A method for formula optimization and performance prediction of brick and stone cultural relic protection materials

By combining the mechanism and data-driven models of CVAE and PINN, and embedding the material physicochemical control equations, the formulation of brick and stone cultural relic protection materials was optimized, solving the problems of multi-objective conflict and small sample size, and realizing high-precision material performance prediction and rapid research and development.

CN123120129APending Publication Date: 2026-11-17HUNAN UNIV OF SCI & ENG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611034960.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for the development of materials for the preservation of brick and stone cultural relics suffer from several problems, including conflicting objectives, long development cycles, poor timeliness, and black-box illusions due to small sample data. As a result, material formulations cannot simultaneously achieve high strength reinforcement and high permeability, and the models cannot perceive the underlying physical world.

Method used

A mechanism- and data-driven dual-engine model combining a conditional variational autoencoder generator network (CVAE) and a physical information neural network (PINN) is constructed. By embedding classical material physicochemical control equations as strong constraints, the material formulation is optimized to achieve high-precision prediction and inverse optimization of long-term weather resistance performance.

Benefits of technology

Achieving high-precision prediction within milliseconds shortens the R&D cycle and outputs material formulations that balance high strength reinforcement and high breathability. The prediction accuracy reaches 96.8%, resolving conflicts among multiple indicators and shortening the R&D cycle by 88.8%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure SMS_28
    Figure SMS_28
  • Figure SMS_73
    Figure SMS_73
  • Figure SMS_77
    Figure SMS_77
Patent Text Reader

Abstract

This invention discloses a method for optimizing the formulation and predicting the performance of brick and stone cultural relic protection materials, belonging to the interdisciplinary field of artificial intelligence and materials science. First, the invention collects the multi-dimensional heterogeneous material matrix parameters and spatiotemporal environmental parameters of the target brick and stone cultural relic, and constructs a feature control matrix after preprocessing. Second, the feature control matrix is ​​forcibly introduced into the latent space of CVAE to constrain the network to generate candidate material formulations. Then, a PINN performance prediction network is built, and the Washburn capillary water absorption kinetic equation and the thermal stress matching strain compatibility equation are embedded into the loss function in the form of physical residuals through automatic differentiation technology, realizing the hard constraint of physical laws on the network weight update. Finally, with a mutually exclusive optimization objective of a compressive strength improvement rate greater than 25% and a water vapor transmission rate retention rate greater than 80%, a multi-objective Pareto evolutionary algorithm is introduced to screen the optimal customized formulation. The proposed solution achieves a compressive strength improvement rate of 28.7% and a water vapor transmission rate retention rate of 84.3% after reinforcement, effectively coordinating the deep conflict between reinforcement performance and permeability, and significantly improving the model's generalization ability under small sample conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and materials science, and specifically relates to a method for optimizing the formula and predicting the performance of brick and stone cultural relic protection materials. Background Technology

[0002] Brick and stone cultural relics (such as grottoes, ancient city walls, and cliff carvings) are precious, non-renewable cultural heritages. Under long-term natural conditions, they inevitably face threats from a variety of complex diseases, including weathering, salt erosion, crumbling, and flaking. Developing and applying high-performance reinforcement and protective materials is a core means of delaying the weathering and damage of these relics. Due to the significant differences in the matrix of brick and stone relics from different regions and with different materials, the development of protective materials must adhere to the principle of precise customization, tailored to each individual relic. How to efficiently and scientifically design intelligent materials that balance high-strength reinforcement with physical compatibility is a core demand in the current field of emergency conservation of cultural relics.

[0003] Currently, the research and development of reinforcement and protection materials for brick and stone cultural relics both domestically and internationally mainly adopts the following two types of solutions: Option 1: Traditional materials science trial-and-error method. This method relies heavily on the personal engineering experience of materials experts for formulation (such as the ratio of ethyl silicate monomer, modified components, and catalysts). After the initial formulation is completed, extensive and long-term indoor accelerated aging and weathering tests must be conducted in the laboratory (such as hundreds of freeze-thaw cycles, wet-dry cycles, and UV weathering). By observing the measured performance degradation, the material formulation can be fine-tuned and finalized to protect it.

[0004] Option 2: Purely data-driven traditional deep learning prediction method. In recent years, with the development of computer technology, some scholars have attempted to introduce purely data-driven models such as ordinary multilayer perceptrons (MLP) or convolutional neural networks (CNN). This approach uses historical, fragmented experimental data (such as different component ratios and cured strength) as input for black-box supervised learning, thereby establishing a mapping network from input formulation to output performance indicators. The trained model is then used to assist in screening material formulations.

[0005] The two existing technical solutions mentioned above have revealed the following insurmountable technical bottlenecks in practical applications: (1) The conflict between multiple objectives is irreconcilable: Traditional experimental trial-and-error methods are extremely difficult to resolve the natural physical contradiction between "strength enhancement and high permeability". When protective materials improve the compressive strength of bricks and stones, they often overfill the micropores of the porous matrix, leading to a sharp deterioration in water vapor permeability. Once the cultural relic loses its permeability, the internal salt and alkali moisture cannot evaporate freely, causing more destructive surface flaking and large-area peeling. It is extremely difficult to find the Pareto optimal balance point in this kind of multi-objective conflict through manual trial and error.

[0006] (2) The research and development cycle is extremely long and the timeliness is poor: the accelerated weathering test in the traditional scheme usually takes several months or even years, resulting in low efficiency in the finalization of new protective materials and failing to meet the time window requirements for the rescue protection of endangered cultural relics.

[0007] (3) "Black box illusion" and physical paradox under small sample data: Due to the preciousness of cultural relics, large-scale destructive sampling experiments cannot be carried out, which makes the amount of material samples that can be trained extremely small (a typical characteristic of small samples). Pure data-driven general AI models are prone to serious overfitting when faced with such small samples. More seriously, because it belongs to pure black box computation, the model cannot perceive the underlying physical world and often generates illegal material formulas that completely violate the common sense of material physics (for example, calculating the physical paradox that capillary penetration depth increases in the opposite direction as porosity decreases), which lacks engineering practical value. Summary of the Invention

[0008] To address or partially resolve the problems existing in related technologies, this invention proposes a method for optimizing the formulation and predicting the performance of brick and stone cultural relic protection materials. It constructs a gradient update process that uses classical material physicochemical control equations as strong constraints, rigidly embedded into a neural network. Under conditions of extremely scarce experimental samples, through a novel paradigm driven by both mechanism and data, it achieves high-precision prediction of the long-term weather resistance performance of brick and stone cultural relic protection materials at the millisecond level. Furthermore, it precisely reverse-optimizes a highly compatible material composition formulation that balances "high-strength reinforcement" and "high-permeability retention," completely eliminating the illusion of an AI black box and shortening the R&D cycle from the traditional months to just a few days.

[0009] This invention provides a method for optimizing the formulation and predicting the performance of brick and stone cultural relic protection materials, comprising the following steps: (1) Collect the matrix parameters of the multi-dimensional heterogeneous materials of the target brick and stone cultural relics, and simultaneously extract the spatiotemporal environmental parameters during the service period of the target brick and stone cultural relics. After standardization and principal component analysis for dimensionality reduction, construct the feature control matrix in the multidimensional space. The matrix parameters of the multi-dimensional heterogeneous materials include mineral composition, density, porosity, and pore size distribution. The spatiotemporal environmental parameters include annual average temperature difference, ultraviolet radiation intensity, and annual average freeze-thaw cycle frequency. (2) Based on the generative model, the protective material formulation is generated in reverse. A CVAE (Conditional Variational Autoencoder) generative network is built. The feature control matrix is ​​forcibly introduced into the latent space of the autoencoder as an external conditional variable. Through the constraint of the conditional control vector, the CVAE model can reversely explore the high-dimensional parameter space for specific materials and environments and generate candidate material formulations in batches. The candidate material formulations include the ratio of tetraethyl orthosilicate, nano silica gel and organotin. (3) Construct a physical information neural network, and use automatic differentiation technology to embed the core physical control equations that control the seepage and mechanical deformation of porous media into its loss function to construct a data-mechanism dual-driven loss function. Then, perform forward evolution prediction on the anti-aging and various physicochemical properties of candidate material formulations under complex service scenarios. (4) Extract the multi-dimensional long-term evolution data output by the physical information neural network, and take the improvement rate of compressive strength after reinforcement and the retention rate of water vapor transmission rate as two deeply conflicting mutually exclusive optimization sub-objectives. Introduce a multi-objective evolution algorithm to optimize the Pareto front solution for all formulas that pass the physical test, and output and lock the final optimal intelligent customized formula.

[0010] Preferably, the standardization method involves processing the original heterogeneous parameter vector using the Z-score standardization operator. For any input feature x, its standardization mapping formula is: In the formula, This is the empirical arithmetic mean of this feature in the historical matrix database. The standard deviation is denoted as . The principal component analysis dimensionality reduction includes: constructing the covariance matrix of the standardized features. Eigenvalue decomposition yields the eigenvalue sequence. And its corresponding orthogonal eigenvectors. Based on the decoupling threshold set according to the cumulative contribution rate (preset to be greater than 95%), the first part is truncated. k Each principal component axis projects the original eigenvectors into a low-dimensional orthogonal space, generating the eigenvalue control matrix. . The innovative aspect of the conditional variational autoencoder generative network architecture described in this invention lies in using the decoupled projection matrix generated by cascading the multi-source heterogeneous material matrix characteristics of cultural relics with multi-factor service climate spatiotemporal parameters as the external strong boundary condition constraint space input. This forces the Gaussian distribution in the variational latent space to... It deeply encodes the intrinsic mechanical and seepage evolution trajectories of specific materials under specific climates, realizing the inverse cross-directional mapping from the hidden variable space to the nonlinear customized material chemical formulation space.

[0011] Preferably, the conditional control vector The feature control matrix described in step (1) The low-dimensional, compact feature vectors are constructed by concatenating the row vectors after flattening. These feature vectors are then directly concatenated to the fully connected layer of the encoder and the latent space sampling vectors during forward propagation. The rear side enables directional guidance for exploration of higher-dimensional space. Preferably, the latent space sampling vector To utilize the reparameterization trick from an adaptive Gaussian distribution The low-dimensional latent variable random feature vector extracted from it is expressed as follows: ,in To follow a standard normal distribution The random noise matrix.

[0012] Preferably, in step (3), the data-mechanism dual-driven total loss function is: Loss =w1L data + w2L physics Among them, L data For purely supervised data loss terms, the mean square error between predicted and measured values ​​in the laboratory is calculated; w1 and w2 are weighting coefficients, which are dynamically adjusted in real time using a gradient balancing algorithm to maintain the gradient magnitude balance between data-driven terms and physical constraint terms during backpropagation.

[0013] The key physical constraint loss term L physics Includes constraints from the Washburn capillary water absorption kinetic equation and the thermal stress matching and strain compatibility equation: ,in and These are the dynamic control weights for the capillary seepage residual term and the thermal strain compatibility residual term, respectively.

[0014] To forcefully incorporate partial differential equations into network updates, the autograd engine of frameworks like PyTorch is used to extract the first derivative of the prediction depth with respect to time t. Construct the root mean square residual loss That is, the Washburn capillary water absorption kinetic equation constrains: ; Indicates all The error of each time step (or data sampling point) is averaged. The actual rate of liquid level rise predicted by the model, where t is time; The theoretical physical speed, This represents the current depth of liquid penetration. r The average micropore radius of the target brick and stone artifact. c The surface tension of the material, i To wet the contact angle, or This represents the dynamic viscosity of the formulation solution.

[0015] If the network produces a black-box illusion prediction that violates the percolation conservation law, the residual will spike exponentially, forcibly pulling back the gradient update direction.

[0016] Thermal stress matching and strain compatibility equation constraints: A thermo-elastic-plastic constitutive equation under alternating temperature conditions is introduced to control the interfacial thermal expansion mismatch rate between the reinforcing layer and the masonry matrix. This constraint forces the network output prediction results to ensure that the strain gradient of the reinforcing layer and the matrix material is within a safe and continuous range when subjected to thermal stress, so as to prevent delamination caused by interfacial shear failure during long-term service.

[0017] Therefore, thermal stress matching and strain compatibility equation constraints : To prevent shear slippage and subsequent delamination between the reinforcement layer and the stone matrix, a one-dimensional interfacial strain continuity compatibility residual equation is introduced: In the formula, This represents the difference in thermal expansion coefficients between material interfaces. The temperature difference is alternating. This constraint forces the strain gradients of both components to remain within the safe threshold shear strain range. ). This represents the total number of discrete time steps corresponding to the alternating temperature cycle. To address the transient thermal stress in the reinforcing layer material at time t, The elastic modulus of the reinforcing layer material, The transient thermal stress of the matrix of the brick and stone cultural relic at time t. It represents the elastic modulus of the matrix of the brick and stone cultural relic.

[0018] Preferably, the multi-objective evolutionary algorithm is the Non-Dominated Sorting Genetic Algorithm (NSGA-II), with parameters including a population size of 100, a maximum number of generations of evolution of 300, a crossover probability of 0.9, a mutation probability of 0.1, and a selection operator based on crowding distance to maintain the diversity of the non-dominated solution set.

[0019] The technical solution provided by this invention has the following beneficial effects: (1) This invention, for the first time, directly incorporates the Washburn capillary flow derivative equation and the thermo-elastic-plastic shear strain conservation control equation for brick and stone cultural relic protection materials into the partial derivative spectrum structure of the Physical Information Neural Network Performance Prediction Model (PINN) in the form of residual penalty. This improvement ensures that even under small sample training, each step of the model's gradient update is subject to the forced trajectory regulation of classical physical science theorems. This fundamentally eliminates the generation of illegal material formulations that are not based on common sense in physics, achieving a prediction accuracy of 96.8%, and realizing a deep integration of scientific mechanisms and deep learning.

[0020] (2) This invention combines the CVAE generation model with the Pareto multi-objective algorithm to automatically decouple the entanglement between multiple indicators in a high-dimensional implicit parameter space strongly constrained by the seepage physics equation. This allows the solidified modified network to be precisely adsorbed onto the micropore walls of porous sandstone to form a nanoscale reinforcing and protective film, which not only provides a gentle and balanced mechanical improvement to the parent material (compressive strength steadily increases by 28.7%), but also completely opens the gas phase transport channels in the center of the porous medium, thereby achieving an extremely high water vapor permeability retention rate of 84.3%, perfectly resolving the technical deadlock of multiple conflicting indicators in the research and development of cultural relic protection materials.

[0021] (3) The method described in this invention can, within a two-day model calculation cycle, derive the decay curves of mechanical and compatibility indicators of candidate formulations after undergoing 100 or even 200 severe freeze-thaw cycles with high precision at the millisecond level. This successfully shortens the entire life cycle development and finalization cycle of new reinforcement and protection materials by more than 88.8%, providing highly timely intelligent technical support for the rescue projects of a large number of endangered and severely weathered brick and stone cultural relics in my country. Detailed Implementation

[0022] To better understand the technical content of this invention, specific embodiments are provided below to further illustrate this invention.

[0023] Example 1 To verify the absolute feasibility of this invention in resolving the aforementioned problems of "multi-index conflict" and "small sample generalization distortion," this embodiment uses the protection of a typical severely weathered sandstone cave in northern my country as a specific example, and designs and conducts scientific experiments: A method for optimizing the formulation and predicting the performance of brick and stone cultural relic protection materials includes the following steps: (1) Initialization of the experimental matrix and service scenario, i.e., construction of a multivariate heterogeneous dataset of the matrix and its environment: Weathered sandstone fragments were collected, and the initial basic material parameters were determined as follows: average porosity 16.5%, average micropore radius r = 2.4. The initial compressive strength is 14.2 MPa, the quartz content is 72%, and the initial water vapor transmission rate is [missing information]. The reference contact angle is 32°, and the density is... The environmental conditions are set as a high-frequency alternating freeze-thaw scenario in an extremely arid and semi-arid climate zone. The spatiotemporal environmental parameters include an average of over 100 freeze-thaw cycles per year and an average annual temperature difference variation. Ultraviolet radiation intensity .

[0024] After standardization and dimensionality reduction using principal component analysis, the initial basic materials science parameters and spatiotemporal environmental parameters are used to construct a feature control matrix in a multidimensional space. This matrix serves as the control condition vector for subsequent customized formulations, specifically including: To eliminate the distortion of high-dimensional spatial distance metrics caused by different physical dimensions (such as porosity percentage and temperature difference in degrees Celsius), the Z-score normalization operator is used to process the original 7-dimensional parameters (mineral composition, density, porosity, pore size distribution, initial compressive strength, initial water vapor transmission rate, and contact angle) and 3-dimensional spatiotemporal environmental parameters (annual mean temperature difference, ultraviolet radiation intensity, and annual freeze-thaw cycle frequency). For any input feature x, its normalization mapping formula is: ; In the formula, This is the empirical arithmetic mean of this feature in the historical matrix database. The standard deviation is denoted as . Then construct the covariance matrix of the normalized features. Eigenvalue decomposition yields the eigenvalue sequence. And its corresponding orthogonal eigenvectors. Based on the decoupling threshold set by the cumulative contribution rate (preset to be greater than 95%), the first k = 4 principal component axes are truncated, and the original 10-dimensional eigenvectors are projected onto a 4-dimensional low-dimensional orthogonal space to generate the feature control matrix. After PCA projection array After multiplication, the feature control vector is obtained: .

[0025] (2) CVAE model training and reverse generation of candidate material formulations based on generative models: First, 45 sets of historical laboratory measured small sample labeled datasets were prepared as supervision signals. The 45 sets of historical laboratory measured small sample labeled datasets include the corresponding mapping pairs of different tetraethyl orthosilicate / nano silica specific gravities, measured strength after curing, measured air permeability retention rate, and simulated freeze-thaw decay rate tested on various porous rocks in the past. They were mapped to [0, 1] using MinMax normalization as supervision signals.

[0026] Secondly, a conditional variational autoencoder generative network is constructed, and the specific architecture of the network is as follows: The encoder consists of three fully connected layers. Ultimately, it differentiates into the mean. With variance Two 2D latent space vector branches.

[0027] Decoder: The function of sampling vectors from the latent space. With conditional control vector Cascaded as a 6-dimensional input, through Finally, the mass percentages of the three components are output through the Softmax layer. The conditional control vector c is derived from the characteristic control matrix described in step (1). The low-dimensional, compact feature vectors are constructed by cascading after row vector flattening. By directly concatenating them to the back of the fully connected layer of the encoder and the latent space sampling vector z during forward propagation, directional guidance for exploring the high-dimensional space is achieved.

[0028] The feature control matrix is ​​forcibly introduced into the latent space of the CVAE encoder as an external condition to jointly train the CVAE network for 500 rounds. The model weights are backpropagated through the Adam optimizer and automatic differentiation to calculate the physical residuals.

[0029] When the absolute value of the fluctuation of the ELBO loss function, which is composed of the latent spatial variational loss and the reconstruction loss, is less than 50 consecutive epochs... The model is considered converged at this point. After convergence, the input feature control vector is... The model outputs 200 sets of modified ethyl silicate composite formulations in batch reverse engineering, which are the candidate material formulations.

[0030] (3) Construct a physical equation loss constraint performance prediction network model (PINN). The specific network framework topology of the physical information neural network (PINN) performance prediction model is as follows: The input layer is 4-dimensional, and the corresponding physical features are: the mass percentage of the three components (TEOS, nano-SiO2 sol, and catalyst mass fraction) and the accelerated aging evolution time. t The hidden layer consists of four cascaded fully connected feedforward layers, each containing 128 neurons. The activation functions of the hidden layers all employ the adaptive hyperbolic tangent (Tanh) function, which is second-order continuously differentiable, to ensure the continuous and smooth propagation of gradients during automatic differentiation. The output layer is two-dimensional, corresponding to the predicted output over time. tThe curves showing the rate of increase in compressive strength and the rate of retention of water vapor permeability under long-term evolution. By using Automatic Differentiation technology, the core physical control equations that control the seepage and mechanical deformation of porous media are rigidly embedded into their loss functions to construct a data-mechanism dual-driven total loss function. Then, the anti-aging and compressive strength improvement rates and water vapor permeability retention rates of 200 modified ethyl silicate composite formulations under complex service scenarios are predicted forward. The total loss function driven by the data and mechanism is as follows: Loss =w1L data + w2L physics Among them, L data For purely supervised data loss, the mean square error between the predicted values ​​of the deep fully connected network and the experimental values ​​is calculated; w1 and w2 are weight coefficients, which are dynamically adjusted in real time using a gradient balancing algorithm to maintain the balance of gradient magnitudes between the data-driven terms and the physical constraint terms during backpropagation.

[0031] Physical constraint loss term L physics Includes constraints from the Washburn capillary water absorption kinetic equation and the thermal stress matching and strain compatibility equation: ,in and These are the dynamic control weights for the capillary seepage residual term and the thermal strain compatibility residual term, respectively. In this embodiment, a dynamic gradient balancing algorithm based on the Neural Tangent Kernel (NTK) is used to adaptively adjust the weights in real time during each training epoch. and The value is determined to maintain the consistency of the gradient magnitude of the two physical constraints during backpropagation.

[0032] To forcefully incorporate partial differential equations into network updates, the autograd engine of frameworks like PyTorch is used to extract the first derivative of the prediction depth with respect to time t. Construct the root mean square residual loss That is, the Washburn capillary water absorption kinetic equation constrains: ; symbol Indicates all The root mean square (RMS) error is calculated from the physical equation residuals of each discrete spatiotemporal sampling point. The actual rate of liquid level rise predicted by the model, where t is time; The theoretical physical speed, This represents the current depth of liquid penetration. r The average micropore radius of the target brick and stone artifact. c The surface tension of the material, i To wet the contact angle, or This represents the dynamic viscosity of the formulation solution.

[0033] If the network produces a black-box illusion prediction that violates the percolation conservation law, the residual will spike exponentially, forcibly pulling back the gradient update direction.

[0034] Thermal stress matching and strain compatibility equation constraints: A thermo-elastic-plastic constitutive equation under alternating temperature conditions is introduced to control the interfacial thermal expansion mismatch rate between the reinforcing layer and the masonry matrix. This constraint forces the network output prediction results to ensure that the strain gradient of the reinforcing layer and the matrix material is within a safe and continuous range when subjected to thermal stress, so as to prevent delamination caused by interfacial shear failure during long-term service.

[0035] Therefore, thermal stress matching and strain compatibility equation constraints : To prevent shear slippage and subsequent delamination between the reinforcement layer and the stone matrix, a one-dimensional interfacial strain continuity compatibility residual equation is introduced: In the formula, This represents the difference in thermal expansion coefficients between material interfaces. The temperature difference is alternating. This constraint forces the strain gradients of both components to remain within the safe threshold shear strain range. ).

[0036] (4) Extract the multi-dimensional long-term evolution prediction data (compressive strength improvement rate and water vapor transmission rate retention rate after 100 freeze-thaw cycles) of 200 candidate material formulations output by PINN. Using a compressive strength improvement rate greater than 25% and a water vapor transmission rate retention rate greater than 80% as two deeply conflicting mutually exclusive optimization sub-objectives, a fast non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy is introduced as a multi-objective evolutionary operator to optimize all candidate material formulations that pass the physical test using Pareto front solutions. In this embodiment, the algorithm parameters are specifically configured as follows: crossover probability is set to 0.9, mutation probability is set to 0.1, initial population size is 100, and maximum number of generations is set to 300. Pareto front solutions are then used to optimize all candidate material formulations that pass the physical test, and the final optimal intelligent customized formulation is output and locked.

[0037] The optimal custom formulation process parameters, determined by the Pareto multi-objective algorithm, are as follows: 82.4% TEOS (main component), 14.1% nano-SiO2 modified sol, and 3.5% dibutyltin dibutyldiocyanate catalyst. After 100 high-frequency freeze-thaw cycles, the optimal custom formulation showed a predicted compressive strength increase of 28.5% and a predicted water vapor transmission rate retention of 84.8%.

[0038] (5) Preparation of physical formulation and verification of accelerated aging effect: According to the intelligent recommended formula ratio: the main component TEOS mass fraction is 82.4%, the mass fraction of nano-SiO2 modified sol is 14.1%, and the mass fraction of dibutyltin dibutyldiocyanate catalyst is 3.5%. The modified material was precisely prepared in the laboratory and applied to experimental sandstone blocks using a capillary impregnation method. After curing at room temperature for 28 days, the reinforced sandstone blocks were placed in a fully automated high and low temperature freeze-thaw test chamber for accelerated weathering resistance testing. Every 10 freeze-thaw cycles (-20 to 20 degrees Celsius) constituted one cycle, and each complete freeze-thaw cycle was set to a standard 4 hours (including 2.5 hours of air freezing at -20°C and 1.5 hours of water thawing at 20°C). Compressive strength and water vapor transmission rate were measured once, until 100 cycles were completed, and the measured data were collected.

[0039] The actual test results show that the compressive strength increased by 28.7% after 28 days of curing, and the water vapor permeability retention rate was 84.3%.

[0040] Through forward extrapolation using the PINN network of this invention, the optimal customized formulation, after undergoing 100 high-frequency freeze-thaw cycles, is predicted to have a compressive strength increase of 28.5% and a water vapor permeability retention of 84.8%. Longitudinal comparison with actual measured data from physical samples after 28 days of curing (strength increase of 28.7%, air permeability retention of 84.3%) shows that the mean square error of absolute extrapolation (MAE) is only 0.2% and 0.5%, respectively. This strongly demonstrates that, under extremely scarce small sample conditions, this invention successfully avoids the black box illusion through strong mechanistic anchoring, achieving an astonishing prediction accuracy of 96.8%.

[0041] To clearly demonstrate the significant technological advancements achieved by this invention, a rigorous horizontal comparison was conducted between the core technical indicators and measured evolution effects of the proposed solution (CVAE + PINN mechanism dual-driven deep learning) and Scheme 1 (traditional experimental trial-and-error method) and Scheme 2 (pure data-driven traditional AI model) in the same sandstone protection example. The specific results are detailed in the table below.

[0042] Option 1: Traditional experimental trial and error method: This traditional method relies entirely on expert experience to arbitrarily adjust the mass fraction. Without the premise of multi-objective overall planning, the sample is configured and then sent into a high and low temperature test chamber for a real freeze-thaw cycle test of up to 180 days. It is an extremely time-consuming and passive exhaustive trial and error method that explores the results after the fact.

[0043] Option 2: Pure Data-Driven Traditional AI Model: This option uses a conventional 4-layer Multilayer Perceptron (MLP). The input is directly coupled with three sets of allocation ratios, and the output is directly correlated with material properties. Pure supervised fitting is performed using the MSE loss function. However, due to the lack of embedding the Washburn capillary flow equation and thermal stress compatibility equation, and the extremely limited sample size of only 45 artifacts, the network neurons suffered severe overfitting and high-dimensional pathological bifurcation. To unilaterally cater to the supervision label of increased compressive strength, the model arbitrarily ordered excessive filling of nano-silica, resulting in a false high-strength formula that completely violated capillary mechanics. This led to a large number of micropores being sealed by inorganic matter in actual measurements, causing the water vapor permeability to plummet to a devastating 65.4%.

[0044] Table 1 Table 1 shows that while Scheme 2 offers faster calculation speed, the lack of constraints based on the physical laws of the underlying materials leads to destructive "severe capillary blockage" in the optimized formula, resulting in a water vapor permeability of only 65.4%. In actual protection projects, this will accelerate the accumulation of salt and alkali beneath the surface of the grotto, directly inducing more severe blistering. The technical solution of this invention, however, perfectly overcomes this technical bottleneck.

[0045] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for optimizing the formulation and predicting the performance of brick and stone cultural relic protection materials, characterized in that, Includes the following steps: (1) Collect the matrix parameters of the multi-dimensional heterogeneous materials of the target brick and stone cultural relics, and simultaneously extract the spatiotemporal environmental parameters during the service period of the target brick and stone cultural relics. After standardization and principal component analysis to reduce the dimensionality of the collected multi-dimensional heterogeneous material matrix parameters and spatiotemporal environmental parameters, construct the feature control matrix in the multidimensional space. The matrix parameters of the multi-component heterogeneous materials include mineral composition, density, porosity, pore size distribution, initial compressive strength, initial water vapor transmission rate, and contact angle; the spatiotemporal environmental parameters include annual average temperature difference, ultraviolet radiation intensity, and annual average freeze-thaw cycle frequency. (2) Construct a conditional variational autoencoder generator network, and forcibly introduce the feature control matrix as an external conditional variable into the latent space of the autoencoder. Through the constraint of the conditional control vector, generate candidate material formulations. The candidate material formulations include the ratio of ethyl silicate, nano silica gel and organotin. (3) Construct a performance prediction network model with physical equation loss constraints, build a deep fully connected network for forward evolution prediction of the anti-aging and various physicochemical properties of candidate material formulations under complex service scenarios, and use automatic differentiation technology to forcibly embed the core physical control equations that control the seepage and mechanical deformation of porous media into the loss function of the deep fully connected network to construct a data-mechanism dual-driven total loss function, forcing the gradient update direction of the network weights to obey the boundary constraints of material physics. (4) Extract the multi-dimensional long-term evolution data of the performance prediction network model output by the physical equation loss constraint. Take the improvement rate of compressive strength after reinforcement greater than 25% and the water vapor transmission rate retention rate greater than 80% as two deeply conflicting mutually exclusive optimization sub-objectives. Introduce a multi-objective evolution algorithm to optimize the Pareto front solution of all candidate material formulations that have passed the physical test, and output and lock the final optimal intelligent customized formulation.

2. The method for formula optimization and performance prediction of brick and stone cultural relic protection materials as described in claim 1, characterized in that, The standardization specification involves processing the original parameter vector using the Z-score standardization operator. For any input feature x, the standardization mapping formula is: In the formula, This is the empirical arithmetic mean of this feature in the historical matrix database. Standard deviation; The original parameter vector includes matrix parameters of multi-component heterogeneous materials and spatiotemporal environment parameters; The principal component analysis dimensionality reduction specifically includes: constructing the covariance matrix of standardized features. , where n represents the number of samples. For an n×d matrix, eigenvalue decomposition of the covariance matrix yields the eigenvalue sequence. and its corresponding orthogonal eigenvectors, where d is the overall feature dimension of the original multi-source heterogeneous parameter vectors; based on the decoupling threshold set according to the cumulative contribution rate, the first... Each principal component axis projects the original eigenvectors into a low-dimensional orthogonal space, generating the eigenvalue control matrix. , where m is the total number of samples or batch size of the current input historical matrix dataset.

3. The method for formula optimization and performance prediction of brick and stone cultural relic protection materials as described in claim 2, characterized in that, The decoupling threshold set for the cumulative contribution rate is greater than 95%.

4. The method for formula optimization and performance prediction of brick and stone cultural relic protection materials as described in claim 1, characterized in that, The conditional control vector is a low-dimensional, compact feature vector constructed by cascading the feature control matrix after flattening the row vectors. By directly splicing it to the back of the encoder's fully connected layer and the latent space sampling vector during forward propagation, it enables directional guidance for exploring the high-dimensional space.

5. The method for formula optimization and performance prediction of brick and stone cultural relic protection materials as described in claim 1, characterized in that, The total loss function driven by the data and mechanism is as follows: Loss =w1 L data + w2 L physics; Among them, L data For purely supervised data loss, the mean square error between the predicted values ​​of the deep fully connected network and the laboratory measured values ​​is calculated; w1 and w2 are weighting coefficients. Physical constraint loss term L physics Including the constraints of the Washburn capillary water absorption kinetic equation and the thermal stress matching and strain compatibility equation, the expression is: ,in and These are the dynamic control weights for the capillary seepage residual term and the thermal strain compatibility residual term, respectively. The Washburn capillary water absorption kinetic equation is constrained as follows: ; Indicates all The error of each time step or data sampling point is averaged. The actual rate of liquid level rise predicted by the model, where t is time; The theoretical physical speed, This represents the current depth of liquid penetration. r The average micropore radius of the target brick and stone artifact. γ The surface tension of the material, θ To wet the contact angle, η The dynamic viscosity of the formulation solution; The constraints of the thermal stress matching and strain compatibility equations are as follows: In the formula, This represents the difference in thermal expansion coefficients between material interfaces. The coefficient of thermal expansion of the reinforcing layer material, The coefficient of thermal expansion of the matrix of the brick and stone cultural relic. For alternating temperature difference, This represents the total number of discrete time steps corresponding to the alternating temperature cycle. To address the transient thermal stress in the reinforcing layer material at time t, The elastic modulus of the reinforcing layer material, The transient thermal stress of the matrix of the brick and stone cultural relic at time t. It represents the elastic modulus of the matrix of the brick and stone cultural relic.

6. The method for formula optimization and performance prediction of brick and stone cultural relic protection materials as described in claim 1, characterized in that, The multi-objective evolutionary algorithm is a fast non-dominated sorting genetic algorithm with an elitist strategy.