A method for optimizing a hydrothermal carbon-based organic fertilizer formula

Abstract

The application provides a hydrothermal carbon-based organic fertilizer formula optimization method, and belongs to the technical field of chemical formula optimization based on computer data processing; in the application, a hydrothermal carbon-based fertilizer multidimensional correlation data set is constructed, the chain relationship between the preparation process and the structural characteristics of raw materials is quantified for the first time and is brought into an optimization system, and a component interaction feature extraction module based on a gate cycle unit and a dynamic hypergraph network is used, so that the complex synergistic effect features inside the formula can be deeply extracted and represented; and a formula reverse optimization module based on a conditional variational autoencoder and a conditional diffusion model realizes consistency verification and optimized generation of the formula under physical constraints. The application realizes deep source correlation modeling of the preparation process and the final fertilizer efficiency, accurately analyzes the complex nonlinear synergistic effect among multiple components, and improves the accuracy and efficiency of the optimization process.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of chemical formula optimization based on computer data processing, and particularly relates to a hydrothermal carbon-based organic fertilizer formula optimization method. BACKGROUND

[0002] As an important direction of agricultural sustainable development, the formula optimization of hydrothermal carbon-based organic fertilizer plays a key role in improving fertilizer efficiency and ensuring environmental safety. Traditional organic fertilizer formula design often relies on experience and simple proportioning, which is difficult to accurately balance multiple goals such as nutrient release, soil improvement and crop growth promotion, and lacks systematic analysis of the complex relationship between raw material characteristics, preparation process and fertilizer performance. Especially in the realization of efficient, safe and stable large-scale production, the existing methods often face the problems of low formula efficiency, high cost of repeated experiments, and difficulty in quantifying the safety control, which restricts the scientific application and performance improvement of hydrothermal carbon-based fertilizer in actual production.

[0003] Data-driven and intelligent control have become an important trend in fertilizer research and development. By introducing advanced data modeling and optimization algorithms, the implicit rules between hydrothermal carbon preparation process, component interaction and fertilizer output can be deeply mined, realizing the leap from empirical proportioning to precise design. The existing fertilizer formula optimization methods are as follows:

[0004] A complex fertilizer formula intelligent optimization system is disclosed in Chinese invention patent No. CN202510372620.6, which includes a central server, a data acquisition module, a data preprocessing module, a model training module, an optimization decision module and a result output module. The scheme disclosed in the invention collects crop, soil and environmental data, uses the extracted unified feature data to predict the yield of crops and the required fertilizer formula through a machine learning model, and realizes multi-objective fertilizer formula decision-making through optimization. A potato special fertilizer formula optimization method and system are provided in Chinese invention patent No. CN202411878846.5, which is applied to a cloud service system including an infrastructure layer, a middleware layer and an application layer. Through a field data acquisition system, multi-source experimental data are collected, including soil detection data, meteorological environment data, growth index data and additive combination data; a fertilizer formula database is constructed in a distributed storage system to store the component ratio, physicochemical properties and application effect of the formula; an evaluation model is constructed through a model service module to score the formula; and an intelligent optimization algorithm is run on a GPU accelerated computing node to optimize the formula.

[0005] However, the above methods and existing technologies have the following limitations:

[0006] 1) Lack of deep correlation modeling of hydrothermal carbon raw material preparation process. The existing technology mainly focuses on field crop and environmental data, without considering the key process parameters in the preparation of hydrothermal carbon and their profound influence on the physical and chemical properties of carbon materials, ignoring the decisive role of raw material properties on the source of final fertilizer efficiency;

[0007] 2) Insufficient ability to analyze the complex synergistic effect between formula components. The existing method mainly uses traditional machine learning model to process formula proportion, which is difficult to effectively capture and quantify the nonlinear, high-dimensional chemical and biological synergistic mechanism between hydrothermal carbon, various auxiliary organic matters and microorganisms, resulting in that the formula design stays at the shallow matching level;

[0008] 3) The optimization process lacks clear target orientation and reverse generation ability. The existing technology usually uses forward prediction or iterative search method for optimization, which cannot directly and efficiently infer and generate the optimal formula proportion that meets all constraint conditions according to the preset fertilizer efficiency, safety and other multiple targets, and the precision and efficiency of the optimization process need to be improved;

[0009] 4) The model architecture has limited ability to fuse and deeply extract features of multi-source heterogeneous data. The existing technology usually performs simple splicing or shallow processing on various data, lacks a special network architecture for deep fusion and deep feature mining of time series process data, static physical and chemical indicators, component proportion data and fertilizer efficiency annotation data, and limits the model's learning ability of complex correlation rules. SUMMARY

[0010] To solve the above problems, the present application provides a hydrothermal carbon-based organic fertilizer formula optimization method. In the method, a hydrothermal carbon-based fertilizer multi-dimensional correlation data set is constructed, and a component interaction feature extraction module based on a gated recurrent unit and a dynamic hypergraph network, and a formula reverse optimization module based on a conditional variational autoencoder and a conditional diffusion model are used to realize consistency verification and optimization generation of the formula under physical constraints.

[0011] The hydrothermal carbon-based organic fertilizer formula optimization method provided by the present application comprises the following steps:

[0012] S1, collecting hydrothermal reaction parameter data of hydrothermal carbon-based fertilizer preparation process, and synchronously detecting physical and chemical indicators of hydrothermal carbon finished product to form hydrothermal carbon characteristic attribute data;

[0013] Recording the mixing ratio of hydrothermal carbon finished product and different auxiliary organic matters to form organic fertilizer formula component proportion data; performing experiments on the organic fertilizer to obtain fertilizer efficiency annotation data;

[0014] S2, input the hydrothermal carbon characteristic attribute data into the hydrothermal carbon physical-chemical attribute deep evolution modeling network, capture the time sequence evolution characteristics of the hydrothermal reaction parameters by using the thermal carbon structure evolution gate cycle unit, splice the hydrothermal carbon physical and chemical indexes, and map the hydrothermal carbon structure efficiency characterization vector through the hydrothermal carbon deep perception machine; at the same time, input the organic fertilizer formula component proportion data into the formula component dynamic hypergraph network, build a dynamic hypergraph with the hydrothermal carbon structure efficiency characterization vector and other component embedding features as nodes, aggregate the features through the dynamic hypergraph convolution layer, capture the synergistic effect between the formula components, and output the formula global representation matrix;

[0015] S3, convert the fertilizer efficiency annotation data into target-oriented latent space variables through the conditional variational autoencoder; then utilize the conditional diffusion model to generate the optimal formula proportion through the residual U-Net structure-based denoising network in the reverse denoising process under the guidance of the target-oriented latent space variables and the formula global representation matrix output by S2.

[0016] Preferably, the hydrothermal reaction parameter data of the hydrothermal carbon-based fertilizer preparation process includes time sequence data of two key parameters, temperature and pressure; and the physical and chemical indexes of the hydrothermal carbon finished product include four physical and chemical index data, porosity, specific surface area, pH value and fixed carbon content.

[0017] Preferably, the fertilizer efficiency annotation data is specifically:

[0018] The leaching amounts of nitrogen, phosphorus and potassium, the three main nutrients, at consecutive time points in a simulated soil environment are periodically determined, and the N-P-K nutrient release curve is drawn based on the determination data, and the maximum release rate reflected by the curve is extracted as the nutrient release rate index; the percentage of different particle size aggregates in the soil after fertilization is determined by experiment, and the contribution value of the aggregate structure is calculated as the soil improvement evaluation; based on the plant pot experiment, the total dry matter accumulation of the aboveground and underground parts is determined by the weighing method, and the relative growth rate quantitative index is extracted by combining the growth curve method to quantify the crop growth potential; the nutrient release rate index, soil improvement evaluation and crop growth potential data are uniformly annotated as the fertilizer efficiency annotation data.

[0019] Preferably, the hydrothermal carbon physical-chemical attribute deep evolution modeling network is specifically:

[0020] The hydrothermal carbon structure evolution gated recurrent unit is used to capture the nonlinear influence of hydrothermal reaction parameters on the formation of hydrothermal carbon physicochemical indicators. It includes an input layer, three gated recurrent unit layers, and an output layer. The input layer receives the collected hydrothermal reaction parameter data and stacks three gated recurrent unit layers. Each gated recurrent unit layer includes two gating mechanisms: an update gate and a reset gate. Each layer has 64 hidden units. Each gated unit layer contains an update gate and a reset gate. The update gate controls how much of the hidden state information from the previous moment is retained in the current moment. The reset gate combines new input information with historical memory. The output sequence of the last gated recurrent unit layer is input to the fully connected layer. The fully connected layer maps the hidden states to a preset temporal evolution feature dimension, outputting the temporal evolution features of the influence of hydrothermal reaction parameters on the hydrothermal carbon physicochemical indicators. ;

[0021] Four hydrothermal carbon physicochemical indicators—porosity, specific surface area, pH value, and fixed carbon content—were used as explicit state characteristics to characterize the final results, along with temporal evolution characteristics. The data are then spliced ​​and input into a hydrothermal carbon depth sensor, which consists of four fully connected layers. Each fully connected layer uses the ReLU activation function and nonlinearly maps high-dimensional physicochemical indices and temporal evolution characteristics. Mapped to a low-dimensional latent feature space, outputting a hydrothermal carbon structure performance characterization vector. .

[0022] Preferably, the dynamic hypergraph network of the formulation components specifically comprises:

[0023] Construct a node feature matrix, where the initial features of the hydrothermal carbon nodes are the hydrothermal carbon structure performance representation vectors. The initial characteristics of other auxiliary organic matter and microbial nodes are obtained by mapping the organic fertilizer formula component ratio data through a linear embedding layer;

[0024] Construct a hyperedge feature matrix whose eigenvectors are derived from the features of the organic fertilizer formula component ratio data. Direct composition; constructing the association matrix of the hypergraph If a node belongs to a hyperedge, the corresponding element in the matrix is ​​1; otherwise, it is 0. During training, the correlation matrix is ​​updated in real time based on the changes in the proportion of organic fertilizer formula components over time or batches. ;

[0025] The formulation component dynamic hypergraph network includes dynamic hypergraph convolutional layers, which perform feature propagation and aggregation through the Laplacian operator on the hypergraph; the Laplacian operator on the hypergraph uses the node degree matrix of the diagonal matrix. Hypergraph Incidence Matrix and the hyperedge weight matrix The matrix operation is performed to obtain; wherein the diagonal elements of the superedge weight matrix are obtained by inputting the eigenvalues of the superedge eigenvalue matrix, i.e., the component proportion data of the organic fertilizer formula, into a learnable adaptive weight mapping layer, and performing normalization processing by a Softmax function and calculation;

[0026] The formula component dynamic hypergraph network aggregates the features of all nodes and superedges through the stacking of four dynamic hypergraph convolution layers, and outputs a formula global representation matrix that fuses the complex chemical-biological interaction relationship between components .

[0027] Preferably, the conditional variational autoencoder includes an encoder and a decoder, and converts the fertilizer efficiency annotation data into a latent space variable of a Gaussian distribution by using the fertilizer efficiency annotation data as a constraint condition ;

[0028] The encoder receives the fertilizer efficiency annotation data and outputs a mean vector and a variance vector ; the structure of the encoder includes four fully connected layers: the first layer receives the fertilizer efficiency annotation data as input, accesses a batch normalization layer and a ReLU activation function; the second and third fully connected layers gradually compress the feature dimension; the fourth layer is divided into two parallel fully connected layer branches, which respectively output the mean vector and the variance vector ;

[0029] Then, the target-oriented latent space variable is sampled from the mean vector and the variance vector by using the reparameterization trick; the reparameterization trick first performs logarithmic processing and division by two on the variance vector , takes the exponential operation to obtain the standard deviation term, then performs element-wise multiplication operation on the random noise vector sampled from the standard normal distribution and the standard deviation term; finally, the result is added to the mean vector to obtain the target-oriented latent space variable , which is used as the back-propagation constraint of the conditional diffusion model.

[0030] Preferably, the conditional diffusion model includes a forward noise adding process and a reverse noise removing generation process.

[0031] First, the forward noise adding process gradually adds Gaussian noise to the existing excellent formula proportion data in a preset number of time steps, until it becomes completely disordered random noise ;

[0032] ​Secondly, the reverse denoising generation process is performed, aiming to iteratively infer and denoise from the random noise under the constraint of the determined target-oriented latent space variable and the guidance of the extracted global formulation representation matrix to generate the optimal formulation proportion by a denoising network based on a residual U-Net structure ; the denoising network consists of an encoder, a bottleneck layer and a decoder, wherein the corresponding layers of the encoder and the decoder are connected by a residual connection to retain detailed information.

[0033] Preferably, the denoising network specifically comprises:

[0034] The encoder receives the noisy formulation proportion data of the current time step ; the encoder consists of four residual blocks connected in series, and each residual block contains two convolutional layers, batch normalization layers, ReLU activation functions and a residual connection; at the same time, a conditional injection module is embedded in each residual block, which is used to fuse the target-oriented latent space variable and the global formulation representation matrix to guide the denoising direction; after the output of each residual block, a down-sampling layer is connected to reduce the spatial size of the feature map and increase the receptive field, and the final output of the encoder enters the bottleneck layer;

[0035] The bottleneck layer consists of two serial residual blocks for capturing the highest level of abstract features.

[0036] The decoder receives the output of the bottleneck layer and consists of four residual blocks and an up-sampling layer, which is symmetrical in structure with the encoder; before processing each residual block, the feature map size is increased by the up-sampling layer, and then the output of the up-sampling layer is fused with the features from the corresponding layer of the encoder through a splicing operation; then, the fused features enter the residual block of the decoder for feature extraction, and the residual block structure is the same as that in the encoder.

[0037] Under the guidance of the target-oriented latent space variable and the global formulation representation matrix , the residual U-Net structure iteratively predicts and removes noise to finally obtain the reverse generated optimal formulation proportion .

[0038] Preferably, the excellent formula proportion data refers to formulas whose comprehensive fertilizer efficiency scores calculated by normalizing and weighted summing the fertilizer efficiency annotation data rank in the top 10%, and the specific calculation method of the comprehensive fertilizer efficiency score is: first, the nutrient release rate index, soil improvement evaluation and crop growth potential three index data are normalized by maximum and minimum value to eliminate the dimensional difference, and then the weighted sum method is used to calculate the comprehensive fertilizer efficiency score of each formula.

[0039] Preferably, in the conditional diffusion model training process of S3, the Lagrange multiplier method is introduced as a constraint term of the loss function to force the generated optimal formula proportion to meet the physical constraints, mainly including: the total sum of formula proportion is 100%, and the heavy metal content index must be within the safety threshold;

[0040] In order to obtain the heavy metal content prediction value, a safety evaluation regression model is constructed, which adopts a three-layer fully connected network structure: the first layer receives the spliced water, heat and carbon characteristic attribute data and organic fertilizer formula component proportion data, and performs initial feature fusion through the full connection layer, batch normalization layer and ReLU activation function; the second layer further extracts nonlinear features through the full connection layer, batch normalization layer and ReLU activation function; the third layer outputs the heavy metal content prediction scalar through a linearly activated full connection layer, and the model takes the measured heavy metal content in the safety threshold annotation as the training label, so that it can predict the heavy metal content prediction value corresponding to any input formula .

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] (1) Deep source correlation modeling of preparation process and final fertilizer efficiency is realized: the prior art usually regards the fertilizer raw material as a static attribute, ignoring the key influence of the preparation of water, heat and carbon raw materials on its structure performance. The present application collects the water, heat and carbon reaction temperature and pressure time series process data, and uses the gating cycle unit to capture the dynamic evolution law of its physicochemical properties such as water, heat and carbon porosity and specific surface area, so as to first deeply quantify the "preparation process-structure characteristic" chain relationship of the raw material and introduce it into the optimization system, solving the problem of unstable formula design basis and inaccurate prediction caused by ignoring the source characteristics of the raw material in the traditional method;

[0043] (2) Accurate analysis of the complex nonlinear synergistic effect among multiple components: The existing formula optimization method mainly uses linear or shallow model to deal with the proportion, which is difficult to describe the complex high-dimensional chemical and biological synergistic effect among water, heat, carbon, multiple organic matters and microorganisms. The invention innovatively introduces a dynamic hypergraph network, models each component and its mixed relationship as a dynamically evolving hypergraph structure, and explicitly aggregates the interaction information among multiple components through a hypergraph convolution layer, so as to deeply extract and represent the complex synergistic effect characteristics in the formula, overcoming the limitations of the existing technology in analyzing the formula mechanism;

[0044] (3) Provide target-oriented formula reverse intelligent generation capability: The existing optimization method mainly depends on forward prediction and iterative trial and error, and cannot directly generate the optimal formula according to the expected fertilizer efficiency, safety and other multiple targets. The invention fuses the conditional variational autoencoder and the conditional diffusion model, first constructs a conditional latent space constrained by the target fertilizer efficiency, and then in the reverse denoising process of the diffusion model, the target latent variable and the global representation of the formula are guided at the same time, realizing end-to-end reverse generation, which significantly improves the accuracy and efficiency of the optimization process, and solves the pain points of the traditional optimization path of detour and long time consumption;

[0045] (4) Ensure the physical feasibility and environmental safety of the optimized formula: The existing scheme often takes the total formula, heavy metal safety and other hard constraints as post-check items, which is easy to produce a theoretically effective but unproducible or risky formula. The invention introduces a physical constraint loss based on the Lagrange multiplier method in the core stage of model training, and takes a lightweight fertilizer efficiency and safety prediction model as a virtual verification link, forcing the generated scheme to meet the actual production and environmental protection requirements such as "proportion sum is 100%" and "heavy metal does not exceed the standard", realizing the consistency of the optimization process and the physical reality, and fundamentally guaranteeing the preparability and environmental friendliness of the generated formula. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The figure is the overall technical route flowchart of the invention.

[0047] Figure 2 The figure is the structure diagram of the component interaction feature extraction module.

[0048] Figure 3 The figure is the structure diagram of the formula reverse optimization module based on the conditional variational autoencoder and the conditional diffusion model.

[0049] Figure 4 The figure is the performance analysis diagram of the component interaction feature extraction module based on the gated recurrent unit and the dynamic hypergraph network in the embodiment.

[0050] Figure 5 The figure is the pure feature spectrum of the formula global representation matrix in the embodiment.

[0051] Figure 6 Performance analysis chart for reverse optimization of formulation and correction of physical constraints in the examples. DETAILED DESCRIPTION

[0052] A hydrothermal carbon-based organic fertilizer formula optimization method is provided, and the overall technical route flowchart is as shown in Figure 1 The specific implementation process of the present application will be further described below in conjunction with specific examples.

[0053] S1, hydrothermal carbon-based fertilizer multi-dimensional correlation dataset construction

[0054] First, for the preparation process of hydrothermal carbon-based fertilizer, the hydrothermal reaction temperature and pressure time series data are collected through the industrial control system, and the porosity, specific surface area, pH value, and fixed carbon content of the hydrothermal carbon product are simultaneously detected, which together constitute the hydrothermal carbon characteristic attribute data. Secondly, the mixing ratio of hydrothermal carbon and different auxiliary organic matters is recorded to constitute the organic fertilizer formula component ratio data. Then, indoor and pot experiments are carried out on the organic fertilizer product to obtain the nutrient release rate index, soil improvement evaluation, and crop growth potential, which are uniformly labeled as fertilizer efficiency labeling data. Simultaneously, heavy metal content detection is carried out and safety threshold labeling is set according to safety standards. Finally, the hydrothermal carbon characteristic attribute data, organic fertilizer formula component ratio data, and fertilizer efficiency and safety threshold labeling are integrated to construct the hydrothermal carbon-based fertilizer multi-dimensional correlation dataset.

[0055] S1-1 Multi-source data acquisition and characteristic attribute definition:

[0056] First, for the preparation process of hydrothermal carbon, the time series data of temperature and pressure, two key parameters of hydrothermal reaction, are collected through the industrial control system. Simultaneously, the generated hydrothermal carbon product is subjected to physicochemical index detection to obtain four physicochemical index data of porosity, specific surface area, pH value, and fixed carbon content. The temperature and pressure time series data of hydrothermal reaction and the four physicochemical index data of porosity, specific surface area, pH value, and fixed carbon content are combined to form the hydrothermal carbon characteristic attribute data. Secondly, the mixing ratio data of hydrothermal carbon and different auxiliary organic matters during the preparation of organic fertilizer is recorded, and the ratio data is used to form the organic fertilizer formula component ratio data. During data collection, all data use a unified sampling frequency and clock synchronization mechanism to ensure accurate alignment of the hydrothermal carbon characteristic attribute data and the organic fertilizer formula component ratio data in the time dimension.

[0057] S1-2 Fertilizer efficiency and safety evaluation labeling processing:

[0058] For the prepared organic fertilizer product, indoor and plant pot experiments are carried out. First, in terms of fertilizer effect evaluation, data are obtained through indoor and plant pot experiments: in the aspect of fertilizer effect evaluation, the continuous monitoring method is used for physical and chemical detection, and the leaching amount of nitrogen (N), phosphorus (P) and potassium (K), the three main nutrients, at continuous time points in the simulated soil environment is periodically determined, and based on the determination data, the N-P-K nutrient release curve is drawn, and the maximum release rate reflected by the curve is extracted as the nutrient release rate index; the percentage of different particle size aggregates in the soil after fertilization is determined by experiment, and the contribution value of the aggregate structure is calculated as the soil improvement evaluation; the plant pot experiment adopts the weighing method to determine the total dry matter accumulation of the aboveground and underground parts, and combines with the growth curve method to extract the relative growth rate quantitative index to quantify the crop growth potential. The nutrient release rate index, soil improvement evaluation and crop growth potential data are uniformly marked as fertilizer effect marking data. At the same time, the heavy metal content of the organic fertilizer product is detected, and according to the limit value requirement in the national organic fertilizer safety standard, the maximum allowable content of key heavy metals such as cadmium, lead, arsenic, chromium and mercury is counted, and whether the measured heavy metal content of each formula exceeds the limit is judged, and a safety threshold value is set. The marking uses a binary label, when the content of any key heavy metal exceeds the limit, the formula is marked as unsafe; otherwise, it is marked as safe. Finally, the fertilizer effect marking data and the safety threshold marking are integrated to serve as the fertilizer effect and safety evaluation marking, at the same time, part of the high-quality samples are reserved as the excellent formula proportion data. These excellent formula proportion data refer to the top 10% of formulas in the comprehensive fertilizer effect score after normalization processing and weighted summation calculation, and the specific calculation method of the comprehensive fertilizer effect score is as follows: first, the nutrient release rate index, soil improvement evaluation and crop growth potential three index data are normalized to eliminate the dimension difference, then the weighted summation method is used to calculate the comprehensive fertilizer effect score of each formula, and finally the score is sorted from high to low for screening.

[0059] S1-3 Data set construction and data closed loop:

[0060] The collected hydrothermal carbon characteristic attribute data, formula component proportion data, and fertilizer effect and safety evaluation marking are integrated to construct a multi-dimensional correlation data set of hydrothermal carbon-based fertilizer. Among them, the hydrothermal carbon characteristic attribute data and the formula component proportion data will be used as the input of the subsequent S2 module feature extraction, which together represents the preparation process and component composition of the hydrothermal carbon-based organic fertilizer. The fertilizer effect and safety evaluation marking will be used as the guide value of the optimization target and the benchmark for loss function calculation in the subsequent S3 module formula optimization.

[0061] S2, component interaction feature extraction module construction based on gated recurrent unit and dynamic hypergraph network

[0062] A component interaction feature extraction module is constructed, as shown in Figure 2 The component interaction feature extraction module includes a hydrothermal carbon physical-chemical attribute deep evolution modeling network and a formula component dynamic hypergraph network. The hydrothermal carbon feature attribute data obtained in S1 is input into the hydrothermal carbon physical-chemical attribute deep evolution modeling network. The hydrothermal reaction parameter time sequence evolution feature is captured by using the hydrothermal carbon structure evolution gated recurrent unit model, and is spliced with the hydrothermal carbon physical and chemical indexes. The hydrothermal carbon structure efficiency representation vector is obtained by mapping through the deep perception machine. At the same time, the formula component proportion data obtained in S1 is input into the formula component dynamic hypergraph network. The dynamic hypergraph is constructed with the hydrothermal carbon structure efficiency representation vector and other component embedded features as nodes. The dynamic hypergraph convolution layer is used to aggregate the features, capture the synergistic effect between the formula components, and output the formula global representation matrix which fuses the complex interaction relationship between the components.

[0063] S2-1 constructs a hydrothermal carbon physical-chemical attribute deep evolution modeling network to obtain a hydrothermal carbon structure efficiency representation vector. The network includes a hydrothermal carbon structure evolution gated recurrent unit model and a hydrothermal carbon deep perception machine model, which are as follows:

[0064] For the hydrothermal carbon feature attribute data collected in S1, first, a hydrothermal carbon structure evolution gated recurrent unit model is constructed to capture the nonlinear influence of the hydrothermal reaction parameters on the formation of the hydrothermal carbon physical and chemical indexes. The network structure of the hydrothermal carbon structure evolution gated recurrent unit model includes an input layer, three layers of gated recurrent unit layers, and an output layer. The input layer receives the time sequence data of the two key parameters of the hydrothermal reaction collected in S1. Three layers of gated recurrent unit layers are stacked. Each gated recurrent unit layer includes two gating mechanisms, an update gate and a reset gate. Each layer is provided with 64 hidden units to ensure sufficient capacity to capture the time sequence dependence. Each gated unit layer includes an update gate and a reset gate. The update gate is responsible for controlling how much of the hidden state information at the previous time is retained to the current time to capture long-term dependencies. The reset gate is responsible for determining how to combine new input information with historical memory to adapt to time sequence changes. Through the three-layer stacking structure, the model can learn the deep and complex dynamic evolution features in the time sequence data of the hydrothermal reaction parameters. The output sequence of the last layer of gated recurrent unit layers is input into a fully connected layer, which is used to map the hidden state to a preset time sequence evolution feature dimension to output the time sequence evolution feature of the influence of the hydrothermal reaction parameters on the hydrothermal carbon physical and chemical indexes . Secondly, the porosity, specific surface area, pH value and fixed carbon content in the hydrothermal carbon feature attribute data in S1 are used as the explicit state features representing the final results, and are combined with the time sequence evolution features The data is stitched and input into a hydrothermal carbon deep perception machine (Multilayer Perceptron, MLP) model. The hydrothermal carbon deep perception machine model is composed of four fully connected layers, each of which uses a ReLU activation function to map high-dimensional physicochemical indicators and time-series evolution characteristics to a low-dimensional latent feature space through non-linear mapping, achieving feature abstraction and dimensionality reduction. This process maps the physicochemical indicators and time-series evolution characteristics of the hydrothermal carbon feature attribute data as input, and the hydrothermal carbon deep perception machine model finally outputs a low-dimensional hydrothermal carbon structural efficiency representation vector . The hydrothermal carbon structural efficiency representation vector characterizes the structural efficiency potential of the hydrothermal carbon in organic fertilizer.

[0065] S2-2 Component synergy effect formula component dynamic hypergraph network construction:

[0066] For the formula component proportion data in S1, a dynamic hypergraph convolutional network (Dynamic Hypergraph Convolutional Networks, DHGCN) is innovatively constructed to capture the non-linear synergistic effect between multiple components. The formula component dynamic hypergraph network includes node feature matrix construction, hyperedge feature matrix construction, correlation matrix construction, and dynamic hypergraph convolutional layer.

[0067] First, the node feature matrix is constructed, where the initial features of the hydrothermal carbon nodes are the hydrothermal carbon structural efficiency representation vectors output by S2-1, and the initial features of other auxiliary organic matter and microbial nodes are obtained by linear embedding layer mapping from the formula component proportion data in S1.

[0068] Secondly, the hyperedge feature matrix is constructed, and its feature vector is directly composed of the features of the formula component proportion data. The features of the formula component proportion data are the mixing proportions of hydrothermal carbon and all auxiliary organic matter recorded in S1, i.e., the original numerical vector of the formula component proportion data.

[0069] Then, the correlation matrix of the hypergraph is constructed. If the node belongs to the hyperedge, the corresponding element of the matrix is 1, otherwise it is 0. During training, the correlation matrix is updated in real time according to the changes of the formula component proportion data over time or batch , making it dynamic. The core of the formula component dynamic hypergraph network is the dynamic hypergraph convolutional layer, which propagates and aggregates features through the Laplacian operator on the hypergraph. The Laplacian operator on the hypergraph is composed of the node degree matrix of the diagonal matrix , the correlation matrix of the hypergraph , and the hyperedge weight matrix is obtained by matrix operation. The diagonal elements of the super-edge weight matrix are obtained by the eigenvalues of the super-edge feature matrix, i.e., the feature of the proportion of formula components is input into a learnable adaptive weight mapping layer and is normalized by a Softmax function to obtain the formula global representation matrix , which automatically learns the importance of super-edges using the component composition features of the formula itself, avoiding the dependence on the S1 fertilizer efficiency annotation data in the feature extraction stage and ensuring the model's reasoning ability for unknown formulas.

[0070] Finally, the formula component dynamic hypergraph network aggregates the features of all nodes and super-edges through the stacking of four dynamic hypergraph convolution layers, and outputs a formula global representation matrix that integrates the complex chemical-biological interaction between components. The formula global representation matrix contains deep features of component synergies.

[0071] S3, Construction of Formula Reverse Optimization Module Based on Conditional Variational Autoencoder and Conditional Diffusion Model

[0072] The formula reverse optimization module based on conditional variational autoencoder and conditional diffusion model, as shown in Figure 3 , inputs the fertilizer efficiency annotation data of S1 into the formula reverse optimization module based on conditional variational autoencoder and conditional diffusion model. First, the conditional variational autoencoder is used to convert the fertilizer efficiency annotation data into target-oriented latent space variables. Then, the conditional diffusion model is used to guide the target-oriented latent space variables and the formula global representation matrix output by S2, and the denoising network based on the residual U-Net structure is used to generate the optimal formula proportion through the reverse denoising process.

[0073] S3-1 Construction of Target-Oriented Conditional Latent Space:

[0074] To achieve target-oriented formula reverse optimization, the fertilizer efficiency annotation data in S1 is first used as a constraint condition, and the conditional variational autoencoder (CVAE) model is used to convert the fertilizer efficiency annotation data into Gaussian distributed latent space variables . The conditional variational autoencoder model consists of an encoder and a decoder.

[0075] The encoder receives the fertilizer efficiency annotation data and outputs the mean vector and the variance vector .

[0076] The structure of the encoder contains four fully connected layers. 1) The first layer receives the fertilizer efficiency labeled data as input, and accesses the batch normalization layer and the ReLU activation function. 2) The second and third layers of fully connected layers gradually compress the feature dimension. 3) The fourth layer is divided into two parallel branches of fully connected layers, which respectively output the mean vector and the variance vector .

[0077] Then, by the reparameterization trick, the target-oriented latent space variable is sampled from the mean vector and the variance vector . Specifically, the reparameterization trick first performs logarithmic processing and division by two on the variance vector , takes the standard deviation item by exponential operation, then performs element-wise multiplication operation on the random noise vector sampled from the standard normal distribution and the standard deviation item; finally, the result is added to the mean vector to obtain the target-oriented latent space variable . The target-oriented latent space variable represents the Gaussian distribution of the target fertilizer efficiency in the latent space, which will be used as the back-propagation constraint of the conditional diffusion model.

[0078] The decoder receives the target-oriented latent space variable and attempts to reconstruct the original fertilizer efficiency labeled data. The structure of the decoder contains four fully connected layers. 1) The first layer maps the target-oriented latent space variable back to a higher dimension. 2) The three layers of fully connected layers (second, third, and fourth layers) gradually restore the feature dimension, and access the batch normalization layer and the ReLU activation function after each layer. 3) The output layer outputs the reconstructed fertilizer efficiency labeled data by the linear activation function.

[0079] S3-2 Recipe inverse generation of the conditional diffusion model based on the residual U-Net structure:

[0080] The conditional diffusion model (CDM) is innovatively introduced for the inverse generation of the recipe proportion. The conditional diffusion model contains two parts: the forward noise adding process and the reverse denoising generation process.

[0081] First, the forward noise adding process is performed on the existing excellent recipe proportion data in S1, and Gaussian noise is gradually added within a preset number of time steps, until it becomes completely disordered random noise .

[0082] Second, the reverse denoising generation process is performed, which aims to generate the target recipe proportion from the random noise In the context of the target-oriented latent space variables determined in S3-1 Under the constraints, and the global representation matrix of the formulation extracted by S2-2 Under the guidance of [the relevant authority], iterative reverse reasoning and noise reduction are performed.

[0083] Finally, the optimal formula ratio is generated. This process is implemented using a denoising network based on a residual U-Net structure. The denoising network does not include an attention mechanism and consists of three parts: an encoder, a bottleneck layer, and a decoder. Residual connections are used between corresponding layers in the encoder and decoder to preserve detailed information.

[0084] The specific processing flow of the denoising network is as follows:

[0085] First, the encoder receives the noise addition formula ratio data for the current time step. The encoder consists of four residual blocks connected in series. Each residual block contains two convolutional layers, a batch normalization layer, a ReLU activation function, and a residual connection. Additionally, a conditional injection module is embedded in each residual block to fuse target-oriented latent space variables. and global representation matrix of the formula This guides the direction of denoising. After the output of each residual block, a downsampling layer is applied to reduce the spatial size of the feature map and increase the receptive field. The final output of the encoder enters the bottleneck layer.

[0086] Secondly, the bottleneck layer consists of two cascaded residual blocks, which do not undergo sampling operations and are specifically used to capture the highest-level abstract features. These highest-level abstract features are the global topological structure inherent in the formulation data regarding the interrelationships between the proportions of each component, as well as the essential latent distribution patterns highly correlated with fertilizer efficiency targets. Then, the decoder receives the output of the bottleneck layer and consists of four residual blocks and an upsampling layer, structurally symmetrical to the encoder. Before processing each residual block, the feature map size is increased through the upsampling layer. Subsequently, the output of the upsampling layer is fused with the features from the corresponding layer of the encoder through a concatenation operation.

[0087] Then, the fused features are fed into the residual block of the decoder for feature extraction. The residual block structure is the same as that in the encoder and also embeds a conditional injection module.

[0088] Finally, the residual U-Net structure is used to define the target-oriented latent space variables. and global representation matrix of the formula Guided by the algorithm, noise is iteratively predicted and removed. Specifically, the noisy recipe data at the current time step is input into the denoising network to predict the noise component it contains. Then, the predicted noise component is subtracted from the current data and a random perturbation term is added, thereby gradually calculating the distribution of the denoised recipe data from the previous time step. Finally, the decoder outputs the optimal recipe ratio. .

[0089] S4. Formulation Consistency Verification and Correction under Physical Constraints and Platform Integration

[0090] Formula consistency verification and correction under physical constraints and platform integration: Formula consistency verification and correction under physical constraints are carried out by first training a lightweight fertilizer efficiency simulator using the S1 dataset to virtually verify the generated optimal formula ratio and predict its expected fertilizer efficiency; secondly, the Lagrange multiplier method is introduced into the training of the conditional diffusion model to construct a physical constraint loss function, which constrains the sum of the formula ratios and the safety threshold of heavy metal content, and the heavy metal content is predicted through a safety assessment regression model. The constraint term is added to the total loss to fine-tune the denoising network; finally, the trained dynamic hypergraph network of formula components and the conditional diffusion model are deployed to the hydrothermal carbon-based fertilizer formulation platform, integrating high-performance computing and human-computer interaction interface.

[0091] S4-1 Formula Virtual Verification and Fertilizer Efficacy Simulation:

[0092] To ensure the optimal formula ratio generated by S3 To assess the feasibility of practical preparation, a physical consistency check module was constructed. First, a lightweight fertilizer efficiency simulator (LES) was trained using hydrothermal char feature attribute data, formulation component ratio data, and fertilizer efficiency annotation data from the multidimensional association dataset of hydrothermal char-based fertilizers collected by S1. This lightweight fertilizer efficiency simulator employs a three-layer fully connected network structure, taking the hydrothermal char feature attribute data and formulation component ratio data from S1 as input and the fertilizer efficiency annotation data as output. The optimal formulation ratio output by S3 was then analyzed. Virtual validation was conducted, using a lightweight fertilizer efficiency simulator to predict the expected fertilizer effects of the formula in terms of nutrient release rate, soil improvement assessment, and crop growth promotion potential. .

[0093] S4-2 Loss Function Correction for Physical Constraints:

[0094] During the training of the conditional diffusion model in S3, the Lagrange multiplier method is introduced as a constraint term in the loss function to force the generation of the optimal recipe ratio. Physical constraints must be met. The main constraints include: the total proportion of ingredients must be 100%, and the heavy metal content must be within safe thresholds.

[0095] To obtain the heavy metal content prediction value, a safety evaluation regression model (SERM) is constructed. The safety evaluation regression model adopts a three-layer fully connected network structure: the first layer receives the spliced hydrothermal carbon characteristic attribute data and the formula component proportion data, and performs initial feature fusion through the fully connected layer, batch normalization layer and ReLU activation function; the second layer further extracts nonlinear features through the fully connected layer, batch normalization layer and ReLU activation function; and the third layer outputs the heavy metal content prediction scalar through a linearly activated fully connected layer. The model takes the measured heavy metal content in the safety threshold labeling of S1 as the training label, and through training, it can predict the heavy metal content prediction value corresponding to any input formula .

[0096] The calculation of the constraint term of the loss function includes two parts: one is the square difference between the sum of the optimal formula proportion and 1 multiplied by the Lagrange multiplier ; and the other is the difference between the heavy metal content prediction value corresponding to the optimal formula proportion and the limit value in the safety threshold labeling set in S1, and the maximum value is compared with zero, and then multiplied by the Lagrange multiplier . The constraint term is added to the denoising loss of the conditional diffusion model to obtain the modified total loss. If the optimal formula proportion generated by S3 does not satisfy the physical constraint , the constraint term is used as a feedback signal to fine-tune the weights in the denoising network of the residual U-Net structure of S3 through backpropagation, realizing consistency verification and correction of the formula, and ensuring that the generated formula has actual preparability.

[0097] S4-3 formula platform integration:

[0098] The formula component dynamic hypergraph network trained by S2 and the conditional diffusion model trained by S3 are deployed and integrated into the hydrothermal carbon-based fertilizer formula platform. The platform has high-performance computing capability and can support real-time inference of complex deep learning models. At the same time, a human-computer interaction interface is established to receive user input parameters and display formula recommendations of the system.

[0099] Experimental verification and analysis:

[0100] In order to verify the effectiveness of the hydrothermal carbon-based fertilizer formula optimization method proposed in the present application, the present experiment takes the hydrothermal carbon-based fertilizer multi-dimensional correlation dataset as the research object, and focuses on evaluating the core performance of the method in hydrothermal carbon feature extraction and formula reverse optimization and physical constraint correction. The experimental indicators include feature reconstruction error (characterizing feature extraction effectiveness), fertilizer efficiency prediction bias (characterizing formula optimization accuracy) and heavy metal content compliance rate (characterizing physical constraint effectiveness).

[0101] 1. Performance analysis of hydrothermal carbon feature extraction module

[0102] The feature reconstruction error of the traditional single GRU and the ordinary graph convolution network (GCN) is compared to verify the capturing ability of the method to the hydrothermal carbon time evolution feature and the component collaborative interaction feature. The feature reconstruction error calculation method is: the mean square error (MSE) of the extracted feature and the true physico-chemical feature, and the lower the error, the better the feature extraction effect.

[0103] As shown in Figure 4 The feature reconstruction error of the method proposed in the application in the whole training cycle is significantly lower than that of the single GRU and the ordinary GCN method:

[0104] In the early training stage (round 1-10), the error of the method decreases fastest, which reflects the efficient capturing ability of the three-layer GRU to the hydrothermal carbon time evolution feature.

[0105] In the late training stage (round 30-50), the error of the method is stable at about 0.028, which is much lower than that of the single GRU (0.085) and the ordinary GCN (0.045), which verifies the effective aggregation of the dynamic hypergraph network to the multi-component nonlinear collaborative interaction feature, and the low-dimensional abstract ability of the deep perception machine to the physico-chemical-time sequence feature.

[0106] The results show that the component interaction feature extraction module based on the gated recurrent unit and the dynamic hypergraph network can completely capture the full-dimensional features of the hydrothermal carbon, and provide high-quality feature input for subsequent formula optimization.

[0107] Figure 5 The pure feature spectrum of the formula global representation matrix finally output by the component interaction feature extraction module based on the gated recurrent unit and the dynamic hypergraph network is visualized. The light and shade distribution of the color in the figure represents the high-dimensional feature activation strength extracted by the deep neural network: the dark area represents the low response background information, and the bright orange-yellow light spots and strips reveal the high-intensity nonlinear synergistic effect formed between the structure performance of the hydrothermal carbon and the auxiliary organic component. This complex texture structure is not random noise, but a dense latent feature space manifold formed by the fusion of discrete formula proportion data and time evolution physico-chemical parameters after the model is processed by dynamic hypergraph convolution aggregation. It accurately quantifies the interaction mode of different formula combinations at the micro level.

[0108] 2. Performance analysis of formula reverse optimization and physical constraint correction

[0109] The formula reverse optimization method proposed in the application compares the fertilizer efficiency prediction deviation (relative error of predicted fertilizer efficiency and actual fertilizer efficiency) in two scenarios of "with physical constraints (Lagrange multiplier method to constrain the sum of formula proportions and heavy metal safety threshold)" and without physical constraints, and simultaneously counts the heavy metal content compliance rate (the proportion of formulas with predicted values below the safety threshold), to verify the improvement effect of physical constraints on formula feasibility.

[0110] As shown in Figure 6 The introduction of physical constraints significantly improves the performance of formula reverse optimization:

[0111] Fertilizer efficiency prediction deviation: the deviation of the scheme with physical constraints is always lower than that of the unconstrained scheme, and stabilizes in the range of 0.005-0.02 as the number of formula tests increases, and the deviation of the unconstrained scheme can only reach 0.03 at the lowest, verifying the improvement effect of the lightweight fertilizer simulator virtual verification and loss function constraint on formula precision;

[0112] Heavy metal compliance rate: the compliance rate of the scheme with physical constraints reaches 100% after the 13th formula, and remains above 90% throughout, proving that the safety evaluation regression model and the Lagrange multiplier method can effectively constrain the heavy metal content within the safety threshold.

[0113] The above only describes the preferred embodiments of the application and is not intended to limit the application. For those skilled in the art, the application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included within the protection scope of the application.

[0114] Although the specific embodiments of the application have been described above, they are not intended to limit the scope of protection of the application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the application without creative labor are still within the protection scope of the application.

Claims

1. A method for optimizing the formulation of a hydrothermal carbon-based organic fertilizer, characterized in that, Comprise the following steps: S1, collect hydrothermal reaction parameter data of hydrothermal carbon-based fertilizer preparation process, and synchronously detect physical and chemical indexes of hydrothermal carbon product to form hydrothermal carbon characteristic attribute data; the hydrothermal reaction parameter data of the hydrothermal carbon-based fertilizer preparation process comprises time series data of two key parameters of temperature and pressure; the physical and chemical indexes of the hydrothermal carbon product include four kinds of physical and chemical index data of porosity, specific surface area, pH value and fixed carbon content; Record the mixing ratio of hydrothermal carbon product and different auxiliary organic matters to form organic fertilizer formula component ratio data; carry out experiments on organic fertilizer to obtain fertilizer efficiency annotation data, including nutrient release rate index, soil improvement evaluation and crop growth potential data; S2, input the hydrothermal reaction parameter data into the hydrothermal carbon physical-chemical attribute deep evolution modeling network, capture the time series evolution characteristics of the hydrothermal reaction parameter by using the thermal carbon structure evolution gate cycle unit, and splice it with the physical and chemical indexes of the hydrothermal carbon, and obtain the thermal carbon structure efficiency characterization vector through the hydrothermal carbon deep perception machine; at the same time, input the organic fertilizer formula component ratio data into the formula component dynamic hypergraph network, build a dynamic hypergraph with the hydrothermal carbon structure efficiency characterization vector and other component embedding features as nodes, the other components include other auxiliary organic matters and microorganisms, aggregate features through the dynamic hypergraph convolution layer, capture the synergistic effect between formula components, and output the formula global representation matrix; S3, convert the fertilizer efficiency annotation data into target-oriented latent space variables through the conditional variational autoencoder; then use the conditional diffusion model to generate the optimal formula ratio through the residual U-Net structure-based denoising network in the reverse denoising generation process under the guidance of the target-oriented latent space variables and the formula global representation matrix output by S2.

2. The method for optimizing the formulation of hydrochar-based organic fertilizer according to claim 1, characterized in that: The fertilizer efficiency annotation data is specifically: Periodically measure the leaching amount of nitrogen, phosphorus and potassium, the three main nutrients, at consecutive time points in a simulated soil environment, draw the N-P-K nutrient release curve based on the measured data, and extract the maximum release rate reflected by the curve as the nutrient release rate index; measure the percentage of different particle size aggregates in the soil after fertilization through experiments, calculate the aggregate structure contribution value as the soil improvement evaluation; based on plant pot experiments, measure the total dry matter accumulation of aboveground and underground parts by weighing method, and combine with the growth curve method to extract the relative growth rate quantitative index to quantify the crop growth potential; the nutrient release rate index, soil improvement evaluation and crop growth potential data are uniformly annotated as fertilizer efficiency annotation data.

3. The method for optimizing the formulation of hydrochar-based organic fertilizer according to claim 1, characterized in that: The hydrothermal carbon physical-chemical attribute deep evolution modeling network is specifically: The hydrothermal carbon structure evolution gating cycle unit is used for capturing the nonlinear influence of the hydrothermal reaction parameters on the formation of the hydrothermal carbon physical and chemical indexes, and comprises an input layer, three layers of gating cycle unit layers and an output layer; the input layer receives collected hydrothermal reaction parameter data, and the three layers of gating cycle unit layers are stacked; each layer of the gating cycle unit layers comprises two gating mechanisms of an update gate and a reset gate; 64 hidden units are arranged in each layer; each gating cycle unit layer comprises the update gate and the reset gate; the update gate is responsible for controlling how much hidden state information at the previous moment is retained to the current moment; the reset gate is used for combining new input information with historical memory; the output sequence of the last layer of gating cycle unit layers is input to a fully connected layer; the fully connected layer is used for mapping the hidden state to a preset time sequence evolution feature dimension to output the time sequence evolution features of the influence of the hydrothermal reaction parameters on the hydrothermal carbon physical and chemical indexes ; The porosity, specific surface area, pH value, and fixed carbon content of the hydrothermal carbon are taken as the dominant state characteristics of the final results, and the time evolution characteristics are spliced and input into a hydrothermal carbon deep perception machine, which is composed of four fully connected layers, each of which uses a ReLU activation function to map high-dimensional physicochemical indicators and time evolution characteristics to a low-dimensional latent feature space, outputting a hydrothermal carbon structure performance characterization vector .

4. The hydrochar-based organic fertilizer formulation optimization method of claim 3, wherein: The formula component dynamic hypergraph network is specifically: A node feature matrix is constructed, wherein the initial features of the hydrothermal carbon nodes are the hydrothermal carbon structure performance characterization vectors and the initial features of the other auxiliary organic matter and microbial nodes are obtained by linear embedding layer mapping from the organic fertilizer formula component proportion data; Constructing the hyperedge feature matrix, whose eigenvectors are the features of the organic fertilizer formula component proportion data Direct composition; constructing the incidence matrix of the hypergraph If the node belongs to the hyperedge, the corresponding element of the matrix is 1, otherwise it is 0; during the training process, the incidence matrix is updated in real time according to the change of the organic fertilizer formula component proportion data over time or batch ; The formula component dynamic hypergraph network comprises a dynamic hypergraph convolution layer, which performs feature propagation and aggregation through a Laplacian operator on a hypergraph; the Laplacian operator on the hypergraph is obtained through matrix operation on a node degree matrix of a diagonal matrix , an association matrix of the hypergraph , and a hyperedge weight matrix ; wherein diagonal elements of the hyperedge weight matrix are calculated by inputting features of a hyperedge feature matrix, i.e., organic fertilizer formula component proportion data to a learnable adaptive weight mapping layer and performing normalization processing through a Softmax function. The formula component dynamic hypergraph network aggregates features of all nodes and hyperedges through a stack of four-layer dynamic hypergraph convolution layers, and outputs a formula global representation matrix that integrates complex chemical-biological interaction relationships among components .

5. The method for optimizing the formulation of hydrochar-based organic fertilizer according to claim 1, characterized in that: The conditional variational autoencoder comprises an encoder and a decoder, and utilizes the fertilizer efficiency labeled data as a constraint condition to convert the fertilizer efficiency labeled data into a latent space variable in a Gaussian distribution ; The encoder receives the fertilizer efficiency labeled data and outputs the mean vector and the variance vector ; the structure of the encoder comprises four fully connected layers: the first layer receives the fertilizer efficiency labeled data as input, accesses a batch normalization layer and a ReLU activation function; the second and third layers of fully connected layers gradually compress the feature dimension; the fourth layer is divided into two parallel branches of fully connected layers, which respectively output the mean vector and the variance vector ; Then, the target-oriented latent space variable is sampled from the mean vector and the variance vector by a reparameterization trick; the reparameterization trick first takes the logarithm of the variance vector and divides it by two, takes the exponential of the result to obtain the standard deviation term, then element-wise multiplies the standard deviation term with a random noise vector sampled from the standard normal distribution; finally, adds the result to the mean vector to obtain the target-oriented latent space variable as the backpropagation constraint of the conditional diffusion model.

6. The hydrochar-based organic fertilizer formulation optimization method of claim 5, wherein: The conditional diffusion model includes a forward noise adding process and a reverse denoising generation process; First, a forward noise adding process is performed to add Gaussian noise to the existing excellent formula ratio data In a preset time step, Gaussian noise is gradually added until it becomes completely disordered random noise ; Secondly, the reverse denoising generation process is carried out, and the target is to generate the optimal formula proportion from the random noise under the constraint of the determined target-oriented latent space variable and the guidance of the extracted formula global representation matrix , iteratively infer and denoise by a denoising network based on a residual U-Net structure ; The denoising network is composed of an encoder, a bottleneck layer and a decoder, wherein the corresponding layers of the encoder and the decoder adopt residual connection to retain detailed information.

7. The hydrochar-based organic fertilizer formulation optimization method of claim 6, wherein: The denoising network is specifically: The encoder receives the noisy recipe proportion data of the current time step The encoder is composed of four layers of residual blocks in series, and each residual block contains two convolution layers, a batch normalization layer, a ReLU activation function, and a residual connection; at the same time, a conditional injection module is embedded in each residual block, which is used to fuse the target-oriented latent space variable And a recipe global representation matrix to guide the denoising direction; after the output of each residual block, a down-sampling layer is accessed to reduce the spatial size of the feature map and increase the receptive field, and the final output of the encoder enters the bottleneck layer; The bottleneck layer is composed of two series of residual blocks for capturing the highest level of abstract features; The decoder receives the output of the bottleneck layer and is composed of four residual blocks and an up-sampling layer, which is symmetrical to the encoder in structure; before the processing of each residual block, the feature map size is increased by the up-sampling layer, and then the output of the up-sampling layer is fused with the features from the corresponding layer of the encoder through a splicing operation; then, the fused features enter the residual block of the decoder for feature extraction, and the structure of the residual block is the same as that of the residual block in the encoder; By guiding the residual U-Net structure in the target-oriented latent space variable and the formula global representation matrix , iteratively predict and remove noise, and finally get the optimal formula ratio of reverse generation .

8. The hydrochar-based organic fertilizer formulation optimization method of claim 6, wherein: The excellent formula proportion data refers to the top 10% of the formula in the comprehensive fertilizer efficiency score calculated by normalizing and weighted summing the fertilizer efficiency annotation data, and the specific calculation method of the comprehensive fertilizer efficiency score is: first, the nutrient release rate index, soil improvement evaluation and crop growth potential three index data are normalized by maximum and minimum value to eliminate the dimensional difference, and then the weighted sum method is used to calculate the comprehensive fertilizer efficiency score of each formula.

9. The hydrochar-based organic fertilizer formulation optimization method of claim 8, wherein: In the conditional diffusion model training process of S3, the Lagrange multiplier method is introduced as a constraint term of the loss function to force the optimal formula proportion generated The physical constraints are met, and the main constraints include that the total of the formula proportion is 100%, and the heavy metal content index must be within the safety threshold; To obtain the prediction value of heavy metal content, a safety evaluation regression model is constructed, which adopts a three-layer fully connected network structure: the first layer receives the spliced water and heat carbon characteristic attribute data and the proportion data of the organic fertilizer formula components, and performs initial feature fusion through a fully connected layer, a batch normalization layer and a ReLU activation function; the second layer further extracts nonlinear features through a fully connected layer, a batch normalization layer and a ReLU activation function; the third layer outputs a heavy metal content prediction scalar through a linearly activated fully connected layer, and the model takes the measured heavy metal content in the safety threshold marking as a training label, so as to predict the heavy metal content prediction value corresponding to any input formula through training .

Images