A multi-modal AI4S basic scientific research method based on small sample generalization learning

Abstract

The application discloses a kind of multi-modal AI4S basic scientific research methods based on small sample generalization learning, comprising: collecting the multi-source multi-modal data under basic scientific research scene and pre-processing;Extracting heterogeneous multi-modal features, introducing cross-modal attention alignment mechanism to map it to the same high-dimensional representation space, to obtain unified multi-modal fusion feature set;Small sample generalization basic model based on meta-learning framework is constructed, scientific research field knowledge graph priori knowledge is injected into feature reasoning layer, and the model is iteratively trained based on metric learning strategy using fusion feature set;The new multi-modal scientific research data to be studied is input into the trained small sample generalization basic model, and the analysis result is output after scientific research task generalization reasoning;The data after artificial checking using analysis result dynamically updates model parameters;The present application realizes the small sample generalization reasoning and model dynamic iterative optimization of multi-modal scientific research data by the meta-learning strategy of cross-modal feature alignment fusion and priori knowledge injection.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-driven scientific research technology, and more specifically to a multimodal AI4S basic scientific research method based on few-sample generalization learning. Background Technology

[0002] Currently, AI-driven scientific research (AI4S) has become the core development paradigm of basic scientific research, gradually shifting from traditional experimental trial-and-error and theoretical derivation to AI-assisted scientific discovery. Faced with the massive amounts of heterogeneous multimodal data generated during scientific research (such as atlas images, paper texts, and physicochemical values), the introduction of deep learning and large-scale model technology to replace inefficient manual analysis has become a mainstream trend. However, basic scientific research generally suffers from long experimental cycles, difficulty in obtaining samples, reliance on expert annotation, and a limited number of reusable samples of the same type. This makes traditional supervised learning methods that rely on large amounts of data and conventional multimodal fusion methods difficult to adapt to the current application situation of small samples, high specialization, and multimodal heterogeneity. Existing small-sample learning methods are mostly limited to general image and text classification and have not been customized or designed for multimodal joint generalization in AI4S scenarios. The closest existing technology is the multimodal AI4S analysis method based on large models, which mainly extracts image and text features through pre-trained large models and performs simple splicing and fusion, relying on massive amounts of publicly labeled data for fine-tuning to achieve conventional classification and retrieval. However, this approach has significant shortcomings: First, it heavily relies on large-scale manually labeled datasets, making it unsuitable for the scarce sample scenarios in basic scientific research; second, it only performs simple feature splicing on multimodal data, lacking cross-modal correlation alignment for heterogeneous characteristics, resulting in low information fusion accuracy; third, the general-purpose large model lacks the injection of prior scientific knowledge, making it prone to overfitting under small sample conditions, and exhibiting weak generalization and cross-task transfer capabilities; fourth, it has not constructed a small-sample generalization mechanism adapted to scientific research logic, and can only handle routine standardized tasks, failing to meet the modeling needs of cutting-edge emerging disciplines; and fifth, the model process is divorced from actual scientific research operations, failing to achieve a closed-loop process from multimodal data input and feature mining to scientific law derivation.

[0003] Therefore, there is an urgent need to build a new AI4S technology framework that integrates prior knowledge in the field, possesses accurate cross-modal correlation alignment and powerful small sample generalization capabilities, and closely aligns with the logic of scientific research, in order to break through the existing data and model bottlenecks and truly realize an intelligent closed loop from multimodal data-driven to scientific law discovery. Summary of the Invention

[0004] The purpose of this invention is to provide a multimodal AI4S basic scientific research method based on few-shot generalization learning. Through a meta-learning strategy of cross-modal feature alignment and fusion and prior knowledge injection, it realizes few-shot generalization inference and dynamic iterative optimization of the model for multimodal scientific research data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A multimodal AI4S basic scientific research method based on few-shot generalization learning includes the following steps: S1. Collect multi-source multimodal data in basic scientific research scenarios, perform preprocessing, and obtain preprocessed multi-source multimodal data; S2. Feature extraction is performed on the preprocessed multi-source multimodal data to obtain heterogeneous multimodal features; a cross-modal attention alignment mechanism is introduced to map the heterogeneous multimodal features to the same high-dimensional representation space to obtain a unified multimodal fusion feature set; S3. Construct a few-shot generalization basic model based on a meta-learning framework, and inject prior knowledge from the knowledge graph of basic scientific research fields into the few-shot generalization basic model. Based on the metric learning strategy, the unified multimodal fusion feature set is used to iteratively train the small-sample generalization base model with injected prior knowledge. S4. Input the multimodal scientific research data to be studied into the trained small sample generalization basic model, perform scientific research task generalization inference, and output a multimodal AI4S basic scientific research analysis report. S5. Using the aforementioned multimodal AI4S basic scientific research analysis report, dynamically update the parameters of the small-sample generalization basic model to achieve iterative optimization of the small-sample generalization basic model.

[0006] Furthermore, in S1, the multi-source multimodal data includes: microstructure images, experimental detection maps, scientific research paper texts, physicochemical experimental values, and industry standard texts; The preprocessing process includes: noise reduction, cropping, and normalization of the microstructure images and experimental detection maps; Perform word segmentation, entity extraction of professional terms, and stop word removal on scientific research paper texts and industry standard texts; Standardize and remove outliers from physicochemical experimental values.

[0007] Furthermore, in S2, the process of generating the unified multimodal fusion feature set is as follows: Shallow feature extraction is performed on the preprocessed multi-source multimodal data; A cross-modal attention alignment mechanism is introduced to establish a correlation mapping between image visual features, text semantic features, and numerical statistical features, and a unified multimodal fusion feature is constructed to obtain a unified multimodal fusion feature set.

[0008] Furthermore, the functional expression for the shallow feature extraction is: ; in, For shallow image features, These are shallow text features. These are shallow numerical features. For image encoders, For text encoders, It is a numerical encoder. For preprocessed image data, For preprocessed text data, This is the preprocessed numerical data.

[0009] Furthermore, the functional expression of the cross-modal attention alignment mechanism is: ; in, , For any two different modal features, d For feature dimension, This is the cross-modal attention weight matrix.

[0010] Furthermore, the functional expression for constructing unified multimodal fusion features is: ; in, For globally unified representation features, This is the cross-modal attention weight matrix between images and text. For shallow image features, This is the cross-modal attention weight matrix between images and numerical values. These are shallow numerical features. This is the cross-modal attention weight matrix between text and numerical values. These are shallow text features.

[0011] Furthermore, in step S3, the process of injecting prior knowledge from the knowledge graph of basic scientific research into the small-sample generalization basic model specifically involves: Construct a knowledge graph in the field of basic scientific research, and take the relationship between scientific theorems, experimental laws and professional terms as prior knowledge; By incorporating domain prior knowledge into the feature inference layer of the few-shot generalization basic model through the knowledge embedding layer, the feature learning direction of the few-shot generalization basic model is constrained, so that the few-shot generalization basic model conforms to the professional logic of basic scientific research.

[0012] Furthermore, in S3, the process of iteratively training the few-sample generalization base model with injected prior knowledge specifically involves: A small sample generalization basic model is built based on the meta-learning framework, and a metric learning strategy is introduced to calculate the similarity distribution between features of different scientific research samples and construct a loss function for small sample classification and regression tasks. Iterative training using labeled samples optimizes the feature extraction and metric parameters of the small-sample generalization base model.

[0013] Furthermore, in S4, the process of generalizing reasoning for the scientific research task includes: feature matching, pattern fitting, structure prediction, and literature knowledge association mining; The multimodal AI4S basic scientific research analysis report includes: scientific research data classification results, experimental trend predictions, and conclusions derived from potential scientific laws.

[0014] Furthermore, the implementation process of iterative optimization of the small-sample generalization base model in S5 is as follows: The results manually verified by researchers are used as incremental small sample data and fed back into the training set of the small sample generalization base model. The parameters of the small-sample generalization basic model are dynamically updated by incremental small-sample fine-tuning, continuously improving the generalization ability of the small-sample generalization basic model in interdisciplinary and new-direction basic scientific research tasks.

[0015] According to specific embodiments provided by the present invention, the present invention has the following technical effects compared to the prior art: This invention maps multi-source heterogeneous features to a unified high-dimensional representation space through a cross-modal attention alignment mechanism, breaking down information silos between different modalities and achieving deep collaboration and complementary fusion of multimodal features, significantly improving the comprehensiveness and accuracy of feature representation. It injects domain-based prior knowledge into the feature inference layer, endowing the model with inherent logical constraints from the scientific domain, overcoming the potential for non-physical compliance inferences in traditional black-box models, and greatly enhancing the interpretability and scientific rationality of model inference. Relying on meta-learning and metric learning strategies, the model possesses the ability to quickly extract common patterns and measure feature differences from a very small number of samples, greatly reducing dependence on massive amounts of labeled data and achieving efficient and accurate generalized inference for novel scientific research tasks. Combined with manual verification feedback for dynamic updating of model parameters, an adaptive closed-loop iterative mechanism of "inference-verification-evolution" is constructed, enabling the model to continuously absorb new knowledge along with the scientific exploration process, ensuring the model's performance in long-term scientific applications. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] The following description, in conjunction with the accompanying drawings, further illustrates the multimodal AI4S basic scientific research method based on few-sample generalization learning according to the present invention. Figure 1This is a schematic diagram of the overall process of the multimodal AI4S basic scientific research method based on few-sample generalization learning in this invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0019] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0020] like Figure 1 As shown, this invention provides a multimodal AI4S basic scientific research method based on few-shot generalization learning, including the following steps: S1. Collect multi-source multimodal data in basic scientific research scenarios, perform preprocessing, and obtain preprocessed multi-source multimodal data; In S1, the multi-source, multi-modal data includes: microstructure images, experimental detection maps, scientific research paper texts, physicochemical experimental values, and industry standard texts; The preprocessing process includes: noise reduction, cropping, and normalization of the microstructure images and experimental detection maps; Perform word segmentation, entity extraction of professional terms, and stop word removal on scientific research paper texts and industry standard texts; Standardize and remove outliers from physicochemical experimental values.

[0021] This embodiment specifically involves standardizing and cleaning multi-source heterogeneous scientific research data to unify its format, eliminating data noise, outliers, and invalid information interference, thus providing a high-quality standardized dataset for subsequent modeling. The specific modality-specific refined preprocessing process is as follows: (1) Multimodal data acquisition: Four types of core data in basic scientific research scenarios are collected in a unified manner, including material / biological microstructure images, physicochemical experimental detection spectrum images, scientific research papers / industry standard texts, and physicochemical experimental numerical data, to construct the original multimodal scientific research dataset.

[0022] (2) Image data preprocessing: Four standardization steps are performed in sequence: 1) Image noise reduction: Gaussian filtering algorithm is used to eliminate salt and pepper noise and Gaussian noise generated during image acquisition and equipment detection, while retaining the effective features of microstructure and spectral peaks; 2) Adaptive cropping: Based on the effective area boundary of the scientific research image, invalid blank areas are automatically cropped to unify the image aspect ratio; 3) Size normalization: All images are uniformly scaled to the model adaptation size (224×224 / 384×384); 4) Pixel normalization: Image pixel values ​​are linearly mapped from [0,255] to the [0,1] interval to eliminate the influence of pixel scale differences.

[0023] (3) Text-type data preprocessing: Perform fine-grained processing on research abstracts, experimental records, and industry standard texts: 1) Professional word segmentation: Adopt a professional word segmenter adapted to the fields of physics, chemistry, biology and materials to achieve accurate word segmentation of scientific research terms and proper nouns, and avoid professional semantic fragmentation caused by general word segmentation; 2) Entity extraction: Extract core scientific research entities such as material components, molecular structures, experimental parameters, and physical features based on a domain dictionary; 3) Stop word removal: Filter general stop words such as "de", "he", "jiyu" and meaningless function words; 4) Text vectorized encoding: Convert the processed professional text into semantic vectors of fixed dimensions to complete standardized text representation.

[0024] (4) Numerical-type data preprocessing: Process quantitative data such as experimental temperature, pressure, concentration, mechanical parameters, and spectral values: 1) Outlier elimination: Adopt the 3σ criterion to eliminate abnormal value points caused by experimental operation errors and equipment jitter; 2) Standardization processing: Use the Z-Score standardization method to complete numerical normalization and eliminate dimensional differences; 3) Missing value imputation: For a small amount of missing experimental data, adopt the mean interpolation method of similar experiments under the same working condition for imputation to maintain the continuity of experimental data.

[0025] Through the above modal-specific fine-grained preprocessing, the input formats and data quality of the three types of heterogeneous data are unified, forming a standardized multi-modal scientific research dataset, which avoids the problems of model training failure and feature extraction distortion caused by data defects.

[0026] S2. Performing feature extraction on the preprocessed multi-source and multi-modal data to obtain heterogeneous multi-modal features; introducing a cross-modal attention alignment mechanism to map the heterogeneous multi-modal features to the same high-dimensional representation space to obtain a unified multi-modal fused feature set; In the step S2, the generation process of the unified multi-modal fused feature set is specifically as follows: Performing shallow feature extraction on the preprocessed multi-source and multi-modal data; Introducing a cross-modal attention alignment mechanism to establish associative mapping among image visual features, text semantic features, and numerical statistical features, and performing unified multi-modal fused feature construction to obtain a unified multi-modal fused feature set The function expression of the shallow feature extraction is: ; Wherein, is the image shallow feature, is the text shallow feature, is the numerical shallow feature, is the image encoder, is the text encoder, is the numerical encoder, is the preprocessed image data, For preprocessed text data, This is the preprocessed numerical data.

[0027] The functional expression for the cross-modal attention alignment mechanism is: ; in, , For any two different modal features, d For feature dimension, This is the cross-modal attention weight matrix.

[0028] The functional expression for constructing unified multimodal fusion features is: ; in, For globally unified representation features, This is the cross-modal attention weight matrix between images and text. For shallow image features, This is the cross-modal attention weight matrix between images and numerical values. These are shallow numerical features. This is the cross-modal attention weight matrix between text and numerical values. These are shallow text features.

[0029] This fusion method abandons the traditional simple splicing mode, strengthens the cross-modal correlation features of scientific research data through attention weights, weakens irrelevant noise features, completely eliminates the semantic gap of heterogeneous data, and outputs a multimodal fusion feature set with unified dimensions and accurate correlation.

[0030] S3. Construct a few-shot generalization basic model based on a meta-learning framework, and inject prior knowledge from the knowledge graph of basic scientific research fields into the few-shot generalization basic model. Based on the metric learning strategy, the unified multimodal fusion feature set is used to iteratively train the small-sample generalization base model with injected prior knowledge. In step S3, the process of injecting prior knowledge from the knowledge graph of basic scientific research into the feature inference layer of the small-sample generalization basic model is as follows: Construct a knowledge graph in the field of basic scientific research, and take the relationship between scientific theorems, experimental laws and professional terms as prior knowledge; By incorporating domain prior knowledge into the feature inference layer of the few-shot generalization basic model through the knowledge embedding layer, the feature learning direction of the few-shot generalization basic model is constrained, so that the few-shot generalization basic model conforms to the professional logic of basic scientific research.

[0031] Specifically, this embodiment involves constructing a knowledge graph of basic scientific research fields such as materials, chemistry, biology, and physics, and using the relationships between scientific research theorems, experimental laws, and professional terms as prior knowledge. Through a knowledge embedding layer, the prior knowledge of the fields is integrated into the feature reasoning layer of a small-sample generalization model, which constrains the feature learning direction of the model, makes the model conform to the professional logic of basic scientific research, and improves the adaptability of the model to professional tasks under small samples.

[0032] In S3, the process of iteratively training the small-sample generalization base model infused with prior knowledge based on the metric learning strategy is as follows: A small-sample generalization basic model is built based on the meta-learning framework. The labeled scientific samples are divided into a meta-training set and a meta-test set by using a support set-query set splitting method. A metric learning strategy is introduced to calculate the similarity distribution among features of different research samples, construct a loss function for small sample classification and regression tasks, and use labeled samples for iterative training to optimize the feature extraction and metric parameters of the small sample generalization base model.

[0033] This embodiment specifically involves: building a small-sample generalization base model based on a meta-learning framework; using a support set-query set splitting method to divide a small number of labeled research samples into a meta-training set and a meta-test set; introducing a metric learning strategy to calculate the similarity distribution between features of different research samples and constructing loss functions for small-sample classification and regression tasks; using a small number of labeled samples to iteratively train the small-sample generalization base model, optimizing the feature extraction and metric parameters of the small-sample generalization base model, and avoiding overfitting during small-sample training.

[0034] For fundamental scientific research in materials, chemistry, biology, and physics, experimental annotation costs are extremely high. Two sets of feasible parameters are selected: In a standard scenario: 5 labeled samples (5-shot) for each type of scientific research task; Extremely scarce scenario: 1 labeled sample per research task (1-shot); if described by the total amount of the dataset: the total number of labeled samples for a single research task does not exceed 50, which is judged as "a small number of labeled samples".

[0035] This embodiment contrasts with scenarios where industrial big data often involves thousands or tens of thousands of labeled data points; when there are fewer than a few dozen samples per class, which cannot support conventional deep learning training, this is considered a "small number" in this embodiment.

[0036] The ratio of training set to test set in a standard 5-shot scenario is as follows: 1. A small number of labeled samples: 5 labeled samples for each type of research task, for a total of 30-50 labeled samples across all categories; 2. Partition ratio (support set: query set = 1:2): (1) Support set S: 5 items per class (training fit); (2) Query set Q: 10 items per class (to test and verify generalization).

[0037] In extremely scarce (1-shot) scenarios, the ratio of training to test sets is as follows: 1. Limited labeled samples: Only one labeled sample per research task; 2. Partition ratio (support set: query set = 1:3): (1) Support set S: 1 item per class; (2) Query set Q: 3 records for each class.

[0038] 1) Optimal ratio: 1:2, 5 support sets + 10 query sets per category; 2) Lightweight ratio (when samples are extremely scarce): 1:1 5 support sets + 5 query sets per class.

[0039] The iterative training process for the small-sample generalization base model is as follows: (1) Small sample dataset splitting rules For scarce labeled research samples, a classic support set-query set splitting model of meta-learning is adopted: a small number of labeled samples for each type of research task are randomly divided into a support set S and a query set Q; the support set is the reference sample for model learning, used to fit the basic feature patterns of the research task; the query set is the test sample, used to verify the model's generalization ability and iteratively optimize the parameters.

[0040] (2) Meta-training process 1) Task Sampling: Randomly sample multiple groups of different research sub-tasks from the multidisciplinary basic scientific research task pool to construct meta-training batches; 2) Feature metric calculation: Introduce a metric learning strategy to calculate the Euclidean similarity of features between the query set samples and the support set samples; ; in, For query set features, To support set features; 3) Loss Function Construction: Combining scientific research classification and regression task characteristics, a joint loss function for small samples is constructed: ; in, For the classification loss of scientific research samples, For experimental numerical regression loss, This is the balance coefficient; 4) Iterative training: Based on a small number of samples, backpropagation is used to iteratively update the model encoder, attention weights, and metric parameters, constraining the model's convergence direction and suppressing overfitting on small samples.

[0041] (3) Meta-testing process 1) Select novel, untrained samples from niche / cutting-edge scientific research tasks as the test set; 2) Input the model to complete feature matching and pattern fitting; 3) Calculate the prediction accuracy and error index of the test set to verify the model's cross-task generalization ability; 4) Fine-tune the meta-learning rate and loss weights based on the test results to optimize the model's generalization performance.

[0042] S4. Input the multimodal scientific research data to be studied into the trained small sample generalization basic model, perform scientific research task generalization inference, and output a multimodal AI4S basic scientific research analysis report. In S4, the generalized reasoning of the scientific research task includes: feature matching, pattern fitting, structure prediction, and literature knowledge association mining; The research analysis results include: research data classification results, experimental trend predictions, and derivation conclusions of potential scientific laws, forming a standardized AI4S basic research analysis report.

[0043] Specifically, this embodiment involves inputting novel multimodal scientific research data to be studied into a trained model, utilizing small-sample generalization capabilities to perform feature matching, pattern fitting, structure prediction, and literature knowledge association mining for unknown samples; and outputting scientific research data classification results, experimental trend predictions, and conclusions deducing potential scientific laws to form a standardized AI4S basic scientific research analysis report.

[0044] S5. Using the data from the scientific research analysis results that have been manually verified, dynamically update the parameters of the small sample generalization base model to achieve iterative optimization of the model.

[0045] This embodiment specifically involves: using the results manually verified by researchers as incremental small-sample data, which is then fed back into the model training set; employing incremental small-sample fine-tuning to dynamically update model parameters and continuously improve the model's generalization ability in interdisciplinary and emerging basic research tasks; the specific parameter update and iteration process is as follows: (1) Incremental data backflow: The scientific reasoning results that have been manually verified and corrected by researchers are used as new incremental small sample data and automatically backflowed into the model training dataset to expand the sample library of niche and cutting-edge scientific research tasks.

[0046] (2) Precise parameter dynamic update: The incremental small sample fine-tuning strategy is adopted to freeze the basic feature extraction parameters of the model and only fine-tune and update four core parameters: cross-modal attention weight parameters, meta-learning metric parameters, knowledge prior embedding weights, and loss function balance coefficients, so as to avoid the catastrophic forgetting problem caused by full parameter training.

[0047] (3) Adaptive adjustment mechanism: The learning rate is adaptively adjusted according to the subject type, sample scarcity and inference error of the incremental data: a small high learning rate is used for fine-tuning of new scientific research tasks, and a low learning rate is used for fine optimization of similar iterative tasks, so as to continuously improve the generalization reasoning ability of the model across disciplines, new scenarios and new tasks, and realize the long-term dynamic evolution of the model.

[0048] Compared with the prior art, the present invention also has the following technical effects: (1) This invention maps heterogeneous scientific research data such as images, text, and numerical values ​​to a unified feature space through a multimodal attention alignment fusion mechanism, which overcomes the defect of low feature correlation caused by simple multimodal splicing in the prior art and effectively improves the accuracy of basic scientific research data pattern mining. (2) The present invention adopts a meta-learning small sample generalization architecture, which only requires a small number of professionally labeled samples to complete model training, thus getting rid of the dependence of existing technologies on large-scale labeled datasets and greatly reducing the cost of scientific research sample labeling and the experimental trial and error cycle; (3) This invention embeds prior knowledge in basic scientific research fields, constrains the model feature learning logic, solves the problems of poor professional adaptability of general large models and easy overfitting of small samples, and significantly improves the model's generalization and transfer ability to unknown frontier scientific research tasks. (4) This invention constructs a standardized AI4S full-process scientific research processing link, which fits the business logic of basic scientific research and can be directly adapted to basic research scenarios in multiple disciplines such as materials, biology, physics, and chemistry, and has strong practicality. (5) The present invention establishes an incremental small sample iterative optimization mechanism, which can continuously optimize the model using incremental scientific research data, and realize the long-term adaptation of the model to the research needs of emerging basic scientific research directions.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multimodal AI4S basic scientific research method based on few-shot generalization learning, characterized in that, Includes the following steps: S1. Collect multi-source multimodal data in basic scientific research scenarios, perform preprocessing, and obtain preprocessed multi-source multimodal data; S2. Extract features from the preprocessed multi-source multimodal data to obtain heterogeneous multimodal features; A cross-modal attention alignment mechanism is introduced to map heterogeneous multimodal features to the same high-dimensional representation space, resulting in a unified multimodal fusion feature set; S3. Construct a few-shot generalization basic model based on a meta-learning framework, and inject prior knowledge from the knowledge graph of basic scientific research fields into the few-shot generalization basic model. Based on the metric learning strategy, the unified multimodal fusion feature set is used to iteratively train the small-sample generalization base model with injected prior knowledge. S4. Input the multimodal scientific research data to be studied into the trained small sample generalization basic model, perform scientific research task generalization inference, and output a multimodal AI4S basic scientific research analysis report. S5. Using the aforementioned multimodal AI4S basic scientific research analysis report, dynamically update the parameters of the small-sample generalization basic model to achieve iterative optimization of the small-sample generalization basic model.

2. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 1, characterized in that, In S1, the multi-source, multi-modal data includes: microstructure images, experimental detection maps, scientific research paper texts, physicochemical experimental values, and industry standard texts; The preprocessing process includes: noise reduction, cropping, and normalization of the microstructure images and experimental detection maps; Perform word segmentation, entity extraction of professional terms, and stop word removal on scientific research paper texts and industry standard texts; Standardize and remove outliers from physicochemical experimental values.

3. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 1, characterized in that, In S2, the process of generating the unified multimodal fusion feature set is as follows: Shallow feature extraction is performed on the preprocessed multi-source multimodal data; A cross-modal attention alignment mechanism is introduced to establish a correlation mapping between image visual features, text semantic features, and numerical statistical features, and a unified multimodal fusion feature is constructed to obtain a unified multimodal fusion feature set.

4. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 3, characterized in that, The functional expression for the shallow feature extraction is: ； in, For shallow image features, These are shallow text features. These are shallow numerical features. For image encoders, For text encoders, It is a numerical encoder. For preprocessed image data, For preprocessed text data, This is the preprocessed numerical data.

5. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 3, characterized in that, The functional expression for the cross-modal attention alignment mechanism is: ； in, , For any two different modal features, d For feature dimension, This is the cross-modal attention weight matrix.

6. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 3, characterized in that, The functional expression for constructing unified multimodal fusion features is: ； in, For globally unified representation features, This is the cross-modal attention weight matrix between images and text. For shallow image features, This is the cross-modal attention weight matrix between images and numerical values. These are shallow numerical features. This is the cross-modal attention weight matrix between text and numerical values. These are shallow text features.

7. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 1, characterized in that, In step S3, the process of injecting prior knowledge from the knowledge graph of basic scientific research into the small-sample generalization basic model is specifically as follows: Construct a knowledge graph in the field of basic scientific research, and take the relationship between scientific theorems, experimental laws and professional terms as prior knowledge; By incorporating domain prior knowledge into the feature inference layer of the few-shot generalization basic model through the knowledge embedding layer, the feature learning direction of the few-shot generalization basic model is constrained, so that the few-shot generalization basic model conforms to the professional logic of basic scientific research.

8. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 1, characterized in that, In S3, the process of iteratively training the small-sample generalization base model with injected prior knowledge is as follows: A small sample generalization basic model is built based on the meta-learning framework, and a metric learning strategy is introduced to calculate the similarity distribution between features of different scientific research samples and construct a loss function for small sample classification and regression tasks. Iterative training using labeled samples optimizes the feature extraction and metric parameters of the small-sample generalization base model.

9. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 1, characterized in that, In S4, the generalized reasoning process of the scientific research task includes: feature matching, pattern fitting, structure prediction, and literature knowledge association mining; The multimodal AI4S basic scientific research analysis report includes: scientific research data classification results, experimental trend predictions, and conclusions derived from potential scientific laws.

10. The multimodal AI4S basic scientific research method based on few-shot generalization learning according to claim 1, characterized in that, In S5, the implementation process of iterative optimization of the small-sample generalization base model is as follows: The results manually verified by researchers are used as incremental small sample data and fed back into the training set of the small sample generalization base model. The parameters of the small-sample generalization basic model are dynamically updated by incremental small-sample fine-tuning, continuously improving the generalization ability of the small-sample generalization basic model in interdisciplinary and new-direction basic scientific research tasks.

Images