AI-based material performance spectrum detection system and method

Abstract

The invention discloses a material performance spectrum detection system and method based on AI, and relates to the field of artificial intelligence, and the system comprises a spectrogram embedding module, a feature fusion module, a performance prediction module, an uncertainty evaluation module and an output analysis module. According to the method, multi-source information such as spectral data, microstructure images and material composition is fused, material performance characteristics are modeled through a multi-modal attention mechanism, and the integrity of characteristic expression and the accuracy of a prediction model are improved. A heterogeneous neural network structure is constructed, effective representation of high-dimensional spectrogram data is realized, and the feature learning ability is enhanced through a residual fusion mechanism. Through an attention hot area map, an attention distribution map and an LIME causal analysis method, key feature interpretation of each prediction result is provided, and scientificity and reliability of a prediction model are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-based material property spectral detection system and method. Background Art

[0002] With the development of hyperspectral technology and microscopic characterization techniques, the absorption, reflection, and emission characteristics of materials in different spectral bands can be accurately captured, providing a wealth of potential information for predicting material properties. However, spectral data is often highly dimensional and redundant, making it difficult to directly reveal complex nonlinear relationships with properties. Furthermore, relying solely on spectral information often fails to fully characterize a material's structural characteristics and macroscopic performance behavior. In recent years, artificial intelligence, particularly deep learning, has demonstrated superior modeling capabilities in fields such as image recognition, speech processing, and medical diagnosis, and its application in materials science has gradually gained attention. For example, some studies have used convolutional neural networks to process microscopic images of materials or fully connected networks for regression prediction of chemical composition. However, existing technologies have the following shortcomings: most methods utilize only a single modality (such as spectral or image data) and fail to achieve collaborative modeling of multi-source information; traditional neural network models lack interpretable mechanisms and cannot reveal causal relationships between characteristics and properties; existing methods typically only output point predictions and lack estimates of model uncertainty or prediction confidence; and they are poorly adaptable to out-of-distribution samples or novel materials, preventing effective dynamic recognition and feedback.

[0003] Therefore, there is an urgent need for a multimodal intelligent analysis method that integrates spectral information, microstructural characteristics and image data, which can efficiently and accurately predict material properties and provide visual explanations and uncertainty assessments to improve the usability and credibility of the prediction results. Summary of the Invention

[0004] In view of the problems of insufficient utilization of existing information, unexplainable predictions, and lack of uncertainty assessment of results, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to make full use of multimodal material information such as spectrum, image and structure to improve the comprehensiveness and accuracy of material performance prediction.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides an AI-based material property spectral detection system, which includes a spectrum embedding module for acquiring material spectral data, inputting the spectral data into a heterogeneous neural network with a dual-channel structure, one channel being a local feature extraction network based on a CNN, and the other channel being a global dependency modeling network based on a transformer, outputting a multi-scale spectrum embedding vector, and performing a weighted combination of the different scale embedding representations through an attention residual fusion module to generate a dynamically regulated spectrum feature vector; The feature fusion module is used to obtain the material's microstructure, chemical composition, and image data. It uses a multimodal attention fusion mechanism to fuse the spectral feature vector and structural features into a unified feature vector for modeling performance prediction. A performance prediction module is used to construct a deep performance prediction network model based on the fused unified feature vector. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework; The uncertainty assessment module is used to estimate the uncertainty of the performance prediction results output by the deep performance prediction network model and evaluate the confidence interval of each performance through the Bayesian regression network. If the model confidence falls below the set threshold, the dynamic correction mechanism is triggered. The output analysis module is used to output the prediction performance indicators in the form of numerical values and credible intervals. It is combined with the visualization module to display the contribution of the spectrum hot zone and the model attention distribution map, and generate a causal explanation for each prediction result based on the LIME method.

[0007] As a preferred solution of the AI-based material property spectral detection system of the present invention, spectral data is input into a heterogeneous neural network with a dual-channel structure, one channel is a local feature extraction network based on CNN, and the other channel is a global dependency modeling network based on a converter, and the multi-scale spectrum embedding vector is output, including: Multiple one-dimensional convolutional layers with different kernel widths are used to extract the characteristic patterns of local bands. Each convolution layer is followed by an activation function ReLU and BatchNorm layer for normalization, nonlinear processing, and stable training. The convolution output is max-pooled to form a multi-layer semantically nested spectral local feature representation. The spectral vector is divided into sequences of length n, and the spectral vector embedding is generated through the embedding layer. The embedded representation is input into the standard transformer encoder module, and the multi-head self-attention mechanism is used to model the global dependency between different bands. The fusion is performed by splicing or weighting, and the attention fusion mechanism is used to calculate the fusion weight α to generate the final spectrogram multi-scale embedding vector:

[0008] Among them, Z is the final fused spectrum embedding vector, α is the weight coefficient, and Z CNN is the spectral feature vector output by the CNN branch, is the spectral feature vector output by the converter branch, and R is the cross-channel residual correction term.

[0009] As a preferred solution of the AI-based material performance spectral detection system of the present invention, wherein: and using the attention residual fusion module to perform weighted combination of embedding representations of different scales to generate a dynamically controlled spectral feature vector, including: The concatenated vector is input into a lightweight attention network module to generate learnable weighted parameters. The attention coefficient is used to dynamically weight the embedding representations of the two channels. The residual compensation module is introduced to generate the residual concatenation form:

[0010] Among them, R is the residual compensation term, γ is the residual weight coefficient; The weighted fusion result is added to the residual term to obtain the final spectrogram feature vector.

[0011] As a preferred solution of the AI-based material performance spectral detection system described in the present invention, the microstructure, chemical composition and image data of the material are obtained, and a multimodal attention fusion mechanism is used to fuse the spectral feature vector and the structural feature into a unified feature vector for modeling performance prediction, including: For the three types of modalities, their embedding representations are extracted separately, including the spectral embedding vector Z spec , image embedding vector Z img and component embedding vector Z chem , perform weighted fusion of the three modal features to generate a unified fusion vector Z fusion , the fusion method is as follows:

[0012] in, , , is the modal attention weight, and the fusion vector Z is input into the performance prediction model f θ , output the corresponding material performance index, the prediction expression is as follows:

[0013] in, is the predicted performance value, f θ Model the trained performance of the network.

[0014] As a preferred solution of the AI-based material performance spectral detection system described in the present invention, a deep performance prediction network model is constructed based on the fused unified feature vector. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework, including: The graph structure is input into the graph convolutional network to perform context modeling on the spectral feature vector. The expression of GCN is as follows:

[0015] in, is the adjacency matrix with self-loops added, for The degree matrix of is the input node feature of the lth layer, is the learnable weight of the lth layer, σ is the activation function; Embed the spectrogram into Z GCN Combine or fuse features with other modalities into a new high-dimensional vector Z all , input a shared deep neural network encoder to generate shared semantic representation; each target performance indicator corresponds to a task head network, and the structure includes regression tasks and classification tasks.

[0016] As a preferred solution of the AI-based material performance spectral detection system described in the present invention, uncertainty estimation is performed on the performance prediction results output by the deep performance prediction network model, and the confidence interval of each performance is evaluated through a Bayesian regression network. If the model confidence falls below a set threshold, a dynamic correction mechanism is triggered, including: Construct a Bayesian regression prediction head for each performance metric and output the predicted mean and variance:

[0017] Among them, μ ti is the predicted value of the i-th performance indicator, To account for forecast uncertainty, is a normal distribution or a Gaussian distribution; After the prediction output is Gaussian distributed, calculate the confidence interval:

[0018] in, is the prediction interval of the material performance index of item i at the 95% confidence level; The feature vector Z of the current sample fusion Calculate the distance from the training set sample distribution:

[0019] in, is the mean vector of the fusion features in the training set, is the straight-line distance between two eigenvectors in high-dimensional space; If D > the preset offset threshold, it is judged as a new sample or abnormal material; The sample and its actual performance indicators are stored in a cache queue. After accumulating multiple low-confidence samples, fine-tuning or incremental learning is triggered.

[0020] As a preferred solution of the AI-based material performance spectral detection system described in the present invention, the predicted performance indicators are output as numerical values and credible intervals, combined with a visualization module to display the spectral hot zone contribution and model attention distribution map, and a causal explanation of each prediction result is generated based on the LIME method, including: Draw a spectrogram heat map based on the weight information calculated by the attention mechanism in the fusion network; extract the attention weight distribution of the converter network in the model and generate an attention visualization image; For each prediction result, the LIME method is applied to perform causal analysis, including generating multiple perturbation samples near the current input; performing model prediction on the samples and establishing a local linear explanatory model; obtaining the positive and negative contributions of the main input features to the prediction results; and outputting a bar chart or ranked list to illustrate the role of each feature in the current prediction.

[0021] An AI-based material performance spectral detection method, characterized by comprising: Acquire material spectral data and input the spectral data into a heterogeneous neural network with a dual-channel structure. One channel is a local feature extraction network based on CNN, and the other channel is a global dependency modeling network based on transformers. The network outputs multi-scale spectral embedding vectors and performs a weighted combination of the different scale embedding representations through the attention residual fusion module to generate a dynamically controlled spectral feature vector. Obtain the material's microstructure, chemical composition, and image data, and use a multimodal attention fusion mechanism to fuse the spectral feature vector and structural features into a unified feature vector for modeling performance prediction; Based on the fused unified feature vector, a deep performance prediction network model is constructed. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework. The uncertainty of the performance prediction results output by the deep performance prediction network model is estimated, and the confidence interval of each performance is evaluated through the Bayesian regression network; if the model confidence falls below the set threshold, the dynamic correction mechanism is triggered; The prediction performance indicators are output in the form of numerical values and credible intervals, combined with the visualization module to display the contribution of the spectrum hot zone and the model attention distribution map, and generate a causal explanation for each prediction result based on the LIME method.

[0022] In a second aspect, an embodiment of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the AI-based material property spectral detection system as described in the first aspect of the present invention are implemented.

[0023] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the AI-based material property spectral detection system as described in the first aspect of the present invention are implemented.

[0024] The present invention integrates multiple sources of information, including spectral data, microstructure images, and material composition, and models material performance characteristics through a multimodal attention mechanism, improving the integrity of feature representation and the accuracy of the prediction model. A heterogeneous neural network architecture is constructed to effectively represent high-dimensional spectral data, and feature learning capabilities are enhanced through a residual fusion mechanism. Using attention heatmaps, attention distribution maps, and the LIME causal analysis method, key feature explanations for each prediction result are provided, enhancing the scientificity and reliability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 Schematic diagram of the structure of the AI-based material performance spectral detection system.

[0027] Figure 2 Flowchart of the AI-based material performance spectral detection method.

[0028] Figure 3 A diagram of a computer device for an AI-based material property spectral detection system and method. DETAILED DESCRIPTION

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0031] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0032] Example 1 Reference Figures 1 to 3 , which is the first embodiment of the present invention, provides an AI-based material property spectral detection method, including: S1: Obtain material spectral data and input the spectral data into a heterogeneous neural network with a dual-channel structure. One channel is a local feature extraction network based on CNN, and the other channel is a global dependency modeling network based on transformer. It outputs multi-scale spectral embedding vectors and performs weighted combination of embedding representations of different scales through the attention residual fusion module to generate a dynamically controlled spectral feature vector.

[0033] Preferably, multiple one-dimensional convolutional layers with different convolution kernel widths are used to extract the characteristic patterns of the local band. Each convolution layer is followed by an activation function ReLU and a BatchNorm layer for normalization, nonlinear processing and stable training. The convolution output is subjected to maximum pooling to form a multi-layer semantically nested spectral local feature representation. The spectral vector is divided into sequences of length n, and the spectral vector embedding is generated through the embedding layer. The embedded representation is input into the standard transformer encoder module, and the multi-head self-attention mechanism is used to model the global dependency between different bands. The fusion is performed by splicing or weighting, and the attention fusion mechanism is used to calculate the fusion weight α to generate the final spectrogram multi-scale embedding vector:

[0034] Among them, Z is the final fused spectrum embedding vector, α is the weight coefficient, and Z CNN is the spectral feature vector output by the CNN branch, is the spectral feature vector output by the converter branch, and R is the cross-channel residual correction term.

[0035] Preferably, the concatenated vector is input into a lightweight attention network module to generate learnable weighted parameters, the attention coefficient is used to dynamically weight the embedding representations of the two channels, and a residual compensation module is introduced to generate a residual concatenation form:

[0036] Among them, R is the residual compensation term, γ is the residual weight coefficient; The weighted fusion result is added to the residual term to obtain the final spectrogram feature vector.

[0037] Furthermore, the standardized spectral vector is input into the one-dimensional convolutional neural network (1D-CNN) branch, which specifically includes the following substructures: multi-scale convolution layer: multiple 1D convolution layers with different convolution kernel widths (such as 3, 5, 7, etc.) are set to extract local pattern features in different bands of the spectrum; activation and normalization processing: each convolution layer is followed by a ReLU activation function and a Batch Normalization layer to increase the nonlinear expression ability and stabilize the training process; downsampling operation: the maximum pooling operation is used to reduce the dimension of the feature map, retain the strongest local signal, and improve the abstractness of the feature expression; stacking to build a deep semantic structure: after multiple convolution blocks are stacked, a hierarchical local feature representation is formed, and the final output is the local feature vector Z of the spectrum. CNN .

[0038] The same spectral vector is fed into the transformer branch for long-range dependency modeling as follows: The spectral vector is divided into several sub-segments of fixed length n (e.g., 10 dimensions per segment), and each segment is mapped into an embedding vector through an embedding layer. The embedding sequence is input into the standard converter encoder module, and the multi-head self-attention mechanism is used to establish the dependency relationship between different bands. After position encoding, feedforward network and layer normalization, the global spectrogram embedding representation Z is output. 转换器 .

[0039] The output vectors of the two branches are fused in the following way: an adjustable weight α∈[0,1] is learned through the attention fusion mechanism to adjust the relative contribution of local features and global features; a cross-channel residual term R is introduced to fine-tune the fusion result to enhance robustness and information expression integrity.

[0040] S2: Obtain the microstructure, chemical composition and image data of the material, and use a multimodal attention fusion mechanism to fuse the spectral feature vector and structural features into a unified feature vector for modeling performance prediction.

[0041] Preferably, for the three types of modalities, their embedding representations are extracted separately, including the spectral embedding vector Z spec , image embedding vector Z img and component embedding vector Zchem , perform weighted fusion of the three modal features to generate a unified fusion vector Z fusion , the fusion method is as follows:

[0042] in, , , is the modal attention weight, and the fusion vector Z is input into the performance prediction model f θ , output the corresponding material performance index, the prediction expression is as follows:

[0043] in, is the predicted performance value, f θ Model the trained performance of the network.

[0044] Furthermore, the following three types of data are collected: infrared or Raman spectra of materials, images of material microstructures, and material composition tables. The spectra are normalized (0-1) and smoothed to reduce noise; the images are grayscaled, center-cropped, and scaled to 224×224; and the chemical composition vectors are Z-score standardized.

[0045] A two-channel neural network architecture is used to extract spectrogram features: the CNN branch uses a 1D-CNN with three different convolution kernel widths (3, 5, and 7) to extract local spectral information; the transformer branch uses a standard transformer encoder to model the dependencies between long-range bands in the spectrogram. The outputs of the two branches are fused to obtain a 64-dimensional spectrogram embedding vector. ResNet18 is used to extract features from the image, and a 64-dimensional image vector is output through a GAP and fully connected layer. The element composition (e.g., [C=43%, O=52%, H=5%]) is mapped to a 3D input. A 64-dimensional embedding representation is output through FC+ReLU+BN.

[0046] The three modal vectors are concatenated and input into the MLP module to generate attention weights. The weighted summation is performed to obtain a unified fusion vector. The modal weights are dynamically allocated to improve robustness and have a certain fault tolerance for missing modal data.

[0047] S3: Based on the fused unified feature vector, a deep performance prediction network model is constructed. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework.

[0048] Preferably, the graph structure is input into the graph convolutional network to perform context modeling on the spectral feature vector. The expression of GCN is as follows:

[0049] in, is the adjacency matrix with self-loops added, for The degree matrix of is the input node feature of the lth layer, is the learnable weight of the lth layer, σ is the activation function; Embed the spectrogram into Z GCN Combine or fuse features with other modalities into a new high-dimensional vector Z all , input a shared deep neural network encoder to generate shared semantic representation; each target performance indicator corresponds to a task head network, and the structure includes regression tasks and classification tasks.

[0050] Furthermore, each band is regarded as a node in the graph; the following types of edges are established to connect band nodes: strong edges are established between adjacent bands (simulating local continuity); weak edges are constructed based on physical or chemical co-occurrence rules (such as belonging to the same functional group); self-loop edges are added for information retention; an adjacency matrix is constructed and normalized to stabilize training; the constructed spectrogram is input into the graph convolutional network; GCN extracts the contextual features of each band in the graph structure through a local weighted aggregation mechanism; after multi-layer GCN processing, global average pooling is used to obtain the final spectrogram embedding vector.

[0051] The spectral embedding vector is obtained by using networks such as GCN; the texture and morphology features of the microstructure image are extracted through a lightweight convolutional neural network; the element composition is input into the MLP network to obtain the component embedding representation; the above three features are uniformly mapped to the same dimensional space; a weight factor α is learned for each mode modality ; The model can adaptively adjust the modal contribution according to the input content; all modal features are weighted and fused into a unified feature vector Z fusion , as the final material representation.

[0052] Based on a shared deep encoder, an independent task head is constructed for each performance indicator. Task head structure types include regression head and classification head. Multiple task heads can be trained in parallel and jointly optimized using a multi-task loss function. Some task heads can be flexibly called during deployment to adapt to specific application scenarios.

[0053] S4: Estimate the uncertainty of the performance prediction results output by the deep performance prediction network model, and evaluate the credibility interval of each performance through the Bayesian regression network; if the model confidence is lower than the set threshold, the dynamic correction mechanism is triggered.

[0054] Preferably, a Bayesian regression prediction head is constructed for each performance metric, outputting the predicted mean and variance:

[0055] Among them, μ ti is the predicted value of the i-th performance indicator, To account for forecast uncertainty, is a normal distribution or a Gaussian distribution; After the prediction output is Gaussian distributed, calculate the confidence interval:

[0056] in, is the prediction interval of the material performance index of item i at the 95% confidence level; The feature vector Z of the current sample fusion Calculate the distance from the training set sample distribution:

[0057] in, is the mean vector of the fusion features in the training set, is the straight-line distance between two eigenvectors in high-dimensional space; If D > the preset offset threshold, it is judged as a new sample or abnormal material; The sample and its actual performance indicators are stored in a cache queue. After accumulating multiple low-confidence samples, fine-tuning or incremental learning is triggered.

[0058] Furthermore, the offset judgment threshold τ=5.0 is set; if D>τ, it is judged as a "offset sample" or "potential new material"; combined with the model output variance σ: if σ>1.5 or the CI span exceeds the set value, the confidence is too low; further strengthen the confidence of offset judgment.

[0059] Add the current low confidence sample to the cache queue Q; if the cache queue Q meets any of the conditions: the number of samples ≥ 20; the average σ 2 > threshold; if the prediction error continues to deviate, the dynamic update process is triggered.

[0060] Samples in the cache queue Q are merged into the training data, and the model is fine-tuned. If the sample types vary significantly, independent "sub-models" are trained. After the update, the confidence interval distribution of all prediction heads is re-estimated. The system ultimately returns the predicted value and its confidence interval for each material property. New material input samples are highlighted and user verification is recommended. Spectral visualizations of important wavelengths and feature attention are provided. All model updates are automatically recorded, supporting backtracking and version comparison.

[0061] S5: Output the prediction performance indicators in the form of numerical values and credible intervals, combine with the visualization module to display the spectrum hot zone contribution and model attention distribution map, and generate a causal explanation for each prediction result based on the LIME method.

[0062] Preferably, a spectrogram heat map is drawn based on the weight information calculated by the attention mechanism in the fusion network; the attention weight distribution of the converter network in the model is extracted to generate an attention visualization image; For each prediction result, the LIME method is applied to perform causal analysis, including generating multiple perturbation samples near the current input; performing model prediction on the samples and establishing a local linear explanatory model; obtaining the positive and negative contributions of the main input features to the prediction results; and outputting a bar chart or ranked list to illustrate the role of each feature in the current prediction.

[0063] Furthermore, the multi-head attention weight matrix A∈R^{H×L×L} is obtained from the k-th layer of the encoder; Where H is the number of heads and L is the length of the input position. Aggregate or plot the attention matrices of different heads in layers; use heatmap visualization to show the degree of attention of a certain band to other bands; support interactive operations, such as selecting a position to view the direction of its attention diffusion. Images are used to analyze the global and local dependencies of different layers and heads; help explain the model's perception mechanism of long-range spectral patterns. Fix the current sample and perform several small feature perturbations near it; set the ±ε fluctuation range for the key dimensions of each modality; generate n perturbation versions. Run the prediction model for each perturbation sample to obtain the corresponding predicted values y'_1, y'_2, ..., y'_n; and record the perturbation position and amplitude at the same time. Fit a local weighted linear model with the perturbation input and prediction results as samples; draw a bar chart to represent the positive and negative contributions of the spectral band / component / image area; The sorted list outputs the most critical features (e.g., the element Fe has a negative effect on thermal conductivity, while the 850nm band has a positive effect).

[0064] This embodiment also provides a computer device suitable for the case of an AI-based material property spectral detection system, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the AI-based material property spectral detection system proposed in the above embodiment.

[0065] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0066] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the AI-based material property spectral detection system proposed in the above embodiment is implemented.

[0067] In summary, the present invention integrates multi-source information such as spectral data, microstructure images, and material composition, and models material performance characteristics through a multimodal attention mechanism, thereby improving the integrity of feature expression and the accuracy of the prediction model. A heterogeneous neural network structure is constructed to achieve effective representation of high-dimensional spectral data, and the feature learning capability is enhanced through a residual fusion mechanism. Through attention heat map, attention distribution map, and LIME causal analysis method, an explanation of the key features of each prediction result is provided, enhancing the scientificity and reliability of the prediction model.

[0068] Example 2 This embodiment is the second embodiment of the present invention. This embodiment provides an AI-based material property spectral detection system. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0069] Specifically, step 1: Represent each spectral vector as a graph structure: nodes are frequency bands (a total of 512); edges connect adjacent frequency bands, and self-loops are added to form a symmetric adjacency matrix; input the graph structure to a two-layer GCN graph convolutional network, and output the graph embedding vector Z spec ∈R^128.

[0070] Step 2: Input the microstructure image into the lightweight CNN to extract the multi-scale features of the intermediate layer; use the attention pooling mechanism to integrate the image context representation as Z img ∈R^128.

[0071] Step 3: Input the component vector into the multi-layer perceptron (MLP) encoder and embed it into Z chem ∈R^128; Enable Dropout and ReLU activation to improve generalization.

[0072] Step 4: The three-modal vector fusion method introduces the residual compensation term R to correct feature offset and improve the fusion representation capability.

[0073] Step 5: Performance prediction network (multi-task) builds a shared deep encoder f encoder →H; the output layer consists of three task heads: thermal conductivity: Bayesian regression prediction head; hardness: regression prediction head; ductility: classification prediction head (3 levels: low / medium / high); each task head outputs the predicted mean + prediction uncertainty or classification probability.

[0074] Step 6: For each prediction, output the Gaussian distribution parameters (μ, σ); calculate the 95% confidence interval and record the prediction confidence; if the confidence is low (for example, the sample confidence probability is < 80%), record it as a "low confidence sample" and trigger the subsequent incremental learning process.

[0075] Experimental results and verification

[0076] In the case of unseen samples, the average confidence interval width was expanded by 1.3 times, enhancing the confidence difference; the model gave low confidence to high deviation samples, with an accuracy rate exceeding 90%. Material A: The predicted thermal conductivity value is 10.2, and the most contributing feature is "spectrum 1200cm -1 Peak Strength"; Material B: Elongation is classified as "High", mainly dominated by "Image Particle Sparsity" and "High Ni Content".

[0077] This example demonstrates the remarkable effectiveness of the method of the present invention in improving the accuracy of material property predictions, enhancing model stability and credibility, and achieving interpretable output. Through the introduction of spectral structure modeling, multimodal fusion, and the LIME interpretation mechanism, the method significantly outperforms existing single-modality or black-box model solutions. The method has the potential to be widely used in material design, property screening, experimental assistance, and other fields.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An AI-based material performance spectrum detection system, characterized in that: include: The spectral embedding module is used to obtain material spectral data and input the spectral data into a heterogeneous neural network with a dual-channel structure. One channel is a local feature extraction network based on CNN, and the other channel is a global dependency modeling network based on transformers. It outputs multi-scale spectral embedding vectors and performs weighted combination of the different scale embedding representations through the attention residual fusion module to generate a dynamically controlled spectral feature vector. The feature fusion module is used to obtain the material's microstructure, chemical composition, and image data. It uses a multimodal attention fusion mechanism to fuse the spectral feature vector and structural features into a unified feature vector for modeling performance prediction. A performance prediction module is used to construct a deep performance prediction network model based on the fused unified feature vector. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework; Uncertainty assessment module, used to estimate the uncertainty of the performance prediction results output by the deep performance prediction network model, and evaluate the credible interval of each performance through the Bayesian regression network; If the model confidence is lower than the set threshold, the dynamic correction mechanism is triggered; The output analysis module is used to output the prediction performance indicators in the form of numerical values and credible intervals. It is combined with the visualization module to display the contribution of the spectrum hot zone and the model attention distribution map, and generate a causal explanation for each prediction result based on the LIME method.

2. The AI-based material performance spectrum detection system according to claim 1, characterized in that: The spectral data is input into a heterogeneous neural network with a dual-channel structure. One channel is a local feature extraction network based on CNN, and the other channel is a global dependency modeling network based on a converter. The multi-scale spectrum embedding vector is output, including: Multiple one-dimensional convolutional layers with different kernel widths are used to extract the characteristic patterns of local bands. Each convolution layer is followed by an activation function ReLU and BatchNorm layer for normalization, nonlinear processing, and stable training. The convolution output is max-pooled to form a multi-layer semantically nested spectral local feature representation. The spectral vector is divided into sequences of length n, and the spectral vector embedding is generated through the embedding layer. The embedded representation is input into the standard transformer encoder module, and the multi-head self-attention mechanism is used to model the global dependency between different bands. The fusion is performed by splicing or weighting, and the attention fusion mechanism is used to calculate the fusion weight α to generate the final spectrogram multi-scale embedding vector: ； Among them, Z is the final fused spectrum embedding vector, α is the weight coefficient, and Z CNN is the spectral feature vector output by the CNN branch, Z Transformer is the spectral feature vector output by the converter branch, and R is the cross-channel residual correction term.

3. The AI-based material property spectral detection system according to claim 1, characterized in that: The weighted combination of different scale embedding representations is performed through the attention residual fusion module to generate a dynamically regulated spectrogram feature vector, including: The concatenated vector is input into a lightweight attention network module to generate learnable weighted parameters. The attention coefficient is used to dynamically weight the embedding representations of the two channels. The residual compensation module is introduced to generate the residual concatenation form: ； Among them, R is the residual compensation term, γ is the residual weight coefficient; The weighted fusion result is added to the residual term to obtain the final spectrogram feature vector.

4. The AI-based material property spectrum detection system according to claim 1, characterized in that: The method acquires the microstructure, chemical composition and image data of the material, and adopts a multimodal attention fusion mechanism to fuse the spectral feature vector and the structural feature into a unified feature vector for modeling performance prediction, including: For the three types of modalities, their embedding representations are extracted separately, including the spectral embedding vector Z spec , image embedding vector Z img and component embedding vector Z chem , perform weighted fusion of the three modal features to generate a unified fusion vector Z fusion , the fusion method is as follows: ； in, , , is the modal attention weight, and the fusion vector Z is input into the performance prediction model f θ , output the corresponding material performance index, the prediction expression is as follows: ； in, is the predicted performance value, f θ Model the trained performance of the network.

5. The AI-based material performance spectrum detection system according to claim 1, characterized in that: Based on the fused unified feature vector, a deep performance prediction network model is constructed. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework, including: The graph structure is input into the graph convolutional network to perform context modeling on the spectral feature vector. The expression of GCN is as follows: ； in, is the adjacency matrix with self-loops added, for The degree matrix of is the input node feature of the lth layer, is the learnable weight of the lth layer, σ is the activation function; Embed the spectrogram into Z GCN Combine or fuse features with other modalities into a new high-dimensional vector Z all , input a shared deep neural network encoder to generate shared semantic representation; each target performance indicator corresponds to a task head network, and the structure includes regression tasks and classification tasks.

6. The AI-based material property spectrum detection system according to claim 1, characterized in that: The uncertainty of the performance prediction results output by the deep performance prediction network model is estimated, and the credible interval of each performance is evaluated through the Bayesian regression network; If the model confidence falls below the set threshold, a dynamic correction mechanism is triggered, including: Construct a Bayesian regression prediction head for each performance metric and output the predicted mean and variance: ； Among them, μ ti is the predicted value of the i-th performance indicator, To account for forecast uncertainty, is a normal distribution or a Gaussian distribution; After the prediction output is Gaussian distributed, calculate the confidence interval: ； in, is the prediction interval of the material performance index of item i at the 95% confidence level; The feature vector Z of the current sample fusion Calculate the distance from the training set sample distribution: ； in, is the mean vector of the fusion features in the training set, is the straight-line distance between two eigenvectors in high-dimensional space; If D > the preset offset threshold, it is judged as a new sample or abnormal material; The sample and its actual performance indicators are stored in a cache queue. After accumulating multiple low-confidence samples, fine-tuning or incremental learning is triggered.

7. The AI-based material property spectrum detection system according to claim 1, characterized in that: The prediction performance indicators are output as numerical values and credible intervals, combined with a visualization module to display the spectrogram hotspot contribution and model attention distribution map, and a causal explanation of each prediction result is generated based on the LIME method, including: Draw a spectrogram heat map based on the weight information calculated by the attention mechanism in the fusion network; extract the attention weight distribution of the converter network in the model and generate an attention visualization image; For each prediction result, the LIME method is applied to perform causal analysis, including generating multiple perturbation samples near the current input; performing model prediction on the samples and establishing a local linear explanatory model; obtaining the positive and negative contributions of the main input features to the prediction results; and outputting a bar chart or ranked list to illustrate the role of each feature in the current prediction.

8. A material performance spectrum detection method based on AI, characterized in that: include: Acquire material spectral data and input the spectral data into a heterogeneous neural network with a dual-channel structure. One channel is a local feature extraction network based on CNN, and the other channel is a global dependency modeling network based on transformers. The network outputs multi-scale spectral embedding vectors and performs a weighted combination of the different scale embedding representations through the attention residual fusion module to generate a dynamically controlled spectral feature vector. Obtain the material's microstructure, chemical composition, and image data, and use a multimodal attention fusion mechanism to fuse the spectral feature vector and structural features into a unified feature vector for modeling performance prediction; Based on the fused unified feature vector, a deep performance prediction network model is constructed. The deep performance prediction network model introduces a graph convolution module to simulate the internal feature correlation of the spectral sequence and introduces a multi-task learning framework. Uncertainty estimation is performed on the performance prediction results output by the deep performance prediction network model, and the credible interval of each performance is evaluated through the Bayesian regression network; If the model confidence is lower than the set threshold, the dynamic correction mechanism is triggered; The prediction performance indicators are output in the form of numerical values and credible intervals, combined with the visualization module to display the contribution of the spectrum hot zone and the model attention distribution map, and generate a causal explanation for each prediction result based on the LIME method.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the AI-based material property spectral detection system according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the AI-based material property spectral detection system according to any one of claims 1 to 7 are implemented.

Images