CN-121725956-B - Reverse synthesis method of carbon-based environment functional material

Abstract

The invention provides a reverse synthesis method of a carbon-based environment functional material, which belongs to the technical field of computer material science, and comprises the steps of firstly carrying out collaborative collection and joint data set construction of multisource heterogeneous data of the carbon-based environment functional material, then building a reverse generation model of a carbon aerogel Kong Tapu structure by relying on a deep generation countermeasure framework, converting macroscopic performance indexes into high-dimensional constraint characteristics by means of a performance coding network, and realizing directional generation of microscopic pores by a self-adaptive modulation mechanism of a synthesis network. Meanwhile, an embedded physical attribute verification and countermeasure cooperative optimization strategy is implemented, an embedded physical regression verification head of the authentication network is multiplexed, and a self-supervision feedback closed loop of the agent-free model is constructed. Finally, based on multidimensional feature decoupling, technological parameter inversion and intelligent decision recommendation are completed, and the optimal microstructure features are mapped into an executable synthesis process formula through morphological quantitative verification and a technological parameter inversion network.

Inventors

• WANG WEILIANG
• ZHOU ZIYU
• ZHAO ERLING
• MA YONGCHAO
• REN LING
• GAO XINGANG
• DONG WENPING
• SUN GAOJIE
• NI CHENBING
• Lan Xuefang

Assignees

• 青岛理工大学

Dates

Publication Date

20260508

Application Date

20260211

Claims (6)

1. 1. The reverse synthesis method of the carbon-based environment functional material is characterized by comprising the following steps of: S1, converting synthesis process parameters into uniform process parameter vectors and establishing unique indexes, acquiring carbon aerogel pore structure images under different processes, and simultaneously acquiring corresponding microscopic pore structure and surface chemical information and environmental performance indexes as performance vector labels; The microscopic pore structure and surface chemical information comprises a specific surface area numerical value, a pore diameter distribution vector, a total pore volume numerical value, a heteroatom content vector and a defect index; The environmental performance indexes comprise adsorption capacity, selectivity coefficient, kinetic rate constant and cycle retention rate; corresponding association is carried out by taking the index number as a main key, and a training data set is established; S2, constructing a carbon aerogel Kong Tapu structure reverse generation model, namely, a carbon aerogel performance coding network, a carbon aerogel Kong Tapu structure synthesis network, a carbon aerogel Kong Tapu structure identification network, a multi-scale pyramid architecture and a carbon aerogel structure analysis network, wherein the performance vector containing microscopic pore structure, surface chemical information and environmental performance indexes is mapped into an implicit carbon aerogel target performance embedding vector; The carbon aerogel performance coding network comprises a cascading characteristic extraction layer and a semantic mapping layer, wherein a set normalized carbon aerogel performance index vector is firstly obtained, the vector is formed by sequentially splicing seven values of a specific surface area numerical value, a pore diameter distribution vector, a total pore volume numerical value, a heteroatom content vector, a defect index, an adsorption capacity, a selectivity coefficient, a dynamic rate constant and a cycle retention rate; The carbon aerogel Kong Tapu structure synthesis network adopts a step-by-step up-sampling architecture based on condition self-adaptive normalization, and specifically comprises a potential space projection layer, four cascade self-adaptive modulation up-sampling modules, a linear full-connection projection layer and an image generation layer; firstly, sampling from standard normal distribution to obtain random noise vectors, inputting the random noise vectors into a potential space projection layer to perform dimensional transformation and space remolding to obtain an initial feature base map with a certain length and width size and channel number, secondly, inputting the initial feature base map into a cascade self-adaptive modulation up-sampling module, wherein each module comprises a deconvolution operation layer and a self-adaptive instance normalization layer; The carbon aerogel Kong Tapu structure identification network adopts a multi-scale characteristic pyramid identification architecture and comprises three identification sub-networks with different scales, a global average pooling layer and a physical regression check head; the method comprises the steps of simultaneously downsampling a carbon aerogel pore structure synthetic image and an original carbon aerogel pore structure image into three resolution versions of an original size, a half size and a quarter size, respectively inputting three identification sub-networks, wherein each identification sub-network is formed by stacking a convolution layer and a spectrum normalization layer, is respectively responsible for capturing sharpness of pore wall edges, uniformity of pore arrangement and connectivity characteristics of an integral skeleton under different receptive field scales, independently outputting an authenticity judgment result matrix corresponding to the carbon aerogel pore structure synthetic image and the original carbon aerogel pore structure image, respectively carrying out weighted summation after calculating losses of the authenticity judgment result matrix and a real label matrix and a counterfeit label matrix, and simultaneously applying antagonism constraint on three layers of fine textures, local modes and global topologies; Respectively inputting different scale feature images output by the tail ends of three identification sub-network stacked convolution layers into a global average pooling layer for global average pooling treatment to obtain three pooled feature vectors with different scales, and splicing the pooled feature vectors with the three different scales in channel dimensions to form a global abstract semantic feature vector representing the whole topological information of the image; S3, constructing an objective function and executing countermeasure training, and guiding the carbon aerogel Kong Tapu structure to synthesize a network to adjust a microstructure by utilizing an attribute regression gradient returned by the carbon aerogel Kong Tapu structure identification network; s4, inputting a desired carbon aerogel performance vector, and generating a model by reversely generating the structure of the carbon aerogel Kong Tapu after S3 training to generate a candidate carbon aerogel pore structure synthetic image set with specific pore distribution characteristics.
2. 2. The method for reversely synthesizing the carbon-based environment functional material according to claim 1, which is characterized by further comprising the following steps: The method comprises the steps of comprehensively scoring images in a candidate carbon aerogel pore structure synthetic image set by utilizing three different-scale identification sub-networks of a trained carbon aerogel Kong Tapu structure identification network and a physical regression check head, screening out a carbon aerogel optimal pore structure synthetic image which is optimal in physical consistency and accords with the preset performance of a user, carrying out micro morphology quantitative verification again, carrying out topology feature extraction on the carbon aerogel optimal pore structure synthetic image to obtain a morphological feature vector, finally carrying out synthesis process parameter reverse mapping, constructing a process parameter inversion network, decoupling micro pore structure features of the carbon aerogel optimal pore structure synthetic image and mapping the micro pore structure features back to a process parameter space, and outputting executable process parameters corresponding to S1.
3. 3. The method for reversely synthesizing the carbon-based environment functional material according to claim 1, wherein the process parameter set comprises six continuous process parameters including an activation temperature, an activation time, an activator-to-precursor mass ratio, a dopant proportion, an atmosphere flow rate and a heating rate, and two discrete process parameters including a precursor type and an activator type; The method comprises the steps of carrying out linear normalization processing on six continuous process parameters according to preset dimension upper and lower bounds respectively, mapping the six continuous process parameters to a numerical value interval from zero to one, carrying out independent thermal encoding processing on two discrete process parameters respectively, converting the two discrete process parameters into a sparse vector format, finally splicing the normalized continuous parameters with the encoded discrete parameters to form a unique process parameter vector of a current sample, and distributing a globally unique sample index number.
4. 4. The method for reversely synthesizing the carbon-based environment functional material according to claim 1, which is characterized in that: Firstly, inputting an original carbon aerogel pore structure image into the identification network to obtain a corresponding true and false identification result matrix, calculating the mean square error between each element value in the matrix and a true label matrix to obtain true image identification loss, inputting a carbon aerogel pore structure synthetic image generated by a carbon aerogel Kong Tapu structure synthetic network into the identification network to obtain a corresponding true and false identification result matrix, calculating the mean square error between each element value in the matrix and a false label matrix to obtain generated image identification loss, secondly, predicting a resampled carbon aerogel pore structure image by using a physical regression check head in the identification network, outputting a predicted carbon aerogel performance vector label, calculating the mean square error between a predicted vector label and the true carbon aerogel performance label corresponding to a data set to obtain physical regression supervision loss, and finally summing the true image identification loss, the generated image identification loss and the physical regression supervision loss according to preset weight to obtain the identifier loss function; The method comprises the steps of defining and calculating a generator loss function aiming at a carbon aerogel Kong Tapu structure synthesis network, wherein the loss function consists of anti-deception loss and attribute consistency loss, firstly inputting a carbon aerogel pore structure synthesis image generated by the carbon aerogel Kong Tapu structure synthesis network into a carbon aerogel Kong Tapu structure identification network, obtaining a corresponding true and false identification result matrix, calculating a mean square error between each element value in the matrix and a true label matrix to serve as the anti-deception loss, wherein the loss forces texture details of a generated image to approach a true sample, secondly, predicting the carbon aerogel pore structure synthesis image by utilizing a physical regression check head, outputting a predicted carbon aerogel performance vector label, calculating a mean square error between the predicted vector label and a carbon aerogel performance index vector serving as an input condition of the carbon aerogel performance coding network to serve as the attribute consistency loss, and finally summing the anti-deception loss and the attribute consistency loss to obtain the generator loss function.
5. 5. The method for reversely synthesizing the carbon-based environment functional material according to claim 2, wherein the method is characterized in that the method screens out the carbon aerogel optimal pore structure synthetic image which has optimal physical consistency and accords with the preset performance of a user, and comprises the following specific processes: Sequentially inputting the generated images in the synthesized image set of the candidate carbon aerogel pore structure into a carbon aerogel Kong Tapu structure identification network, and performing two-dimensional assessment in parallel, wherein on one hand, identification sub-networks with different scales output an authenticity identification result matrix of each image, and arithmetic average values of all elements in the matrix are calculated to be used as visual reality scores, on the other hand, a physical regression check head outputs the predicted carbon aerogel performance corresponding to each image, and calculates multi-dimensional Euclidean distance between a predicted vector and a carbon aerogel performance vector input by a user to be used as response precision to a design target, the visual reality scores and the response precision are weighted and fused into comprehensive confidence scores according to preset weights, and the candidate image sets are arranged in descending order according to the scores, so that the image with the highest comprehensive confidence is automatically intercepted to be used as the optimal pore structure design scheme of the finally recommended carbon aerogel.
6. 6. The method for reversely synthesizing the carbon-based environment functional material is characterized in that a self-adaptive wiener filtering algorithm is utilized to remove high-frequency multiplicative noise in an image of an optimal pore structure design scheme of the carbon aerogel, a maximum inter-class variance method is adopted to calculate a global optimal threshold value, a gray level image is divided into a binary pore mask and a skeleton mask; The method comprises the steps of constructing and utilizing a lightweight process parameter inversion network, taking an output morphological feature vector as an input, outputting a corresponding standardized process parameter vector, wherein the process parameter inversion network consists of a six-layer convolution feature extractor, a fully connected regression layer and a Softmax classification layer, and performing inverse normalization processing on the standardized process parameter vector outputted by inversion to restore the process parameter with an actual physical dimension.

Description

Reverse synthesis method of carbon-based environment functional material Technical Field The invention belongs to the technical field of computer material science, and particularly relates to a reverse synthesis method of a carbon-based environment functional material. Background In the technical fields of environmental functional materials and green low carbon, the carbon-based environmental functional materials are widely applied to a plurality of key directions such as water pollution control, soil remediation, atmospheric pollutant adsorption, electrochemical catalytic degradation, environmental energy storage coupling and the like due to wide sources, strong structural adjustability, high chemical stability and good environmental adaptability. Such materials typically include, but are not limited to, activated carbon, biomass carbon, carbon aerogel, porous carbon materials, heteroatom doped carbon materials, and the like, whose environmental functional properties are primarily manifested in adsorption capacity and selectivity for specific contaminants, reactivity and stability, electrochemical response characteristics, and structural retention under multiple cycles. The functional performance of the macro environment is not determined by a single factor, but is fundamentally limited by the synergistic effect of a complex microstructure system inside the material, and the microstructure features particularly relate to multi-scale and multi-level topological features such as pore size distribution form, specific surface area, pore channel connectivity, pore wall thickness, surface functional group type and spatial distribution, heteroatom doping position, carbon skeleton defect structure and the like. However, in practical development and engineering application processes, the design and preparation of carbon-based environmental functional materials still highly depend on the traditional forward trial-and-error development paradigm. Researchers usually preset synthesis conditions such as biomass or chemical precursor types, types and amounts of activators, pyrolysis or activation temperatures, heating rates, reaction atmospheres, heat preservation time, post-treatment process parameters and the like based on experience or existing literature, then prepare samples through experiments, and characterize microstructure and macroscopic properties of the materials by means of scanning electron microscopy, specific surface area testing, adsorption property testing and the like. In the forward research and development mode, optimization of material performance often depends on a large number of repeated experiments and artificial experience accumulation, and an efficient and systematic design path is difficult to form. The traditional forward research and development method exposes a series of significant defects in the field of carbon-based environmental functional materials, meanwhile, in an actual application scene, specific performance indexes, such as selective adsorption capacity under complex water chemistry conditions or catalytic stability in a specific potential window, are often required to be preset in advance for specific pollutant systems or service conditions, and the corresponding ideal microstructure morphology and synthesis path thereof are difficult to reversely derive from the target performance indexes directly by the existing method. In addition, a large amount of carbon-based material formulas, process parameters and microstructure data accumulated by different scientific research institutions and manufacturing enterprises in the long-term research and development process are generally stored in respective local systems in a scattered manner, and often core technology secrets or business sensitive information is involved, so that centralized sharing and unified modeling are difficult to realize on the premise of not revealing the privacy. The reality condition objectively restricts the construction of a large-scale high-quality data set, and further restricts the popularization and application of the data driving type material design model in actual industrial scenes. Therefore, how to realize the reverse design of the intelligent material crossing the main body on the premise of ensuring the data privacy and the safety becomes a key technical problem to be solved in the field. From the prior art means, the material design field mainly adopts a molecular dynamics simulation and traditional machine learning prediction method. Although molecular dynamics simulation can reveal the material formation mechanism from atomic or molecular scale, the method has high computational complexity and limited time scale, and is difficult to effectively simulate a true pore structure network of a carbon-based material in a micron or even larger scale, and a micro morphology result which can be used for guiding process regulation cannot be directly generated. The traditional machine learning method is usual