validates the prediction results but also provides in-depth chemical design principles and mechanistic insights, guiding subsequent molecular optimization directions.

[0056] 4. Complete the research and development of end-to-end integration and automated closed-loop system: By seamlessly integrating multiple previously isolated processes such as data collection, feature engineering, AI modeling, virtual screening, theoretical verification, experimental testing, and data feedback, this invention constructs an automated and intelligent complete R&D closed loop. This system achieves end-to-end connectivity and self-iterative capabilities from "design" to "verification" to "optimization," fundamentally changing the traditional R&D model.

[0057] 5. Produce outstanding and proven application results: Novel SAM molecules (such as MTP-S), discovered and validated through the aforementioned closed-loop system, have achieved a champion photoelectric conversion efficiency exceeding 26% and excellent operational stability in perovskite solar cells. This breakthrough performance demonstrates the powerful ability and practical effectiveness of this closed-loop system in solving real-world technical problems.

[0058] 6. Pioneering a new paradigm for scalable intelligent R&D: This invention provides not only a solution for SAM materials, but also a general framework integrating mechanisms, data, and experiments. This core framework can be extended to the design and development of other functional molecules such as organic semiconductors, luminescent materials, and catalysts, providing a transformative and intelligent new paradigm for research and development in the field of materials science.

[0059] On the other hand, this invention also proposes a closed-loop development method for self-assembled monomolecular interface passivation layers based on artificial intelligence, executed according to the aforementioned closed-loop development system, the method comprising: Based on the interfacial interaction mechanism of self-assembled single molecules, a feature vector integrating molecular electronic structure, physicochemical properties and domain rules is constructed to form a standardized dataset. A machine learning model is trained using the standardized dataset to obtain a machine learning model that can predict the performance indicators of corresponding photovoltaic devices based on molecular features. Based on the trained machine learning model, candidate molecules generated from the modular fragment library are virtually screened and sorted. First-principles calculations and / or simulations were performed on the selected candidate molecules to theoretically verify their feasibility in forming a dense, ordered monolayer and optimizing interface properties. Candidate molecules that have passed theoretical verification are experimentally prepared and their performance is tested to obtain actual photovoltaic device performance data; The actual performance data and its corresponding molecular features are fed back as new data to the standardized dataset and used to update the machine learning model, thereby achieving closed-loop iterative optimization.

[0060] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0061] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0062] Example 2 This embodiment fully demonstrates the specific application process of the self-assembled single-molecule interface passivation layer closed-loop development system based on artificial intelligence. It successfully designed and verified a high-performance SAM molecule (MTP-S) and applied it to perovskite solar cells, achieving significant performance improvement.

[0063] Specifically, the following steps are included: Data collection and feature engineering: First, system initialization and data preparation were performed. Over 10,000 standardized data pairs were collected from 320 publicly available papers. Each data pair contained a clearly defined SAM molecular structure, corresponding photovoltaic device structural information, and performance indicators such as photoelectric conversion efficiency (PCE). Subsequently, a mechanism-guided feature engineering process was initiated: for each molecule, eight key electronic features were calculated using density functional theory (DFT), 14 physicochemical descriptors were extracted using the cheminformatics tool RDKit, and 14 additional rule descriptors were introduced based on functional group knowledge and basis type. Finally, a standardized dataset containing 36 input features was constructed to provide high-quality, interpretable input for model training. The efficiency distribution and feature correlation of the dataset are shown below. Figure 2 As shown.

[0064] AI Model Training and Virtual Screening: Using the standardized dataset described above, an artificial intelligence prediction model was trained. The TabPFN model was employed, with the data divided into training and test sets at a ratio of 85% and 15%, respectively, and 10-fold cross-validation was performed to optimize model robustness. After training, the model achieved a prediction accuracy of R²=0.942 on the independent test set. Subsequently, a virtual screening stage was initiated: a candidate library of 64 virtual molecules was generated based on a pre-built modular fragment library. The feature vectors of the candidate molecules were input into the trained AI model for batch prediction, and the molecules were sorted according to their predicted PCE values. The screening results showed that the molecule 5-(methylthio)-1,3,4-benzothiadiazole-2-ylphosphonic acid (MTP-S) had the highest prediction efficiency (26.12%) and was therefore selected as the priority target for validation.

[0065] First-principles calculation verification: Before proceeding with the costly experimental phase, in-depth first-principles calculations were performed to verify MTP-S, the top candidate molecule selected by AI. DFT calculations showed that the MTP-S molecule possesses a near-planar rigid structure, which is conducive to its formation of a dense and ordered monolayer on the substrate surface. Its dimer binding energy was calculated to be -1.81 eV, significantly higher than that of the commonly used comparative molecule MeO-4PACz (-1.24 eV), indicating stronger intermolecular interactions and better film-forming stability. Electronic structure analysis revealed that its HOMO orbitals are delocalized throughout the molecular framework, which is beneficial for the lateral transport of interfacial charges. Further simulations showed that MTP-S in NiO... x A higher coverage and more ordered monolayer structure can be formed on the substrate. These theoretical results strongly support the predictions of the AI ​​model at the mechanistic level, confirming the potential advantages of MTP-S. Figure 3 As shown, DFT calculations indicate that the MTP-S molecule possesses a near-planar rigid structure, stronger intermolecular binding energy, and delocalized HOMO orbitals in NiO. x It exhibits a superior adsorption configuration on the substrate.

[0066] Experimental synthesis, characterization, and device performance testing: Based on the positive results of theoretical verification, experimental verification was conducted. First, the MTP-S molecule was successfully synthesized, and its SAM layer was systematically characterized. Contact angle tests showed that the MTP-S-modified NiO... xThe surface water contact angle increased to 78°, indicating the formation of a more hydrophobic and dense film. In-situ spectroscopic analysis confirmed that MTP-S effectively extended the crystallization window of the perovskite precursor solution, which is beneficial for the growth of high-quality films. Grazing incidence wide-angle X-ray scattering (GIWAXS) and scanning electron microscopy (SEM) characterization revealed that the MTP-S-based perovskite film had better crystal orientation, larger grain size, and fewer pores. X-ray photoelectron spectroscopy (XPS) confirmed the formation of Pb-S coordination at the interface, achieving effective defect passivation. Time-resolved fluorescence spectroscopy (TRPL) showed that the average carrier lifetime significantly increased from 426 ns to 761 ns, indicating that nonradiative recombination at the interface was effectively suppressed. Figure 4 As shown, the perovskite thin films based on MTP-S exhibit superior crystal orientation, larger grain size, and significantly extended carrier lifetime. Figure 5 As shown, interface analysis confirms that it can effectively regulate the surface potential and achieve ideal energy level alignment.

[0067] Subsequently, inverted perovskite solar cell devices were fabricated. The champion device based on MTP-S SAM, with a small area (0.043 cm²), achieved a photoelectric conversion efficiency of 26.83% (third-party certified efficiency 26.54%), while the large-area (1.05 cm²) device also achieved an efficiency of 25.33%. Stability tests showed that after 2000 hours of maximum power point tracking (MPPT) operation in a nitrogen environment, the device still maintained 93.8% of its initial efficiency; after 1200 hours in an air environment with 60% relative humidity, the efficiency retention rate was 91.9%, demonstrating excellent operational and environmental stability.

[0068] Data feedback and model iteration: To complete the closed loop and enhance the system's continuous evolution capability, the complete structural information, all computational features, and excellent device performance data (PCE=26.83%, stability data, etc.) of the MTP-S molecules obtained in this experiment were added as new samples after standardization. Subsequently, the AI ​​prediction model was retrained using the updated database. Through iteration, the model's learning of characteristic patterns of high-performance SAM molecules with rigid conjugated backbones and specific functional groups was enhanced, further improving its future virtual screening accuracy and generalization ability, thus completing a complete "prediction-verification-feedback-optimization" closed-loop iteration.

[0069] Ultimately, as Figure 6 As shown, the champion device based on MTP-S achieved a photoelectric conversion efficiency exceeding 26% and exhibited excellent long-term operational stability. See Table 1 for 0.043 cm⁻¹. 2 The performance parameters of the area-specific solar cell are shown in Table 2, which is 1.05 cm².2 Performance parameters of area-based batteries.

[0070] Table 1 Table 2 This embodiment, through the full-process application of a closed-loop system, not only efficiently discovered and verified the high-performance SAM molecule MTP-S, achieving simultaneous improvement in the efficiency and stability of perovskite solar cells, but also demonstrated the effectiveness and advancement of the system in deeply integrating artificial intelligence, theoretical calculation, and experimental research and development, providing a complete solution for the intelligent development of functional molecular materials.

[0071] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A closed-loop development system for self-assembled monomolecular interface passivation layers based on artificial intelligence, characterized in that, Includes the following steps: The feature engineering and dataset construction module is used to extract and construct feature vectors containing molecular electronic structure, physicochemical properties and domain knowledge rules from historical data based on the interfacial interaction mechanism of self-assembled single molecules, forming a standardized dataset for model training. The artificial intelligence prediction model module is connected to the feature engineering and dataset construction module. It is used to train a machine learning model based on the standardized dataset and predict the corresponding photovoltaic device performance indicators based on the input molecular feature vector after training the machine learning model. The virtual screening and molecular design module is connected to the artificial intelligence prediction model module. It is used to generate candidate molecules based on the modular fragment library, call the trained machine learning model to predict and rank the performance of the candidate molecules, and output a list of screened candidate molecules. The theoretical calculation and verification module is connected to the virtual screening and molecular design module and is used to perform first-principles calculations or simulations on the screened candidate molecule list to verify its theoretical feasibility in forming a dense and ordered monolayer and optimizing interface performance. The experimental verification and feedback iteration module is connected to the theoretical calculation verification module and the feature engineering and dataset construction module, respectively. It is used to perform experimental preparation and performance testing on the theoretically verified candidate molecules, and feed the obtained experimental results as new data back to the feature engineering and dataset construction module to update the standardized dataset and iteratively optimize the machine learning model.

2. The closed-loop development system according to claim 1, characterized in that, The feature engineering and dataset construction module includes: The data collection unit is used to collect historical data including the molecular structure and electronic properties of SAM and the corresponding photovoltaic device performance parameters. First-principles calculation unit, used to obtain at least one electronic structure descriptor of frontier orbital energy level, dipole moment and polarizability of a molecule through density functional theory calculation. The chemical descriptor extraction unit is used to extract physicochemical descriptors from molecular structures using cheminformatics tools. The rule feature encoding unit is used to construct and encode rule features related to interface passivation and energy level matching based on functional group type and substrate material knowledge.

3. The closed-loop development system according to claim 1, characterized in that, The artificial intelligence prediction model module includes: The model training unit is used to train a machine learning model with the feature vectors in the standardized dataset as input and the photoelectric conversion efficiency of the photovoltaic device as the output target. The model application unit is used to input the feature vectors of the candidate molecules to be predicted into the trained machine learning model and output the performance prediction values.

4. The closed-loop development system according to claim 3, characterized in that, The machine learning model is a model built based on a predictive neural network for tabular data.

5. The closed-loop development system according to claim 1, characterized in that, The virtual screening and molecular design module includes: The fragment library management unit is used to manage modular fragment libraries containing head functional groups, bridging groups, and anchoring groups; Molecular splicing units are used to splice fragments selected from different fragment libraries according to chemical rules to generate complete candidate molecular structures; The feature mapping and prediction unit is used to calculate feature vectors for the complete candidate molecular structures and call the machine learning model to perform batch prediction and sorting.

6. The closed-loop development system according to claim 1, characterized in that, The theoretical calculation verification module is configured to perform at least one of the following calculations: The adsorption configuration, adsorption energy, and interfacial electronic structure of candidate molecules on the target substrate were calculated using density functional theory. Molecular dynamics simulations were used to model the self-assembly behavior and monolayer order of candidate molecules at interfaces.

7. The closed-loop development system according to claim 1, characterized in that, The experimental verification and feedback iteration module includes: The synthesis and characterization unit is used to synthesize target SAM molecules and characterize their film-forming properties and their effects on the morphology and defects of perovskite films. The device fabrication and testing unit is used to fabricate photovoltaic devices containing SAM layers and test their photoelectric conversion efficiency and stability parameters. The data feedback unit is used to add the test results and their corresponding molecular characteristic data to the standardized dataset after standardization processing.

8. A closed-loop development method for self-assembled monomolecular interface passivation layers based on artificial intelligence, characterized in that, Based on the closed-loop development system according to any one of claims 1 to 7, the method includes: Based on the interfacial interaction mechanism of self-assembled single molecules, a feature vector integrating molecular electronic structure, physicochemical properties and domain rules is constructed to form a standardized dataset. A machine learning model is trained using the standardized dataset to obtain a machine learning model that can predict the performance indicators of corresponding photovoltaic devices based on molecular features. Based on the trained machine learning model, candidate molecules generated from the modular fragment library are virtually screened and sorted. First-principles calculations and / or simulations were performed on the screened candidate molecules to theoretically verify their feasibility in forming a dense, ordered monolayer and optimizing interface properties. Candidate molecules that have passed theoretical verification are experimentally prepared and their performance is tested to obtain actual photovoltaic device performance data; The actual performance data and its corresponding molecular features are fed back as new data to the standardized dataset and used to update the machine learning model, thereby achieving closed-loop iterative optimization.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, The processor implements the method of claim 8 when executing the computing program.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of claim 8.

Images