Material synthesis process intelligent recommendation method based on small sample data

Abstract

The invention discloses a material synthesis process intelligent recommendation method based on small sample data, and relates to the field of material synthesis processes, and the method comprises the steps: obtaining small sample structure-process data, and carrying out the preprocessing to obtain a training sample; constructing a process feasibility discrimination model based on a pre-trained structure understanding enhancement model by using the training sample; utilizing the training sample to generate a model based on a pre-trained material structure, and obtaining a process generation model through instruction fine tuning training; and inputting the structural data of a to-be-synthesized material into the trained process generation model to generate candidate synthesis processes, and inputting the structural data and the candidate synthesis processes into the trained process feasibility discrimination model to obtain discrimination results of the candidate synthesis processes, and recommending the processes. According to the invention, through the pre-trained material structure understanding enhancement model and the pre-trained material structure generation model, the high-precision process discrimination model and the high-precision process generation model can be obtained based on small sample structure-process data, and intelligent recommendation of the material synthesis process is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of material synthesis processes, and more specifically, to an intelligent recommendation method for material synthesis processes based on small sample data. Background Art

[0002] Currently, the methods for determining material synthesis processes mainly include theoretical modeling and numerical simulation, empirical process trial-and-error methods, online process monitoring and feedback control, as well as high-throughput screening and data-driven optimization. These methods complement each other in material synthesis to jointly address the highly coupled problem of process parameters under the action of complex multi-variables. In recent years, with the wide application of artificial intelligence and machine learning technologies in the field of materials science, process exploration methods driven by small sample data have gradually become a hot topic.

[0003] However, the material synthesis process is complex and the process conditions are variable. Material synthesis involves complex multi-scale and multi-factor influences. For example, small changes in the types of reactants, ratios, temperature curves, etc. will significantly affect the quality and morphology of the material. In addition, since the synthesis and growth mechanisms of many materials are not fully understood, it is difficult to construct an accurate mathematical model, and the available experimental data is relatively scarce, with small amounts of laboratory data and inconsistent quality. These all pose challenges to the application of artificial intelligence and machine learning models.

[0004] Therefore, how to realize the intelligent generation and feasibility discrimination of material synthesis processes based on small sample "structure-process" labeled data, and then guide material synthesis is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent recommendation method for material synthesis processes based on small sample data, aiming to construct an association mapping between the material structure and its synthesis process, and then to generate a process recommendation scheme for laboratory synthesis only by giving the material information of the target physical properties.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The present invention discloses an intelligent recommendation method for material synthesis processes based on small sample data, and the specific steps are as follows:

[0008] Obtain small sample structure-process data and perform preprocessing to obtain training samples containing material structure data and corresponding synthesis process data;

[0009] Use the training samples to construct a process feasibility discrimination model based on a pre-trained structure understanding enhancement model;

[0010] Using the training samples, based on a pre-trained material structure generation model, a process generation model is obtained through instruction fine-tuning training;

[0011] Input the structure data of the material to be synthesized into the trained process generation model to generate candidate synthesis processes. Input the structure data of the material to be synthesized and the candidate synthesis processes into the trained process feasibility discrimination model to obtain the discrimination results of whether each candidate synthesis process can be synthesized, and use the synthesizable candidate synthesis processes as recommended processes.

[0012] Furthermore, the process feasibility discrimination model includes a structure understanding enhancement model, a process neural network, and a multi-layer perceptron; the structure understanding enhancement model extracts features based on the structure data; the process neural network extracts features based on the synthesis process data; the multi-layer perceptron fuses the features extracted by the structure understanding enhancement model and the process neural network and outputs the discrimination results.

[0013] Furthermore, the small-sample structure-process data includes: the structure data of the material and synthesis process information.

[0014] Furthermore, the preprocessing includes:

[0015] Using one-hot encoding, map the synthesis method to a k-dimensional vector; using one-hot encoding, map each element of the reactant to an n×m-dimensional vector, use numerical encoding to record the proportion information of each element to obtain a p-dimensional vector; use numerical encoding to record the reaction temperature of q segments and the corresponding time to generate a 2q-dimensional vector; use one-hot encoding to map the final operation information to an r-dimensional vector.

[0016] Furthermore, the structure understanding enhancement model consists of an encoder Encoder and a decoder Decoder; the encoder Encoder consists of 8 stacked Transformer encoders, and each layer includes a multi-head attention mechanism module, an addition and normalization module, a feed-forward neural network, and an addition and normalization module connected in sequence; the decoder Decoder consists of 4 stacked Transformer decoders, and each layer includes a multi-head attention mechanism module, an addition and normalization module, a feed-forward neural network, and an addition and normalization module.

[0017] Furthermore, the structure understanding enhancement model is pre-trained using a masked re-prediction algorithm, specifically: masking local feature points in the material feature sequence with a set probability, the encoder Encoder encodes the input material structure features to obtain a structure feature sequence, and the decoder Decoder decodes the masked atomic feature points from the encoded structure feature sequence to obtain a material structure feature vector.

[0018] Further, the process neural network includes three sequentially connected feature extraction modules and an additional linear transformation layer;

[0019] Each of the feature extraction modules includes a linear transformation layer, a normalization layer, and a non-linear activation function. The formula of the i-th feature extraction module is expressed as:

[0020] h i =σ(BN(W i h i-1 +b i ), i = 1, 2, 3;

[0021] Wherein, h i represents the feature vector extracted by the i-th feature extraction module, h0 is the input feature vector, W i and b i respectively represent the weight matrix and bias vector of the i-th layer, BN represents the batch normalization operation, and σ represents the activation function;

[0022] The additional linear transformation layer maps the feature vector extracted by the last feature extraction module to a two-dimensional space. The formula is:

[0023]

[0024] Wherein, y represents the feature vector output by the process neural network, W4 and b4 respectively represent the weight matrix and bias vector of the additional linear transformation layer, and h3 is the output feature vector of the three-layer feature extraction module.

[0025] Further, the pre-training process of the material structure generation model is specifically as follows:

[0026] Convert the material structure information into a formatted symbolic string, and construct a domain knowledge base for structure-process association; use the symbolic string to perform secondary pre-training on the general language model using the next-word prediction task to obtain a pre-trained material structure generation model.

[0027] Further, the instruction fine-tuning is: by constructing an instruction fine-tuning data set, clarify the input and output formats of the material structure generation model;

[0028] The instruction fine-tuning data set includes input instructions and output instructions. The input instructions include: chemical formula of the material, material structure information; the output instructions include: material synthesis process information.

[0029] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an intelligent recommendation method for material synthesis processes based on small-sample data. Under the condition of small-sample material structure-process data, through a pre-trained material structure understanding enhancement model and a material structure generation model, a high-precision process discrimination model and a process generation model are obtained, thereby realizing the intelligent generation and feasibility discrimination of material synthesis processes, and being able to recommend appropriate feasible processes to guide material synthesis. The present invention realizes the intelligent recommendation of material synthesis processes, without the need to manually construct a candidate process data set based on intuition and conduct a large number of experimental verifications, which helps to accelerate the synthesis progress of new materials and reduce the cost of a large number of process experimental verifications at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings.

[0031] Figure 1 It is a schematic diagram of the overall process of the embodiment of the present invention.

[0032] Figure 2 It is a schematic diagram of the collaborative chain of the process discrimination and generation models of the embodiment of the present invention.

[0033] Figure 3 It is a schematic diagram of the process neural network structure of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0035] The embodiment of the present invention discloses an intelligent recommendation method for material synthesis processes based on small-sample data, as Figure 1 and Figure 2 shown, and the specific steps are as follows:

[0036] Obtain small-sample structure-process data and perform preprocessing to obtain training samples containing material structure data and corresponding synthesis process data;

[0037] Using the training samples, based on the pre-trained structure understanding enhancement model, construct a process feasibility discrimination model ( Figure 2Chinese SynBert);

[0038] Using training samples, based on a pre-trained material structure generation model, a process generation model is obtained through instruction fine-tuning training ( Figure 2 Chinese SynGPT);

[0039] Input the structure data of the material to be synthesized into the trained process generation model to generate candidate synthesis processes. Input the structure data of the material to be synthesized and the candidate synthesis processes into the trained process feasibility discrimination model to obtain the discrimination results of whether each candidate synthesis process can be synthesized, and use the synthesizable candidate synthesis processes as recommended processes.

[0040] Specifically, the process feasibility discrimination model aims to determine whether a certain process can successfully prepare the target material. Compared with laboratory trial-and-error verification, the cost of model prediction is extremely low. Since there are both feasible and infeasible solutions among the candidate processes, accurately distinguishing between the two is crucial for guiding actual synthesis. Based on the given target material structure data and candidate synthesis process information, the feasibility of the candidate processes is determined, and referring to the key process information in the dataset, the overall framework of the model is designed, and its input includes the material structure data and the corresponding synthesis process information.

[0041] Aiming at the huge differences between the process and material structure information, a network structure design scheme based on differential input is proposed. For the material structure part, a structure understanding enhancement model is adopted. This model can deeply capture the internal characteristics of materials through pre-training on a large amount of material structure data and perform well under small sample conditions. For the candidate material synthesis processes, since they involve various process parameters (such as reaction conditions, solvent types, temperature, time, raw material ratio, etc.), their representation forms are significantly different from the material structure and it is difficult to directly fuse with the structure features output by the structure understanding enhancement model. Therefore, the present invention specifically designs a customized data preprocessing method and network architecture to process the process information and extract the key information, so as to achieve a higher-precision prediction of the feasibility of the candidate processes. Finally, the key features obtained from the structure understanding enhancement model and the process neural network are fused, and a multi-layer perceptron is used as the fusion layer to complete the final feature integration and classification prediction. The fusion strategy of the present invention makes full use of the advantages of the two types of information, realizes the collaborative learning of cross-modal features, and provides a scientific and effective basis for the feasibility judgment of the material synthesis process.

[0042] The process generation model is used to generate feasible synthetic processes. Based on small-sample "structure-process" labeled data, a process generation model based on the material structure generation model is constructed. Process data is difficult to directly model due to its complexity and format uncertainty, while large models such as GPT have natural advantages in dealing with high-dimensional complex generation tasks. The material structure generation model is a material structure generation model built based on Llama3-70B. It adopts secondary pre-training of material structure text data and is fine-tuned on this basis to focus on generating complex process data text. Using a generation large model not only reduces the difficulty of generating high-dimensional complex processes but also avoids the limitations brought by directly generating such information using regression models, thus effectively alleviating the challenges of the process generation task.

[0043] Through the collaboration of the process feasibility discrimination model and the process generation model, intelligent recommendation of material synthesis processes is realized. First, the process generation model is used to generate a large number of potentially highly probable feasible synthetic processes in batches. Then, the generated processes and the material structure information are input into the process feasibility discrimination model together to screen out the most likely successful synthesis scheme. Since the structure understanding enhancement model and the material structure generation model are not limited to a specific material system in the pre-training stage, this joint method can exhibit good generalization performance, be trained on different material structures and process data, and recommend the best synthesis process.

[0044] In a specific embodiment, the process feasibility discrimination model includes a structure understanding enhancement model, a process neural network, and a multi-layer perceptron; the structure understanding enhancement model extracts features based on structure data; the process neural network extracts features based on synthetic process data; the multi-layer perceptron fuses the features extracted by the structure understanding enhancement model and the process neural network and outputs a discrimination result.

[0045] In a specific embodiment, the small-sample structure-process data includes: structure data of the material, and synthetic process information (such as synthetic growth method, reactants, reaction temperature, final operation information, etc.).

[0046] In a specific embodiment, the preprocessing includes:

[0047] Using one-hot encoding to map the synthesis method into a k-dimensional vector; using one-hot encoding to map each element of the reactants into an n×m-dimensional vector, and using numerical encoding to record the proportion information of each element to obtain a p-dimensional vector; using numerical encoding to record the reaction temperature of q segments and the corresponding time to generate a 2q-dimensional vector; using one-hot encoding to map the final operation information into an r-dimensional vector.

[0048] Specifically, in actual process conditions, there may be significant differences in the synthesis processes of different materials. To fully demonstrate the effectiveness of the model of the present invention in process information processing and prediction, the focus is on single-crystal synthesis process data, and representative experimental records are selected for analysis. An example of the process data information of one record is shown in Table 1.

[0049] Table 1 Process Data Information

[0050]

[0051] Based on this example and other experimental data, first, the original text data is systematically sorted out, and the unstructured process descriptions are converted into structured information. During the sorting process, key process parameters are extracted, mainly including: growth method, reactants, reaction temperature, and final operation. Among them, the growth methods cover dry burning, flux method, vapor transport method, composite method, and solid-state reaction method, etc.; the reactant information records the elements participating in the reaction and their corresponding proportions; the reaction temperature part details the temperature changes and durations at each stage; the final operation includes steps such as centrifugation, crystallization at the high-temperature end, annealing, quenching, and no operation.

[0052] In order to convert the key information in the extracted process information into a data form that can be processed by a neural network, after completing the data sorting, further preprocessing of various types of process information data is carried out to convert the original descriptions into quantitative preliminary features. As shown in Table 2, for the growth method, one-hot encoding is used, and "dry burning method", "flux method", "vapor transport method", "composite method", and "solid-state reaction method" are respectively mapped to a 5-dimensional vector; in the reactant part, one-hot encoding is performed on the elements participating in the reaction (generating a vector such as 4×25 dimensions), and numerical encoding is used to record the proportion information of each element (constituting a 4-dimensional vector); the reaction temperature is recorded using numerical encoding, and each sample contains at most 7 stages, and two numerical values of temperature and time are recorded at each stage, generating a total of 14-dimensional vectors; finally, for the final operation, one-hot encoding is also used, and "centrifugation", "annealing", "quenching", and "no operation" are mapped to a 4-dimensional vector. After different preprocessing methods are adopted for each part of the process parameters, the total input vector dimension is calculated as 5+(4×25)+4+14+4 = 127.

[0053] Table 2 Process Preprocessing Methods

[0054]

[0055] One-hot encoding is a commonly used method for encoding discrete variables. Its principle is to represent each category as a binary vector, in which only one component is 1 and the rest are 0. For example, for 5 categories of growth methods, if "dry burning" is encoded, it can be represented as [1, 0, 0, 0, 0]; while "flux" corresponds to [0, 1, 0, 0, 0], and for the rest, one position in the vector is randomly assigned as 1 and the rest are 0. This encoding method can effectively avoid the interference of the order relationship between categories, enabling the neural network to independently learn the features of each category when processing discrete information.

[0056] Through the data collation and preprocessing process, the present invention not only retains the detailed information of each key process parameter in the original experimental data, but also realizes the standardization and structuring of the data. This provides high-quality input data for the subsequent process discrimination and generation model based on the neural network, thereby further improving the model's processing ability and generalization performance for complex process information.

[0057] In a specific embodiment, the structure understanding enhancement model consists of an encoder Encoder and a decoder Decoder; the encoder Encoder consists of 8 stacked Transformer encoders, and each layer includes a multi-head attention mechanism module, an addition and normalization module, a feed-forward neural network, and an addition and normalization module connected in sequence; the decoder Decoder consists of 4 stacked Transformer decoders, and each layer includes a multi-head attention mechanism module, an addition and normalization module, a feed-forward neural network, and an addition and normalization module.

[0058] In a specific embodiment, the mask re-prediction algorithm is used to pre-train the structure understanding enhancement model. Specifically: a local feature point in the material feature sequence is masked with a set probability, the encoder Encoder encodes the input material structure features to obtain a structure feature sequence, and the decoder Decoder decodes the masked atomic feature points from the encoded structure feature sequence to obtain a material structure feature vector.

[0059] Specifically, the structure understanding enhancement model is a pre-trained model that enhances the understanding of structure and can achieve high-precision prediction of small-sample material-physical properties. The model architecture mainly consists of two parts: Encoder and Decoder. The Encoder part uses an 8-layer stacked Transformer encoder, each layer containing multi-head self-attention mechanism (Multi-HeadAttention), addition and normalization (Add&Norm), feed-forward neural network (Feed Forward), and addition and normalization again. This structure can effectively capture the correlation between different atomic feature points in the material structure. The Decoder part is stacked by 8 layers of decoders similar to the Encoder structure, responsible for decoding the masked atomic feature points from the encoded structure feature sequence. Through the unsupervised pre-training process of masking-reprediction, the model can deeply understand the complex relationship between the integrity of the material structure and local features.

[0060] The unsupervised pre-training of the structure understanding enhancement model uses the masking-reprediction algorithm. In the pre-training stage, the model masks local feature points in the material feature sequence with a certain probability. The masking operation not only removes the target feature points but also zeros out the atomic features of the neighboring feature points related to them to avoid information leakage. Subsequently, the model uses the Encoder-Decoder structure to predict the masked feature points using context information. The loss function uses mean squared error (MSE), and the specific form is:

[0061]

[0062] where N represents the number of local feature point sequences in the material, M i represents the set of masked feature point indices in the th feature point sequence, P ij and are the true and predicted feature point values respectively. By minimizing this loss function, the model can effectively learn the relationship between local feature points, thereby improving the ability to understand the material structure.

[0063] In a specific embodiment, in order to fully excavate and extract integrated structured process data, after completing the preprocessing of process data, a special process neural network is designed to encode deep process features for classification and judgment. The process neural network includes three sequentially connected feature extraction modules and an additional linear transformation layer; as Figure 3 shown, each feature extraction module contains a linear transformation layer, a normalization layer, and a non-linear activation function, corresponding to Linear, BatchNorm, and ReLU in the figure respectively. The formula for the i-th layer feature extraction module is expressed as:

[0064] h i= σ(BN(W i h i-1 + b i ))), i = 1, 2, 3;

[0065] Among them, h i represents the feature vector extracted by the i-th layer feature extraction module, h0 is the input feature vector, W i and b i respectively represent the weight matrix and bias vector of the i-th layer, BN represents the batch normalization operation, and σ represents the activation function;

[0066] The additional linear transformation layer maps the feature vector extracted by the last layer feature extraction module to the two-dimensional space, and the formula is:

[0067]

[0068] Among them, y represents the feature vector output by the process neural network, W4 and b4 respectively represent the weight matrix and bias vector of the additional linear transformation layer, and h3 is the output feature vector of the three-layer feature extraction module.

[0069] The feature vector output by the process neural network retains relatively rich process detail information, and is fused with the material structure feature vector generated by the structure understanding enhancement model in the subsequent stage, so as to achieve cross-modal feature integration and more refined subsequent processing. The main advantages of this design are as follows. First, the neural network structure is customized for the input dimension to ensure that the high-dimensional process information after data preprocessing can be fully expressed. Second, by adopting repeated linear layers, normalization layers and activation layer modules, the network has strong feature abstraction and representation capabilities. Finally, the extracted 64-dimensional feature vector provides strong support for cross-modal fusion, thus improving the overall performance and generalization ability of the model in subsequent process discrimination and generation tasks.

[0070] In a specific embodiment, when training the process feasibility discrimination model, in order to evaluate the prediction performance of the model, the cross-entropy loss function is used to measure the difference between the probability distribution predicted by the model and the probability distribution of the true label. The formula is: Among them, y i and are respectively the true value and predicted value of the sample, and N is the total number of samples. In fact, only need to determine whether the process is feasible or not, so only two categories need to be analyzed.

[0071] In a specific embodiment, the pre-training process of the material structure generation model is specifically:

[0072] Convert material structure information (such as lattice parameters, lattice angles, and atomic fractional coordinates, as well as synthesis processes) into a formatted symbolic string, and construct a domain knowledge base for structure-process association; use the next-word prediction task to perform secondary pre-training on the general language model with the symbolic string to obtain a pre-trained material structure generation model. The pre-trained material structure generation model can give detailed material structure data and corresponding synthesis process information according to the input material chemical formula and composition.

[0073] In a specific embodiment, the instruction fine-tuning is as follows: by constructing an instruction fine-tuning data set, clarify the input and output formats of the material structure generation model;

[0074] The instruction fine-tuning data set includes input instructions and output instructions. The input instructions include: the chemical formula of the material, material structure information (such as the lattice constant value, angle, periodic boundary conditions, and atomic position list of the material); the output instructions include: material synthesis process information (such as the growth method, element ratio, and temperature curve in the material synthesis conditions).

[0075] Specifically, the large language model based on fine-tuning has become the core technology for content generation and discrimination tasks. By fine-tuning on the basis of the pre-trained model, the large language model based on fine-tuning can adapt to the requirements of specific tasks and show excellent performance.

[0076] The present invention realizes process discrimination and generation by constructing a process generation model through instruction fine-tuning of the material structure generation model. The material structure generation model is based on the Llama3-70B architecture and performs in-depth domain adaptation on more than 120,000 crystal structures through domain adaptive pre-training: first, convert the lattice parameters, lattice angles, and atomic fractional coordinates into a strictly formatted symbolic string, and construct a domain knowledge base for "structure-process" association; then, use the crystal structure string as the input and perform secondary pre-training on the general language model using the next-word prediction task to enable it to implicitly learn the complex mapping relationship between crystal structure features and synthesis process parameters. This domain adaptive strategy based on structured encoding effectively solves the domain adaptability defect of the natural language pre-trained model in the process generation task and lays a scalable underlying model foundation for the process reverse design under small sample conditions.

[0077] The present invention uses instruction fine-tuning to complete model fine-tuning in order to achieve process generation tasks. During the instruction fine-tuning process of the model, the instructions are task descriptions or commands expressed in natural language, aiming to guide the model to generate specific outputs. These instructions usually contain clear task objectives and related input information, and may be accompanied by examples to help the model understand the specific requirements of the task; the role of the instructions is to enhance the model's understanding and execution ability of specific tasks, enabling it to more accurately generate the expected response according to the instructions given by humans. Here, taking the constructed instructions as an example, the input instruction for the model is:

[0078] “The target material for synthesis is {formula,abc,angles,Sites}, Please provide the synthesis conditions, including: Growth method, Element ratio, and Temperature curve.”

[0079] The output instruction is: “{Growth Method, Element Ratio, Temperature Curve}”, and the complete expression is “Growth Method: Dry Burning

[0080] Element Ratio: [[['Ce', 'Pt', 'Bi'], [0.75, 0.75, 1.0]]]

[0081] Temperature Curve:

[0082] - Heating, Starting temperature: 30 °C, Ending temperature: 1100 °C, Duration: 15 hours

[0083] - Holding, Starting temperature: 1100 °C, Ending temperature: 1100 °C, Duration: 20 hours

[0084] - Cooling, Starting temperature: 1100 °C, Ending temperature: 800 °C, Duration: 100 hours;”

[0085] Instructions designed for process generation and classification tasks first require clarity and precision: the instructions should be specific and unambiguous, ensuring that every word serves the purpose of conveying task details. For example, "The target material for synthesis is Full Formula(Ce12 Bi16 Pt12)" directly clarifies the target material for synthesis and its chemical formula at the beginning; secondly is the richness of context: to help the model better understand the task background, the instructions should include all necessary background information. As shown in the case, in addition to giving the basic information of the material, it also includes its crystallographic parameters (lattice constant values, angles, periodic boundary conditions) and a list of atomic positions; the model is required to provide synthesis conditions based on the given material information, including growth methods, element ratios, and temperature curves.

[0086] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same and similar parts between the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0087] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent recommendation method for material synthesis process based on small sample data, characterized in that, The specific steps are as follows: Obtain small-sample structure-process data and perform preprocessing to obtain training samples containing material structure data and corresponding synthesis process data; Utilize the training samples to construct a process feasibility discrimination model based on a pre-trained structure understanding enhancement model; Utilize the training samples to obtain a process generation model through instruction fine-tuning training based on a pre-trained material structure generation model; Input the structure data of the material to be synthesized into the trained process generation model to generate candidate synthesis processes. Input the structure data of the material to be synthesized and the candidate synthesis processes into the trained process feasibility discrimination model to obtain discrimination results on whether each candidate synthesis process can be synthesized, and use the synthesizable candidate synthesis processes as recommended processes.

2. The intelligent recommendation method for a material synthesis process based on small-sample data according to claim 1, wherein The process feasibility discrimination model includes a structure understanding enhancement model, a process neural network, and a multi-layer perceptron; the structure understanding enhancement model performs feature extraction based on the structure data; The process neural network performs feature extraction based on the synthesis process data; The multi-layer perceptron fuses the features extracted by the structure understanding enhancement model and the process neural network and outputs a discrimination result.

3. The intelligent recommendation method for a material synthesis process based on small-sample data according to claim 1, wherein, The small-sample structure-process data includes: the structure data of the material and synthesis process information.

4. An intelligent recommendation method for a material synthesis process based on small sample data according to claim 3, characterized in that, The preprocessing includes: Using one-hot encoding to map the synthesis method to a k-dimensional vector; using one-hot encoding to map each element of the reactant to an n×m-dimensional vector, using numerical encoding to record the proportion information of each element to obtain a p-dimensional vector; using numerical encoding to record the reaction temperature of q segments and the corresponding time to generate a 2q-dimensional vector; using one-hot encoding to map the final operation information to an r-dimensional vector.

5. An intelligent recommendation method for a material synthesis process based on small-sample data according to claim 1, characterized in that, The structure understanding enhancement model consists of an encoder Encoder and a decoder Decoder; the encoder Encoder consists of 8 stacked Transformer encoders, and each layer includes a multi-head attention mechanism module, an addition and normalization module, a feed-forward neural network, and an addition and normalization module connected in sequence; the decoder Decoder consists of 4 stacked Transformer decoders, and each layer includes a multi-head attention mechanism module, an addition and normalization module, a feed-forward neural network, and an addition and normalization module.

6. The intelligent recommendation method for a material synthesis process based on small-sample data according to claim 5, wherein, The structure understanding enhancement model is pre-trained using a masked re-prediction algorithm. Specifically, local feature points in the material feature sequence are masked with a set probability. The encoder Encoder encodes the input material structure features to obtain a structure feature sequence, and the decoder Decoder decodes the masked atomic feature points from the encoded structure feature sequence to obtain a material structure feature vector.

7. An intelligent recommendation method for a material synthesis process based on small sample data according to claim 1, characterized in that, The process neural network includes three sequentially connected feature extraction modules and an additional linear transformation layer; Each feature extraction module includes a linear transformation layer, a normalization layer, and a non-linear activation function. The formula for the i-th feature extraction module is expressed as: h i = σ(BN(W i h i-1 + b i ), i = 1, 2, 3; Among them, h i represents the feature vector extracted by the i-th layer feature extraction module, h0 is the input feature vector, W i and b i respectively represent the weight matrix and bias vector of the i-th layer, BN represents the batch normalization operation, and σ represents the activation function; The additional linear transformation layer maps the feature vector extracted by the last feature extraction module to a two-dimensional space, and the formula is: where y represents the feature vector output by the process neural network, W4 and b4 respectively represent the weight matrix and bias vector of the additional linear transformation layer, and h3 is the output feature vector of the three-layer feature extraction module.

8. An intelligent recommendation method for a material synthesis process based on small-sample data according to claim 1, characterized in that, The pre-training process of the material structure generation model is specifically as follows: Convert the material structure information into a formatted symbolic string, and construct a domain knowledge base for structure-process association; use the symbolic string to perform secondary pre-training on the general language model using the next-word prediction task to obtain a pre-trained material structure generation model.

9. An intelligent recommendation method for a material synthesis process based on small-sample data according to claim 1, characterized in that, The instruction fine-tuning is as follows: by constructing an instruction fine-tuning data set, clarify the input and output formats of the material structure generation model; The instruction fine-tuning data set includes input instructions and output instructions. The input instructions include: the chemical formula of the material, material structure information; the output instructions include: material synthesis process information.

Images