A machine learning method for high-throughput screening of cuprous iodide cluster scintillators and applications thereof

Abstract

The application provides a machine learning high-throughput screening method for copper iodide cluster scintillators and application thereof, and mainly overcomes the defects of low prediction accuracy of a single machine learning model, dimension explosion of multi-feature splicing, and easy overfitting of a small data set. The screening method specifically comprises the following steps: collecting Cu(I)-I cluster crystals from a crystal structure database to construct a multi-dimensional feature data set; converting organic ligands into SMILES strings to batch generate multi-class molecular fingerprints; extracting ligand atomic bonding graph features by using a directed message passing graph neural network, eliminating redundant features and performing standardization and normalization; performing feature compression on the molecular fingerprints by using a one-dimensional convolution layer and a global maximum pooling layer, and fusing the molecular fingerprints with graph features and descriptors into a unified low-dimensional comprehensive feature vector; selecting an isomorphic base learner and a meta-learner to construct a Stacking stacking ensemble model, and obtaining an optimal prediction model through hyperparameter optimization; and inputting the features of candidate materials, and screening target high-performance scintillators according to the prediction values.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical fields of radiation detection functional materials, high-throughput screening of machine learning materials, and flexible optoelectronic imaging devices. Specifically, it relates to a method for high-throughput screening of cuprous iodide cluster scintillators using machine learning and its application. Background Technology

[0002] X-ray scintillators are core optoelectronic functional materials used in medical radiological imaging, industrial non-destructive testing, and public security inspection. Their core function is to convert ionizing radiation into visible light signals; imaging sensitivity, resolution, and irradiation stability directly determine the performance of the equipment. Currently, commercially available mainstream scintillation single crystals include lead tungstate (PbWO4) and bismuth germanate (Bi3Ge4O). 12 There are three types of lanthanum cerium doped with cerium (PbWO4), PbWO4 contains the toxic heavy metal lead and has extremely low light yield, requiring a high radiation dose during imaging; PbWO4 single crystals must be prepared by high-temperature pulling, which is costly, brittle, and difficult to process flexibly; and LaBr3:Ce is extremely hygroscopic, has a complex packaging process, and has scarce raw materials.

[0003] While emerging metal halide perovskite scintillators can be prepared in solution, their luminescence performance degrades significantly under water-oxygen environments and continuous X-ray irradiation, and lead-based systems also pose environmental risks. Cuprous iodide clusters (Cu(I)-I), with their high X-ray absorption capacity, near 100% luminescence quantum efficiency, lead-free and non-toxic nature, and low-temperature synthesis capabilities, have become highly promising candidates for next-generation scintillator materials. However, the combination space for inorganic copper-iodine cores and organic ligands is vast, and traditional trial-and-error synthesis screening processes are lengthy and extremely costly, making it difficult to rapidly obtain high-performance target materials. Existing machine learning material prediction techniques have significant limitations: using only a single type of feature input, direct splicing of multi-source molecular, crystal, and optical features easily leads to dimensionality explosion and model overfitting; using only a single regression algorithm results in insufficient prediction accuracy; the highest determination coefficient of existing models for scintillator output is only 0.78, indicating limited generalization ability and inability to achieve high-precision, high-throughput screening. In addition, existing flexible scintillation films mostly use perovskite or rare earth powders as luminescent fillers, which have problems such as dynamic imaging ghosting, low spatial resolution, easy degradation of fillers by irradiation, and difficulty in large-area uniform film formation. They cannot meet the actual needs of industrial online dynamic flaw detection and low-dose medical dynamic imaging.

[0004] Therefore, developing a high-precision machine learning method for screening cuprous iodide cluster scintillators has significant industrial value for high-performance cuprous iodide cluster scintillators and high-resolution, shadow-free flexible X-ray imaging devices. Summary of the Invention

[0005] The primary objective of this invention is to overcome the shortcomings of existing single machine learning models, such as low prediction accuracy, dimensionality explosion from multi-feature splicing, and overfitting on small datasets. This invention provides a high-throughput screening method for cuprous iodide cluster scintillators using machine learning. Specifically, it is a high-throughput screening method for cuprous iodide iodide cluster scintillators based on CNN feature fusion and Stacking ensemble learning, which significantly shortens the new material development cycle and accurately predicts the X-ray light output performance of candidate materials.

[0006] The second objective of this invention is to provide a high-performance cuprous iodide scintillation crystal Cu6I8(bu-ted)2, which is selectively discovered using the above-mentioned screening method, to solve the problems of low light yield, high detection limit, toxicity, and poor stability of existing scintillators.

[0007] The third objective of this invention is to provide a flexible PDMS composite X-ray imaging film using the scintillation crystal as a filler, enabling high-resolution static imaging and high-speed dynamic ionizing radiation imaging without afterimages.

[0008] The fourth objective of this invention is to provide an X-ray imaging inspection device equipped with the flexible thin film, which is suitable for various inspection scenarios such as industrial flaw detection, biological live organism detection, and public security inspection.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] Technical Solution 1: A method for high-throughput screening of cuprous iodide cluster scintillators using machine learning, characterized by the following steps:

[0011] 1) Dataset Construction

[0012] Data on Cu(I)-I cluster crystals were collected from a crystal structure database to construct a multi-dimensional feature dataset.

[0013] The input features of the feature dataset include organic ligand molecular structure features and descriptors;

[0014] The descriptor includes crystal geometry cell parameters and effective atomic number (Z). eff ), bandgap width (E) g Solid-state photoluminescence quantum yield (PLQY) and X-ray radioluminescence (RL) spectral characteristics;

[0015] The X-ray radiation emission spectrum light output area is used as the sole target variable for model prediction.

[0016] 2) Multi-dimensional feature extraction and preprocessing

[0017] Organic ligand molecular structure features are converted into SMILES strings, and multiple types of two-dimensional molecular fingerprints are generated in batches. Then, a Directed Message Passing Neural Network (D-MPNN) is used to perform deep learning on the molecular topology graph composed of ligand atoms and chemical bonds to extract deep embedding features of ligand atom bonding graph (i.e., features of atomic bonding and spatial connection graph within the ligand).

[0018] The correlation coefficients of all input features in the feature dataset (i.e., molecular fingerprint output, D-MPNN graph features, cell length, angle, cell volume, effective atomic number, band gap, solid-state photoluminescence quantum yield, and X-ray radiation emission spectrum features, etc.) are calculated. Redundant features with correlations higher than a set threshold are removed (i.e., Pearson correlation coefficients are calculated pairwise for all input features, and the absolute value of the correlation coefficient is set to 0.85 as the judgment threshold to remove highly linearly correlated redundant features to eliminate multicollinearity. This step only completes the removal of linearly correlated redundant features). All remaining descriptors after removing redundancy are uniformly standardized and normalized preprocessed to eliminate differences in the dimensions of different features and avoid interference from different physical dimensions on model training, resulting in a canonical descriptor feature matrix that can be directly input into the machine learning model.

[0019] 3) Feature fusion processing using Convolutional Neural Networks (CNNs)

[0020] The molecular fingerprint is sequentially passed through a one-dimensional convolutional layer and a global max pooling layer for dimensionality reduction and compression to obtain a fingerprint compression vector. The ligand atom bond graph deep embedding feature and the normalized descriptor are respectively mapped to the same feature dimension through a fully connected layer, and then vector concatenation and fusion are performed with the fingerprint compression vector to obtain a low-dimensional comprehensive feature vector with unified dimension, avoiding the dimensionality explosion and data sparsity problems caused by direct concatenation of multiple features.

[0021] 4) Stacked ensemble machine learning model training

[0022] The sample dataset composed of low-dimensional integrated feature vectors is randomly divided into a training subset and an independent test subset according to a preset ratio. On the training subset, the Stacking ensemble model is fitted and trained using a K-fold cross-validation strategy. The model hyperparameter combination is then traversed using a Bayesian hyperparameter optimization algorithm. The optimal hyperparameter combination is determined with the minimum prediction error of the test subset as the objective. Finally, the optimal Stacking ensemble prediction model for quantitative evaluation of material properties is obtained.

[0023] The Stacking ensemble model is built with a two-layer architecture, consisting of multiple heterogeneous base learners at the bottom layer and a single meta learner at the top layer. The output of the heterogeneous base learners at the bottom layer is used as the input features of the meta learners at the top layer to complete the final prediction output.

[0024] 5) Screening of candidate materials

[0025] For the Cu(I)-I cluster crystal to be developed, steps 2) and 3 are executed sequentially to complete the generation of organic ligand molecular structure feature fingerprints, graph network feature extraction, redundant feature removal, descriptor normalization, and multi-source feature vector splicing and fusion to obtain the low-dimensional comprehensive feature vector of the sample to be predicted.

[0026] The low-dimensional integrated feature vector of the sample to be predicted is input into the optimal Stacking integrated prediction model obtained in step 4). The high-performance cuprous iodide cluster scintillator candidate materials are selected by sorting the predicted X-ray radiation emission spectrum light output area values ​​from high to low.

[0027] Furthermore, in step 1), the crystal structure database is the Cambridge Crystallographic Data Centre (CCDC), which collects 61 types of Cu(I)-I cluster crystals to ensure that the dataset covers a sufficiently diverse range of copper-iodine cluster structure types, providing sufficient sample support for model training.

[0028] In step 2), the multiple types of two-dimensional molecular fingerprints include three discrete fingerprints: molecular access system (MACCS) fingerprint, Morgan fingerprint, and topological (TP) molecular fingerprint. These three types of fingerprints describe the topological structure and functional group characteristics of ligand molecules from different perspectives, and are highly complementary, which helps to improve the model's ability to recognize the diversity of ligand structures; the redundant feature that is removed is the unit cell γ angle.

[0029] Furthermore, step 3) specifically involves:

[0030] 3.1) The molecular fingerprints of different types are uniformly reshaped into a one-dimensional convolutional network input size. The one-dimensional convolutional layer Conv1D is used to extract local substructure features of the molecules. Then, the high-dimensional fingerprint vector is compressed and reduced by a global max pooling layer (GMP) to output a fingerprint compressed feature vector of fixed length.

[0031] 3.2) For the ligand atom bond graph features and the normalized descriptors, build independent fully connected layer FNNs to map the heterogeneous features with different dimensions and data types into low-dimensional one-dimensional vectors of the same length. Then, perform concatenation and fusion on the mapped vectors of equal length. Since the initial dimensions of the original features are inconsistent, they cannot be directly concatenated. Dimension alignment must be performed through a fully connected layer before vector concatenation to avoid dimension mismatch errors and feature weight imbalance.

[0032] 3.3) The fingerprint compression vector output by convolutional pooling in 3.1) is concatenated and fused with the materialized feature vector after fully connected mapping and concatenation in 3.2) to obtain a low-dimensional comprehensive feature vector with uniform dimension. This effectively avoids the defects caused by direct concatenation of multi-source original features, such as dimension explosion, sparse sample features, and overfitting of model training.

[0033] Further, step 4) specifically involves:

[0034] 4.1) Dataset partitioning: The dataset corresponding to the fused low-dimensional comprehensive feature vector is randomly split into a training subset and an independent test subset in a 9:1 ratio. A 10-fold cross-validation mode is used to reduce the randomness of data partitioning.

[0035] 4.2) Combination of Heterogeneous Base Learners: Three heterogeneous algorithms, namely XGBoost regressor, Random forest regressor (RF), and LightGBM regressor, are selected as the bottom-level base learners; the linear regression (LR) regressor is used as the upper-level meta-learner, and a two-layer stacking architecture is built. The three heterogeneous algorithms, XGBoost, Random forest, and LightGBM, are based on gradient boosting, randomized ensemble, and histogram acceleration strategies, respectively, to fit the data distribution from different perspectives. The linear regression meta-learner learns the optimal linear combination weights of the prediction results of the three base learners, effectively offsetting the inherent bias of each individual algorithm.

[0036] 4.3) Hyperparameter Grid Optimization: A grid search is performed on all hyperparameters, including base learner tree depth, learning rate, sample sampling ratio, and meta-learning regularization coefficients, to select the optimal hyperparameter combination; the optimal model's predictive coefficient of determination R for the test light collection output is determined. 2 ≥0.88 and MSE≤0.97 indicate that the model's prediction accuracy for the X-ray radiation emission spectral output area is significantly better than existing single regression models (maximum R-value). 2 =0.78).

[0037] Technical Solution 2: A cuprous iodide cluster scintillator, characterized in that: it is obtained by screening using the aforementioned high-throughput screening method of machine learning; its chemical formula is Cu6I8(1-butyl-1,4-diazabicyclo[2.2.2]octane-1-onium)2, which can be abbreviated as Cu6I8(bu-ted)2; the crystal belongs to the monoclinic crystal system, the space group is P21 / n, and the microstructure is a one-dimensional micron rod structure.

[0038] The crystal exhibits the following core optical and radiometric properties: it emits 530 nm green phosphorescence under 365 nm ultraviolet light excitation, with a solid-state photoluminescence quantum yield of 97.3%; its room-temperature luminescence lifetime is 8.4 μs, extending to 10.3 μs at a low temperature of 77 K. This microsecond-level luminescence lifetime is the physical basis for achieving high-speed dynamic X-ray imaging without afterimages. The rapid decay of luminescence after the X-ray is turned off adapts to the dynamic real-time imaging requirements of high-speed rotating samples. Its X-ray radiation luminescence intensity is 56 times that of commercial lead tungstate single crystals (PbWO4) and twice that of BGO single crystals; the lower limit of X-ray air detection is as low as 19.6 nGy. air s -1 With an effective atomic number of 46.7, it exhibits excellent X-ray blocking ability; its luminescence intensity does not decrease after 1 hour of continuous X-ray irradiation, demonstrating excellent irradiation stability; the raw materials are free of lead and thallium, making it green and environmentally friendly, and it can be synthesized in a low-temperature solution.

[0039] The method for preparing the above-mentioned cuprous iodide cluster scintillator is characterized by including the following steps:

[0040] S1. Dissolve cuprous iodide and organic ligand molecules in an alcohol solvent to prepare a mixed reaction solution;

[0041] The organic ligand molecule is any one or more of alkyl-substituted 1,4-diazabicyclo[2.2.2]octane halide, pyridine and its derivatives, imidazole-based nitrogen heterocyclic ligands, and triphenylphosphine; for example, it may be 1-butyl-1,4-diazabicyclo[2.2.2]octane iodide, 1-propyl-1,4-diazabicyclo[2.2.2]octane iodide, or pyridine;

[0042] The alcohol solvent is anhydrous ethanol, or one or a mixture of methanol, isopropanol and acetonitrile;

[0043] S2. The mixed reaction solution obtained in S1 is stirred at a constant temperature of 25-70 ℃ for 0.5-12 h. After the reaction is completed, it is allowed to stand at room temperature and the solid crude product is precipitated by natural cooling crystallization. The preferred process parameters are: reaction temperature 30-60 ℃, stirring reaction time 1-6 h.

[0044] S3. After S2 crystallization, the reaction system was centrifuged to separate the solid and liquid phases, the bottom solid precipitate was collected, washed multiple times with organic solvent, and then vacuum dried to finally obtain pure cuprous iodide cluster scintillator.

[0045] The above preparation method adopts a low-temperature wet chemical solution synthesis route, which does not require a high-temperature single crystal furnace, and the process is simple and suitable for large-scale preparation.

[0046] Furthermore, in S1, the proportions of each raw material are limited as follows:

[0047] The molar ratio of cuprous iodide to organic ligand molecules is 1:0.5-1:4, with the preferred molar ratio being: cuprous iodide: 1-butyl-1,4-diazabicyclo[2.2.2]octane iodide = 1:1.2;

[0048] Anhydrous ethanol is added according to the liquid-solid ratio. For every 1 mmol of cuprous iodide, 5-30 mL of anhydrous ethanol is added, preferably 15 mL of anhydrous ethanol for every 1 mmol of cuprous iodide; for every 1 mmol of 1-butyl-1,4-diazabicyclo[2.2.2]octane iodide, 1-5 mL of anhydrous ethanol is added, preferably 2 mL of anhydrous ethanol for every 1 mmol of 1-butyl-1,4-diazabicyclo[2.2.2]octane iodide.

[0049] In S2, the solid product was washed three times with anhydrous ethanol, and after washing, it was vacuum dried at 60 °C to constant weight.

[0050] The present invention also provides the application of the above-mentioned cuprous iodide cluster scintillator in the preparation of X-ray scintillator thin films and X-ray imaging.

[0051] Technical Solution 3: A flexible X-ray composite scintillation imaging film, characterized in that it comprises a polydimethylsiloxane flexible matrix and the aforementioned cuprous iodide cluster scintillator filler dispersed in the flexible matrix; the mass fraction of the cuprous iodide cluster scintillator filler is 4-6% (preferably 5%, at which filler content achieves optimal balance between film transmittance and luminescence, avoiding the contradiction between high filler agglomeration and low filler insufficient brightness); the thickness of the film is 0.2-0.4 mm (preferably 0.3 mm, which effectively controls scattering while ensuring X-ray absorption efficiency); the film can be bent arbitrarily, and its imaging performance shows no degradation after 2000 bending fatigue cycles; the static X-ray imaging spatial resolution reaches 20 lp mm. -1 The light emission decays at the microsecond level, 20° s -1High-speed rotating imaging eliminates image ghosting and trailing, making it suitable for dynamic real-time X-ray detection. The polydimethylsiloxane flexible matrix has high light transmittance and can be bent arbitrarily. The microrods are uniformly dispersed in the matrix, resulting in weak scattering effects and balancing high light transmittance with high luminous brightness.

[0052] The method for preparing the above-mentioned flexible X-ray composite scintillation imaging film is characterized by including the following steps:

[0053] (1) Mix the polydimethylsiloxane prepolymer and curing agent at a mass ratio of 10:1, add cuprous iodide cluster scintillator, and ultrasonically disperse evenly;

[0054] (2) After the dispersion obtained in step (1) is allowed to stand under vacuum to remove bubbles, it is poured into a flat mold for film laying;

[0055] (3) The film obtained in step (2) is left to stand at room temperature to solidify and form, and then demolded to obtain the flexible X-ray composite scintillation imaging film.

[0056] Ultrasonic dispersion can effectively eliminate the agglomeration of micron rod particles, while vacuum static degassing can avoid internal bubble defects in the film. Together, they ensure the uniformity of film imaging.

[0057] Technical Solution 4: An X-ray imaging detection system, comprising an X-ray radiation source, a reflector, an image acquisition camera, and a data processing terminal, characterized in that it also includes the aforementioned flexible X-ray composite scintillation imaging film. The X-ray radiation source emits ionizing radiation that penetrates the sample under test and irradiates the flexible X-ray composite scintillation imaging film; the flexible X-ray composite scintillation imaging film converts high-energy rays into 530 nm visible light; the reflector reflects the visible light to the image acquisition camera; the image acquisition camera acquires image signals and transmits them to the data processing terminal; the data processing terminal processes the received image signals and outputs a static imaging image or a dynamic real-time video image.

[0058] The concept and principle of this invention:

[0059] This invention first constructs a multi-dimensional copper-iodine cluster dataset containing molecular fingerprints, graph network features, and descriptors. It then uses one-dimensional CNN convolution to compress and fuse high-dimensional molecular features, solving the dimensionality explosion and overfitting problems caused by feature splicing. Next, it employs XGB, RF, and LGBM as base learners and linear regression as a meta-learner to construct a stacking ensemble model, offsetting the fitting bias of a single algorithm, accurately predicting the X-ray light output performance of materials, and selecting Cu6I8(bu-ted)2 micrometer rod scintillation crystals. These crystals rely on the high atomic number Cu / I to efficiently absorb X-rays, with excitons moving towards the cluster center in a triplet state (…). 3The migration of CC generates microsecond-level green phosphorescence, and the coordination cross-linked lattice has excellent water resistance and radiation resistance. The microrod filler is combined with PDMS to make a flexible film that can uniformly convert X-rays into visible light, and complete static, high-speed, and afterimage-free dynamic ionizing radiation imaging with imaging devices.

[0060] Advantages of this invention:

[0061] The CNN feature fusion and stacking ensemble screening method of this invention has multiple advantages over traditional single machine learning models:

[0062] ① High-dimensional molecular fingerprints are compressed using one-dimensional convolution and global max pooling, and then fused with ligand atom bonding graph features extracted by a directed message passing graph neural network and crystal numerical features to form a unified low-dimensional comprehensive feature vector. After multi-source feature fusion, the model R 2 The standard deviation decreased from 0.12 to 0.04, significantly reducing the risk of overfitting on small sample datasets and significantly improving stability;

[0063] ② The combination of XGB, RF, and LGBM heterogeneous base learners with LR meta-learners can offset the inherent prediction bias of a single algorithm, resulting in the optimal model R. 2 The accuracy reached 0.88, far exceeding the highest prediction accuracy of 0.78 of the traditional single model, achieving high-precision and high-throughput prediction and screening of X-ray light output of cuprous iodide cluster scintillators;

[0064] ③ No need for a large number of pre-synthesized candidate crystals; the scintillation performance of materials can be predicted solely through databases and ligand theoretical parameters. The development cycle of new materials is shortened by more than 70%, significantly saving reagent, manpower, and time costs, and enabling high-throughput targeted discovery of cuprous iodide scintillators.

[0065] The Cu6I8(bu-ted)2 scintillation crystal of this invention offers a comprehensive performance improvement over existing commercially available and reported copper-iodine cluster materials:

[0066] ① The radiation intensity is 56 times that of PbWO4, and the X-ray dose used is only 1 / 250 of that of traditional equipment under the same imaging clarity, which greatly reduces the ionizing radiation damage to the human body in medical and security inspection scenarios.

[0067] ② The detection limit is as low as 19.6 nGy air s -1 It can identify ultra-weak X-ray signals and is suitable for low-dose biological in vivo imaging;

[0068] ③ It possesses excellent water stability and long-term X-ray irradiation stability, is lead-free and non-toxic, and avoids the environmental and aging defects of perovskite and lead-based single crystals.

[0069] ④ Low-temperature wet chemical solution synthesis eliminates the need for a high-temperature single crystal furnace above 1000 ℃, resulting in a simple preparation process and low mass production cost.

[0070] This invention, a flexible composite scintillation film based on PDMS, addresses a key shortcoming of existing flexible imaging screens:

[0071] ① The micron rod filler is uniformly dispersed in the PDMS matrix, resulting in weak film scattering effect and a static resolution of up to 20 lpmm. -1 It can clearly distinguish fine structures such as micron-level circuits on precision PCBs and the delicate skeletons of fish.

[0072] ② The luminescence lifetime is on the order of microseconds. The luminescence decays rapidly after the radiation is turned off. High-speed dynamic imaging has no afterimages or ghosting. It is suitable for online non-destructive testing and dynamic interventional X-ray imaging in production lines.

[0073] ③ The film can be prepared in large areas (7 cm width), can be bent, and its performance does not decrease after thousands of bends. It is suitable for the inspection of irregularly shaped workpieces and portable security inspection equipment. A filler content of 5 wt% can balance light transmittance and luminous brightness, and there is no contradiction between high filler agglomeration and low filler insufficient brightness.

[0074] The X-ray imaging detection device of this invention integrates the above-mentioned high-performance flexible film, making the whole device lightweight, adaptable to curved surfaces, with clear static imaging details and no motion blur in dynamic imaging. It can cover ionizing radiation imaging scenarios such as medical low-dose radiation, industrial PCB / casting flaw detection, small animal live imaging, and public security inspection, significantly expanding the applicability of the device. Attached Figure Description

[0075] Figure 1 This is a flowchart illustrating the overall process of the machine learning-based high-throughput screening method for cuprous iodide cluster scintillators of this invention. It is divided into three main stages: dataset construction, model training, and prediction validation. First, 61 cuprous iodide cluster crystal structures are screened and organized from the Cambridge Crystal Database (CCDC), and ligand configurations and effective atomic numbers (Z) are collected. eff Crystal geometry parameters, band gap (E) g A database was established using characteristic parameters such as photoluminescence quantum yield (PLQY) and luminescence intensity, with the light output set as the model's prediction target value. Subsequent data cleaning and feature extraction were performed. The dataset was randomly divided into training and independent test subsets at a 9:1 ratio, and 10-fold cross-validation was conducted. The coefficient of determination (R²) was used to determine the model's performance. 2 The performance of various models was evaluated using two metrics: mean squared error (MSE). The stacked ensemble model with XGBoost, Random Forest (RF), and LightGBM as the basic learners and Linear Regression (LR) as the fusion learner was determined to be the best performing. Finally, the prediction results of the optimized model for novel cuprous iodide cluster materials were verified through synthesis experiments.

[0076] Figure 2 This is the model training process for the present invention;

[0077] a is a schematic diagram of various data cleaning and feature extraction methods;

[0078] b is a schematic diagram of a feature fusion architecture based on a convolutional neural network (CNN);

[0079] c represents the coefficient of determination (R²) for traditional machine learning models such as XGB, RF, SVR, ABR, GBR, LR, KNR, LGBM, NNR, RR, CBR, and ETR. 2 )contrast;

[0080] d represents a performance comparison of models built using four combinations of heterogeneous base learners: RR, CBR, and SVR; LR, NNR, and KNR; ABR, GBR, and ETR; and XGB, RF, and LGBM.

[0081] Figure 3 The following are a series of characterization spectra of the Cu6I8(bu-ted) 2-micron scintillation crystal structure, photoluminescence, X-ray response, detection limit, and irradiation stability of the present invention.

[0082] a is a schematic diagram of the organic-inorganic hybrid material Cu6I8(bu-ted)2 micrometer rod crystal structure;

[0083] b is the two-dimensional excitation-photoluminescence spectrum of Cu6I8(bu-ted)2 microrods, and the inset is a photograph of the microrods emitting light under 365 nm ultraviolet light irradiation.

[0084] c represents the temperature-dependent fluorescence decay curve of the green luminescence of Cu6I8(bu-ted)2;

[0085] d is in the range of 0.688~278 μGy air s -1 The radiative emission (RL) spectra of Cu6I8(bu-ted)2 micrometer rods under different excitation dose rates within the dose range;

[0086] The radiative emission intensity of a Cu6I8(bu-ted)2 μm rod varies with excitation dose; the X-ray detection limit (DL) of the material was calculated to be 19.6 nGy using the 3σ / slope method. air s -1 ;

[0087] f represents the photocurrent curve of a Cu6I8(bu-ted)2 μm rod film with and without X-ray irradiation; the tests were performed under multiple applied voltages, with an irradiation dose rate of 14.4 mGy. air s -1;

[0088] X-ray tolerance test of Cu6I8(bu-ted)2 micrometer rods: The above figure shows the continuous X-ray irradiation test (dose rate 4.01 mGy). air s -1 The figure below shows the intermittent cyclic X-ray irradiation test (equal dose rate, 20s interval between irradiations); the radiative emission intensity at 530 nm was monitored throughout the experiment to characterize the irradiation stability of the microrod material.

[0089] Figure 4 The present invention includes physical objects and images of a 5 wt% Cu6I8(bu-ted)2 doped PDMS flexible thin film, a self-made X-ray imaging test platform, standard line-pair card resolution imaging, PCB static imaging, fish skeleton bioimaging, and capsule dynamic afterimage-free imaging.

[0090] a is a real photo of a 5 wt% Cu6I8(bu-ted) 2-micron rod-doped PDMS flexible thin film under natural light;

[0091] b is a photograph of the film emitting light under 365 nm ultraviolet light excitation;

[0092] c is a schematic diagram of a self-made X-ray imaging test platform. The system uses a digital camera to capture images on a flexible shimmering screen.

[0093] d is the image from the standard X-ray resolution test chart, with the test line pair ranging from 11 to 20 lp mm. -1 ;

[0094] e is an X-ray image of the timing circuit board;

[0095] f shows the bright-field image of the fish sample and its corresponding X-ray image;

[0096] g and h are optical images and dynamic X-ray images of the internal metal spring capsule; the capsule has a 20 o s -1 When rotating at a certain speed, dynamic X-ray imaging exhibits no afterimage or trailing phenomenon. Detailed Implementation

[0097] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Unless otherwise specified, all raw materials used in the following embodiments are commercially available products.

[0098] Example 1: Cu6I8(bu-ted)2 material screening based on stacked ensemble machine learning

[0099] This embodiment describes a high-throughput screening method for cuprous iodide cluster scintillators based on CNN multi-source feature fusion and Stacking ensemble machine learning. Figure 1 and Figure 2 The process is shown below.

[0100] The specific filtering method is as follows:

[0101] Step 1: Dataset Construction

[0102] Sixty-one Cu(I)-I cluster crystals with complete crystal structure information were selected from the Cambridge Crystal Database to construct a dedicated sample dataset. The dataset input features included organic ligand molecular structure features and descriptors. Specifically, the descriptors included crystal geometric cell parameters, effective atomic number, band gap, solid-state photoluminescence quantum yield, and X-ray emission spectrum characteristics. The X-ray emission spectrum light output area was used as the sole prediction target variable for the model.

[0103] Step 2: Multi-dimensional feature extraction and preprocessing

[0104] All organic ligand molecules were converted into SMILES strings, and three types of discrete fingerprints were generated in batches: Molecular Access System Fingerprints (MACCS), Morgan fingerprints, and Topological Molecular Fingerprints (TP). Directed Message Passing Graph Neural Network (D-MPNN) was used to perform deep learning on the molecular topology graph composed of ligand atoms and chemical bonds, and deep embedding features of ligand atom bonding graph (i.e., the bonding and spatial connection graph features of atoms inside the ligand) were extracted.

[0105] Pearson correlation coefficients were calculated pairwise for all input features in the dataset (i.e., molecular fingerprint output, D-MPNN graph features, cell length, angle, cell volume, effective atomic number, band gap, solid-state photoluminescence quantum yield, and X-ray radioluminescence spectral features, etc.). The results showed a high correlation between cell angles and other cell parameters, and these were discarded as redundant features. All retained descriptors underwent standardized and normalized preprocessing to eliminate interference from different physical dimensions on model training, resulting in a canonical descriptor feature matrix that can be directly input into a machine learning model.

[0106] Step 3: Convolutional Neural Network Feature Fusion Processing

[0107] The three types of molecular fingerprints are uniformly reshaped into the standard input size of a one-dimensional convolutional network and fed into a one-dimensional convolutional layer (Conv1D) to extract local substructure features of molecules. Then, the high-dimensional fingerprint vector is compressed and reduced in dimensionality by a global max pooling layer (GMP) to obtain a fingerprint compressed vector of fixed length.

[0108] The molecular graph features extracted by D-MPNN and the normalized descriptors are dimensionally mapped through independent fully connected layers (FNNs). After converting the two types of heterogeneous features into one-dimensional vectors of the same length, vector concatenation and fusion are performed to obtain the aligned materialized feature vectors.

[0109] The aforementioned fingerprint compression vector is then fused with the materialized feature vector mapped and concatenated through a fully connected layer to obtain a low-dimensional comprehensive feature vector with uniform dimensions. This hierarchical and step-by-step fusion strategy fundamentally avoids the defects caused by direct concatenation of multi-source original features, such as dimensionality explosion, sparse sample features, and model overfitting. The standard deviation of the model's prediction determination coefficient is optimized and reduced from 0.12 to 0.04.

[0110] Step 4: Stacked Ensemble Machine Learning Model Training

[0111] The sample dataset, composed of the fused low-dimensional integrated feature vectors, is randomly divided into a training subset and an independent test subset at a 9:1 ratio. A 10-fold cross-validation mode is used to reduce the randomness of data partitioning. Three heterogeneous algorithms, namely XGBoost regressor, random forest regressor (RF), and lightGBM regressor, are selected as the bottom-level base learners, and linear regression (LR) is used as the upper-level meta-learner to build a two-layer stacking ensemble architecture.

[0112] A grid search hyperparameter traversal was performed to optimize the decision tree depth, learning rate, sample row sampling ratio, feature column sampling ratio of the base learner, and the regularization coefficient of the meta-learner. The optimal stacked ensemble model after parameter optimization achieved a coefficient of determination R0 on the test set. 2 =0.88, Mean Square Error (MSE)=0.97; combined with the appendix Figure 2 As shown in the performance comparison bar chart in c, this result significantly outperforms the best performance of existing techniques using any single regression model of XGBoost, RF, or LightGBM (the highest performance of a single model is only R). 2 =0.78), fully demonstrating the improved prediction accuracy of the stacked fusion architecture in this application.

[0113] Step 5: Candidate Material Screening

[0114] Five groups of novel copper-iodine cluster candidate materials containing bu-ted ligands were collected, and the SMILES strings, crystal structure parameters, and optical performance characterization parameters of each group of materials were extracted. Feature engineering was completed in strict accordance with the feature extraction and multi-source feature fusion process in steps 2 and 3 above. The final fused features were input into the optimal Stacking ensemble prediction model that has been trained and optimized for hyperparameters for high-throughput prediction.

[0115] The model output results show that the predicted X-ray fluorescence output intensity of Cu6I8(bu-ted)2 ranks the highest among the five candidate structures, so this compound is selected as the target scintillator material for priority experimental synthesis.

[0116] Example 2: Preparation of Cu6I8(bu-ted)2 micron rod scintillation crystal

[0117] This embodiment describes the preparation and performance characterization of Cu6I8(bu-ted)2 micron rod scintillation crystals, corresponding to... Figure 3 The characterization results are shown.

[0118] Preparation of raw materials: Weigh cuprous iodide (CuI) and 1-butyl-1,4-diazabicyclo[2.2.2]octane iodide (bu-ted iodide) at a molar ratio of 1:1.2; mix 1 mmol of cuprous iodide with 15 mL of anhydrous ethanol and 1 mmol of cuprous iodide with 2 mL of anhydrous ethanol. Stir magnetically at room temperature for 30 min until the solid is completely dissolved. After mixing, prepare a homogeneous and transparent precursor reaction solution.

[0119] Low-temperature solvothermal reaction: The above reaction solution was placed on a constant temperature magnetic stirrer and stirred at 40 °C for 4 h to allow cuprous iodide and bu-ted iodide to react fully in the ethanol system. After the reaction was completed, the reaction solution was cooled to room temperature and allowed to stand at room temperature to crystallize naturally to precipitate solid products.

[0120] Purification and drying: After the solid product is fully separated, the solid is collected by centrifugation, washed three times with ethanol to remove residual ligands, and dried under vacuum at 60°C to obtain white one-dimensional micron rod crystals;

[0121] Performance verification: combined with appendix Figure 3 The characterization results of each sub-graph are used to verify the Cu6I8(bu-ted)2 crystal: such as Figure 3 As shown in figure a, the structure and XRD results confirm that it is a monoclinic P21 / n pure phase target crystal; as Figure 3 As shown in b, Cu6I8(bu-ted)2 emits 530 nm green light under 365 nm ultraviolet excitation; as Figure 3 As shown in d, the X-ray emission peak of Cu6I8(bu-ted)2 is also located at 530 nm, indicating that Cu6I8(bu-ted)2 crystal has the same luminescence center under ultraviolet light and X-ray excitation. Figure 3 As shown in Figure e, the dose-linear response determination coefficient of Cu6I8(bu-ted)2 reaches 0.99, and its detection limit is calculated to be as low as 19.6 nGyair s based on the 3σ / slope method. -1 ;at the same time, Figure 3The f-photocurrent test proves that Cu6I8(bu-ted)2 can achieve X-ray photoelectric conversion; such as Figure 3 As shown in g, continuous long-term and cyclic X-ray irradiation tests both indicate that the material's luminescence does not significantly decrease, demonstrating excellent irradiation stability.

[0122] Example 3: Preparation of 5 wt% Cu6I8(bu-ted)2-doped PDMS flexible imaging thin film

[0123] This embodiment describes the preparation and imaging performance testing of a 5 wt% Cu6I8(bu-ted)2-doped polydimethylsiloxane flexible imaging film. Figure 4 The results are shown.

[0124] Dispersion preparation: PDMS prepolymer and curing agent are mixed at a mass ratio of 10:1, and 5 wt% of dried Cu6I8(bu-ted)2 micron rods are added. The mixture is ultrasonically dispersed for 30 min to eliminate particle agglomeration and obtain a uniform dispersion.

[0125] Degassing and film laying: Transfer the above dispersion to a vacuum drying oven and let it stand under vacuum for 1 hour to fully remove air bubbles from the mixture and avoid air bubble defects inside the film. After degassing, pour the dispersion evenly into a flat mold for film laying;

[0126] Room temperature curing: The mold after film deposition is left to stand at room temperature to fully cure and solidify, and then demolded to obtain a transparent flexible scintillation film with a thickness of 0.3 mm. The film is translucent under natural light and emits bright green fluorescence under 365 nm ultraviolet light excitation.

[0127] Imaging test: based on Figure 4 The X-ray imaging system shown in figure c was tested using... Figure 4 Calibration with the d-standard line can achieve a static imaging resolution of up to 20 lp mm for flexible thin films. -1 ; Figure 4 The d-axis can clearly distinguish micron-level circuitry on PCB boards. Figure 4 The image fully reveals the fine skeletal structure of fish, and has excellent static imaging detail resolution.

[0128] by Figure 4 The g-banded metal spring capsule is used for dynamic imaging of rotating samples. Figure 4 The time-series images at different times show that, due to the material's microsecond-level luminescence lifetime, the luminescence decays rapidly after the ray is turned off. o s -1 High-speed rotating dynamic imaging eliminates afterglow and ghosting phenomena.

[0129] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A method for high-throughput screening of cuprous iodide cluster scintillators using machine learning, characterized in that, Includes the following steps: 1) Dataset Construction Data on Cu(I)-I cluster crystals were collected from a crystal structure database to construct a multi-dimensional feature dataset. The input features of the feature dataset include organic ligand molecular structure features and descriptors; The descriptor includes crystal geometry cell parameters, effective atomic number, band gap width, solid-state photoluminescence quantum yield, and X-ray radiation emission spectral characteristics; 2) Multi-dimensional feature extraction and preprocessing The molecular structural features of organic ligands are converted into SMILES strings, and multiple types of two-dimensional molecular fingerprints are generated in batches. Then, a directional message passing graph neural network is used to extract deep embedding features of ligand atom bonding graphs. The correlation coefficients of all input features in the feature dataset are calculated, and redundant features with correlations higher than a set threshold are removed. Then, all remaining descriptors after removing redundancies are uniformly standardized and normalized to eliminate dimensional differences and obtain a standardized descriptor feature matrix. 3) Convolutional Neural Network Feature Fusion Processing The molecular fingerprint is sequentially passed through a one-dimensional convolutional layer and a global max pooling layer for dimensionality reduction and compression to obtain a fingerprint compression vector. The ligand atom bond graph deep embedding feature and the normalized descriptor are respectively mapped to the same feature dimension through a fully connected layer, and then vector concatenation and fusion are performed with the fingerprint compression vector to obtain a low-dimensional comprehensive feature vector with unified dimension. 4) Stacked ensemble machine learning model training The sample dataset composed of low-dimensional integrated feature vectors is randomly divided into a training subset and an independent test subset according to a preset ratio. On the training subset, the Stacking ensemble model is fitted and trained using a K-fold cross-validation strategy. The model hyperparameter combination is then traversed using a Bayesian hyperparameter optimization algorithm. The optimal hyperparameter combination is determined with the minimum prediction error of the test subset as the objective, and finally the optimal Stacking ensemble prediction model is obtained. The Stacking ensemble model is built with a two-layer architecture, consisting of multiple heterogeneous base learners at the bottom layer and a single meta learner at the top layer connected in series. The output of the heterogeneous base learners at the bottom layer is used as the input feature of the meta learner at the top layer. 5) Screening of candidate materials Steps 2) and 3) are performed sequentially to obtain the low-dimensional comprehensive feature vector of the sample to be predicted for the Cu(I)-I cluster crystal to be developed; The low-dimensional integrated feature vector of the sample to be predicted is input into the optimal Stacking integrated prediction model obtained in step 4). The high-performance cuprous iodide cluster scintillator candidate materials are selected by sorting the predicted X-ray radiation emission spectral area values ​​from high to low according to the model output.

2. The method for high-throughput screening of cuprous iodide cluster scintillators using machine learning according to claim 1, characterized in that: In step 1), the crystal structure database is the Cambridge Crystal Database, which collects 61 types of Cu(I)-I cluster crystals; In step 2), the various types of two-dimensional molecular fingerprints include molecular access system fingerprints, Morgan fingerprints, and topological molecular fingerprints; the redundant features that are removed are the unit cell γ angles.

3. The method for high-throughput screening of cuprous iodide cluster scintillators using machine learning according to claim 1 or 2, characterized in that: In step 4), the plurality of heterogeneous base learners include an XGBoost regressor, a random forest regressor, and a LightGBM regressor; the meta-learner is a linear regressor. The dataset is divided into a training subset and an independent test subset in a 9:1 ratio; The cross-validation method used is 10-fold cross-validation. The Bayesian hyperparameter optimization algorithm uses a grid search method. The coefficient of determination R of the optimal prediction model 2 ≥0.

88.

4. A cuprous iodide cluster scintillator, characterized in that: It was obtained by screening using any one of the methods described in claims 1-3; its chemical formula is Cu6I8(bu-ted)2, and its crystal is monoclinic. The space group has a morphology of one-dimensional microrods.

5. The method for preparing the cuprous iodide cluster scintillator according to claim 4, characterized in that, Includes the following steps: S1. Dissolve cuprous iodide and organic ligand molecules in an alcohol solvent to prepare a mixed reaction solution; The organic ligand molecule is any one or more of the following: alkyl-substituted 1,4-diazabicyclo[2.2.2]octane halide, pyridine and its derivatives, imidazole-based nitrogen heterocyclic ligands, and triphenylphosphine. The alcohol solvent is anhydrous ethanol, or one or a mixture of methanol, isopropanol and acetonitrile; S2. The mixed reaction solution obtained in S1 is stirred at a constant temperature of 25-70 ℃ for 0.5-12 h. After the reaction is completed, it is allowed to stand at room temperature and the solid crude product is precipitated by natural cooling crystallization. S3. After crystallization of S2, the reaction system was centrifuged to separate the solid and liquid phases. The bottom solid precipitate was collected, washed multiple times with organic solvent, and then dried under vacuum to finally obtain pure cuprous iodide cluster scintillator.

6. The preparation method according to claim 5, characterized in that: In S1, the proportions of each raw material are limited as follows: The molar ratio of cuprous iodide to organic ligand molecules is 1:0.5-1:4 in terms of molar amount. Anhydrous ethanol was added according to the liquid-solid ratio, with 5-30 mL of anhydrous ethanol added for every 1 mmol of cuprous iodide; and 1-5 mL of anhydrous ethanol added for every 1 mmol of 1-butyl-1,4-diazabicyclo[2.2.2]octane iodide. In S2, the solid product was washed three times with anhydrous ethanol, and after washing, it was vacuum dried at 60 °C to constant weight.

7. The application of the cuprous iodide cluster scintillator of claim 4 in the preparation of X-ray scintillator films and X-ray imaging.

8. A flexible X-ray composite scintillation imaging thin film, characterized in that: The film comprises a polydimethylsiloxane flexible matrix and the cuprous iodide cluster scintillator filler of claim 4 dispersed in the flexible matrix; the mass fraction of the cuprous iodide cluster scintillator filler is 4-6%; and the thickness of the film is 0.2-0.4 mm.

9. The method for preparing a flexible X-ray composite scintillation imaging thin film according to claim 8, characterized in that, Includes the following steps: (1) Mix the polydimethylsiloxane prepolymer with the curing agent, add the cuprous iodide cluster scintillator, and ultrasonically disperse it evenly; (2) After the dispersion obtained in step (1) is allowed to stand under vacuum to remove bubbles, it is poured into a flat mold to lay a film; (3) The film obtained in step (2) is left to stand at room temperature to solidify and form, and then demolded to obtain the flexible X-ray composite scintillation imaging film.

10. An X-ray imaging detection system, comprising an X-ray radiation source, a reflector, an image acquisition camera, and a data processing terminal, characterized in that: It also includes the flexible X-ray composite scintillation imaging film as described in claim 8.

Images