A method and system for acquiring porous material adsorption data based on physical constraint artificial intelligence

Abstract

The present application relates to a kind of porous material adsorption data acquisition method and system based on physical constraint artificial intelligence, the actual characteristics of the method in combination with porous material descriptor, define a kind of new adsorption amount prediction method with physical meaning, based on the method can quickly obtain the adsorption performance data of target material in given temperature and pressure range, to assist researchers to quickly assess the application potential of material, ultimately reach the purpose of accurate prediction, reduce material development cost, improve research efficiency.

Description

Technical Field

[0001] This invention relates to a method and system for acquiring adsorption data of porous materials based on physical constraint artificial intelligence, belonging to the field of artificial intelligence material design technology. Background Technology

[0002] The development of adsorption separation materials is of great significance to industries such as chemical engineering, energy, and environmental protection. Through the rational design of porous materials, they can achieve efficient separation of complex mixtures, such as low-carbon hydrocarbon separation, pollutant adsorption, and carbon capture and storage, ultimately contributing to the clean energy transition and sustainable development. Currently, most related research relies on individual experiments and simulations of materials. Experiments include material synthesis, characterization, and testing, while simulations include method selection and model building. Both methods have certain limitations, necessitating the development of more efficient methods and systems for evaluating material adsorption performance to reduce research costs and improve efficiency. In evaluating the gas adsorption performance of materials, both experimental and simulation methods have shortcomings. Experimental evaluation involves material synthesis, characterization, and testing, especially when examining the adsorption performance of multiple gases, requiring multiple tests, resulting in a lengthy and costly process. Simulation evaluation also involves individual modeling of each material, which is inefficient and heavily reliant on computational power. Artificial intelligence is an effective tool for solving the above problems, but most artificial intelligence is actually a "black box" at present, lacking physical interpretability, which limits its extrapolation ability, or even lacks extrapolation ability at all, and makes it even more impossible to accurately predict the adsorption performance of target materials. Summary of the Invention

[0003] To address the problems of lengthy and expensive experimental cycles, low simulation efficiency, limited application scope, and poor extrapolation capabilities of traditional artificial intelligence models, this invention innovatively adopts a physical constraint-based artificial intelligence technology to establish a more efficient and accurate method and system for evaluating the adsorption performance of porous materials, effectively reducing the R&D cost of adsorption separation materials and improving R&D efficiency.

[0004] A method for predicting the adsorption performance of porous materials based on physical constraint artificial intelligence includes:

[0005] 1) Obtain the material description data of the target material and determine the adsorption performance label to be predicted and the prediction operating parameters, including the prediction temperature, one or more pressure points, and the type of adsorbate to be predicted; construct a descriptor vector X based on the material descriptor data; obtain the adsorption Henry's constant K of the target material for the adsorbate to be predicted at the prediction temperature. 𝐻 ;

[0006] 2) Determine whether the material descriptor data meets the preset input conditions. The preset input conditions include at least missing value checks and outlier checks. When the material descriptor data includes adsorption isotherm data, the fitability check of the adsorption isotherm data is also included. If the conditions are not met, output a data supplementation or modification prompt and receive the supplemented material descriptor data.

[0007] 3) Extract sample materials of the same type as the target material from the database, obtain the material descriptor data of the sample materials, and construct a descriptor vector 𝑋. 𝑠 And the Henry's constant for adsorption of the sample material for the adsorbate to be predicted at the predicted temperature. 𝐻,𝑠 The actual adsorption amount Q of the sample material under the same conditions as the predicted operating parameters is used as the adsorption performance label data.

[0008] 4) Perform consistency processing on the data obtained in step 3) to ensure that the descriptor vector, pressure point, adsorption Henry's constant, and adsorption performance label data of each sample material meet the model input requirements;

[0009] 5) Construct training and testing sets based on the descriptor vectors and adsorption performance label data of the sample materials;

[0010] 6) Train the physical constraint artificial intelligence model, and use the descriptor vector X in the training set. s Pressure point P and adsorption Henry's constant K 𝐻,𝑠 The physical constraint AI model is input, and the physical constraint AI model is passed through parameterizable functions. Predicted adsorption amount by combining physical constraint relationships and according to Compared with the actual adsorption amount The error between them is used to construct a loss function to optimize the loss function. The parameters, wherein the physical constraint relationship is:

[0011]

[0012] Where P is the pressure point in the predicted operating condition parameters, KH is the adsorption Henry's constant corresponding to the predicted temperature and the adsorbate to be predicted, 𝑋 is the descriptor vector, and 𝑒 −10 This is a stable term used to avoid a denominator of zero;

[0013] 7) Evaluate the training results to obtain a physical constraint artificial intelligence model that meets the preset evaluation criteria;

[0014] 8) Using the physical constraint artificial intelligence model, based on the descriptor vector 𝑋 of the target material and the adsorption Henry's constant KH The predicted operating conditions parameters are used to predict the adsorption performance of the target material under the predicted operating conditions parameters, and the prediction results are output.

[0015] The parameterizable function The function can be linear or nonlinear; the nonlinear function includes at least one or more of the following: polynomial function, exponential function, logarithmic function, and power function; wherein the linear function includes at least: The The parameters to be trained, For descriptor components.

[0016] The physical constraint artificial intelligence model is a physical constraint neural network; the physical constraint neural network includes an input layer, at least one fully connected hidden layer, and an output layer, and the modified linear unit ReLU is used as the activation function between adjacent fully connected layers; the physical constraint neural network uses descriptor vectors Input and output for calculation core parameters Or output used for characterization The function parameters.

[0017] The loss function in step 6) is the sum of the losses on the training set for Q. pred An error function obtained by summing or averaging the error with the true adsorption amount Q, wherein the error function includes one or more of the following: mean square error, mean absolute error, and root mean square error.

[0018] The similar materials mentioned in step 3) are a collection of materials in the parent material category to which the target material belongs. The parent material category includes one or more of the following: metal-organic framework materials, covalent organic framework materials, zeolites, porous carbon materials, porous organic cages, and ionic porous materials.

[0019] The material descriptor data includes one or more of the following: structural descriptors, chemical descriptors, interaction descriptors, integration descriptors, adsorption isotherm data, and adsorption Henry's constant; wherein, the adsorption isotherm data is the adsorption amount of a first adsorbate by the target material or sample material at multiple pressure points at a reference temperature, and the first adsorbate is different from the adsorbate to be predicted; the adsorption Henry's constant includes the adsorption Henry's constant K of the target material or sample material for the adsorbate to be predicted at the predicted temperature. 𝐻 The K 𝐻 Obtained through molecular simulation, experimental testing, or artificial intelligence prediction.

[0020] The standardization process in step 4) includes at least one or more of the following: temperature condition alignment, pressure point alignment, unit unification, normalization, standardization, and maximum / minimum scaling.

[0021] Step 7) uses preset evaluation criteria, including one or more of the following: coefficient of determination, mean square error, root mean square error, mean absolute error, and mean absolute percentage error. Based on these preset evaluation criteria, the optimal physical constraint AI model is determined from one or more candidate physical constraint AI models.

[0022] The prediction results include one or more of the following: adsorption isotherms, adsorption capacity at a single pressure point, working capacity within a given pressure range, and separation performance indicators; the separation performance indicators include adsorption selectivity or selectivity calculated based on the ideal adsorption solution theory IAST; wherein, when an adsorption isotherm is output, the predicted operating parameters include multiple pressure points. And calculate them respectively .

[0023] Henry constant The adsorption Henry's constant of the target material at the temperature corresponding to the predicted operating conditions is used as an input parameter in the adsorption amount prediction.

[0024] A system for predicting the adsorption performance of porous materials based on physical constraint artificial intelligence includes a user terminal, a data processing and storage layer, and an artificial intelligence training and prediction layer.

[0025] The user terminal includes a data acquisition module, a judgment and processing module, a feedback confirmation module, and a display module, and is configured to execute steps 1) and 2) of the method of claim 1 and output the prediction result of step 8); the data processing and storage layer is configured to execute steps 3) to 5) of the method of claim 1; and the artificial intelligence training and prediction layer is configured to execute steps 6) to 8) of the method of claim 1.

[0026] The beneficial effects of this invention are as follows: Combining the practical characteristics of porous material descriptors, a novel method for predicting adsorption capacity with physical meaning is innovatively defined. This method can quickly obtain adsorption performance data of a target material within a given temperature and pressure range based on the material's structural characteristics, thereby assisting researchers in rapidly assessing the material's application potential and ultimately achieving accurate prediction, reducing material development costs, and improving development efficiency. Unlike commonly used descriptor types in existing technologies, this invention innovatively supports multiple descriptors as input, including but not limited to traditional structure-chemical descriptors and the recently emerging integrated descriptors, for training artificial intelligence models. This invention requires extremely little data (less than 1%), reducing the development cost of adsorption materials and improving development efficiency. Attached Figure Description

[0027] Figure 1This is a schematic diagram of the structure of a porous material adsorption data acquisition system based on physical constraint artificial intelligence provided in an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of a user terminal in a porous material adsorption data acquisition system based on physical constraint artificial intelligence provided in an embodiment of the present invention;

[0029] Figure 3 This is a schematic diagram of the data processing and storage layer in a porous material adsorption data acquisition system based on physical constraint artificial intelligence provided in an embodiment of the present invention;

[0030] Figure 4 This is a schematic diagram of the artificial intelligence training and prediction layer in a porous material adsorption data acquisition system based on physical constraint artificial intelligence provided in an embodiment of the present invention;

[0031] Figure 5 This is a flowchart of a method for acquiring adsorption data of porous materials based on physical constraint artificial intelligence, provided by an embodiment of the present invention;

[0032] Figure 6 The present invention provides an analysis of the prediction results of CO2 in ionic porous materials at room temperature using a physical constraint artificial intelligence model. (a) Comparison of adsorption prediction value and actual value, where the color bars represent pressure (unit: Pa), and comparison with the traditional neural network in (b) numerical accuracy and (c) isotherm trend accuracy.

[0033] Figure 7 This is a comparison chart of the predicted and actual values ​​of the propylene / propane selectivity of zeolite at room temperature (40 bar) using the physical constraint artificial intelligence model provided in this embodiment of the invention.

[0034] Figure 8 This is a comparison chart of the predicted and actual values ​​of the adsorption amount of hexane isomers in metal-organic framework materials at room temperature (3E+06 Pa) using the physical constraint artificial intelligence model provided in this embodiment of the invention. Detailed Implementation

[0035] This invention provides a method and system for acquiring adsorption data of porous materials based on physical constraint artificial intelligence, which can help researchers reduce material development costs and improve development efficiency. To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0036] In this invention, the target material specifically refers to an adsorption separation material with a porous structure. The target material includes, but is not limited to, metal-organic frameworks (MOFs), covalent organic frameworks (COFs), zeolites, porous carbons, porous organic cages (POCs), and other porous solid materials with similar adsorption mechanisms. These materials typically have well-defined pore size distributions and specific surface areas.

[0037] In this invention, material description data (or descriptor data) refers to feature vectors that can quantitatively or qualitatively characterize the physicochemical properties of a target material and are correlated with the material's adsorption performance. Specifically, the material description data includes one or more of the following categories:

[0038] Structural descriptors: such as pore size (PLD, LCD), specific surface area, pore volume, porosity, etc.

[0039] Chemical descriptors: such as atomic composition, functional group type, metal center type, etc.;

[0040] Integrated descriptor: refers to performance data obtained through experiments or simulations that can indirectly reflect the porosity characteristics of a material, such as adsorption isotherm data of the material under reference conditions (such as different pressure points) (e.g., methane adsorption data).

[0041] Physicochemical constants: such as Henry's constant, heat of adsorption, etc.

[0042] In this invention, the adsorption performance label refers to a target variable used to characterize the adsorption behavior of a target material for a specific adsorbate under specific operating parameters (such as temperature and pressure). During the model training phase, it serves as the true value for supervised learning; during the prediction phase, it is the output of the model. Specifically, the adsorption performance label includes, but is not limited to, the following forms:

[0043] Absolute adsorption capacity: refers to the amount of gas adsorbed per unit mass or unit volume of material at a given temperature and pressure point (e.g., mmol / g, cm³ / g, mg / g, etc.).

[0044] Adsorption isotherms: These are data sequences showing how the amount of adsorption changes with pressure at a constant temperature. In model training, this is represented as a set of adsorption amount labels corresponding to different pressure points.

[0045] Separation performance indicators: such as adsorption selectivity, which is the ratio of the amount of two or more gases adsorbed or the ideal adsorption solution theoretical (IAST) selectivity calculated based on the amount of adsorption.

[0046] The data source for the adsorption performance label can be experimental test data or high-precision molecular simulation data.

[0047] like Figure 1 As shown in the figure, this embodiment of the invention provides an artificial intelligence service system, including: a user terminal 101, a data processing and storage layer 102, and an artificial intelligence training and prediction layer 103; wherein, the user terminal 101 is connected to both the data processing and storage layer 102 and the artificial intelligence training and prediction layer 103; the data processing and storage layer 102 is connected to both the user terminal 101 and the artificial intelligence training and prediction layer 103; the artificial intelligence training and prediction layer 103 is connected to both the user terminal 101 and the data processing and storage layer 102; the user terminal 101 is used for data processing and prediction. The data is collected, processed, and interacted with, and the information is sent to the data processing and storage layer 102, as well as received and displayed by the artificial intelligence training and prediction layer 103. The data processing and storage layer 102 is used for data storage, data extraction, data standardization, dividing the training set and test set, and transmitting the data to the artificial intelligence training and prediction layer 103. The artificial intelligence training and prediction layer 103 is used for training artificial intelligence models, evaluating training effects, predicting one or more adsorption properties of the target material, and transmitting the data to the user terminal 101 for display and output.

[0048] like Figure 2 As shown in one embodiment of the present invention, the user terminal 101 includes: a data acquisition module 1011, a judgment and processing module 1012, a feedback confirmation module 1013, and a display module 1014; wherein, the data acquisition module 1011 is used to collect material descriptor data from the user and determine one or more adsorption performance labels to be predicted; the judgment and processing module 1012 is used to determine whether the data meets the conditions, i.e., whether it can be smoothly connected into a curve, and whether there are missing values ​​and outliers; the feedback confirmation module 1013 is used to provide data supplementation and modification suggestions to the user or transmit data that meets the conditions to the data processing and storage layer 102; the display unit 1014 is used to receive data from the artificial intelligence training and prediction layer 103 and display the data.

[0049] Preferably, the data storage and processing layer 102, such as Figure 3As shown, it includes: a data storage module 1021, a data extraction module 1022, a data standardization module 1023, and a training set-test set partitioning module 1024; wherein, the data storage module 1021 is used to store descriptor and label data collected from the user terminal, as well as a large amount of existing data used for training the artificial intelligence model; the data extraction module 1022 is used to extract descriptor and label data of the same type of material under the same temperature conditions from the database based on the descriptor and label information collected from the user; the data standardization module 1023 is used to perform various standardization operations on the extracted adsorption data, such as unifying dimensions, normalization, and maximum / minimum scaling; the training set-test set partitioning module 1024 partitions the training set and test set according to a preset sampling strategy and passes them into the artificial intelligence training and prediction layer 103.

[0050] Preferably, the artificial intelligence training and prediction layer 103, such as Figure 4 As shown, it includes: a model training module 1031, a model evaluation module 1032, a performance prediction module 1033, and a data output module 1034; wherein, the model training module 1031 is used to train multiple physical constraint artificial intelligence models of different sizes based on the input descriptor and label data; the model evaluation module 1032 is used to evaluate the training results of multiple physical constraint artificial intelligence models of different sizes, select the optimal model, and input it into the performance prediction module 1033; the performance prediction module 1033 is used to predict one or more adsorption properties of the target material; the data output module 1034 is used to package the prediction results and transmit them to the user terminal for display and output.

[0051] In this patent, the training sample data includes various types of adsorption materials, and the dataset must include other relevant descriptors of these adsorption materials, including but not limited to structural descriptors, chemical descriptors, interaction descriptors, and integrated descriptors based on gas adsorption isotherms.

[0052] When building a model, the first step is to determine the dataset to be used. For example, a large class of adsorbent materials and their adsorption performance data can be selected as training data. During training, the input variables are the numerical values ​​corresponding to the descriptors of these adsorbent materials, while the output variables employ physical constraints, which are implemented using the following formula:

[0053]

[0054] The input values ​​for the artificial intelligence model are x1, x2, x3, x4...x nThen, based on the specific functional form of f(x), the adsorption amount data Q is obtained through the above formula. During the calculation process, the parameters in the function f(x) are obtained by minimizing the loss function (the difference between the true value and the calculated value Q).

[0055] In one example, P represents the external pressure; K H To obtain the adsorption Henry's constant, K can be obtained through computational simulation, experimental testing, and artificial intelligence prediction. H F(x1, x2, x3, x4...x) n These are the core parameters of the model, which can be obtained through methods such as numerical fitting and artificial intelligence prediction. The following is one of the reference methods:

[0056]

[0057] Where x1, x2, x3, x4...x n The descriptors include, but are not limited to, structural descriptors, chemical descriptors, interaction descriptors, and integrated descriptors based on gas adsorption isotherms. It is worth noting that the core parameters shown in the above model are obtained only through linear fitting; in practice, various artificial intelligence models can be used, including multiple fitting formulas, all of which should be included within the scope of this patent protection.

[0058] For example, the function form in f(x) can also be changed to a polynomial, exponential function, logarithmic function, power function, etc.

[0059] Depending on the actual programming needs, it can be transformed into:

[0060]

[0061]

[0062] In the form of Q, Q', and Q'', the adsorption amounts are denoted as Q, Q', and Q''.

[0063] In the above embodiments, the data storage module 1021 stores more than 150,000 material structures, as well as corresponding descriptor and tag data, involving structural data such as pore size and specific surface area (approximately 900,000 records), Henry's constant (approximately 50,000 records), CO2 (approximately 710,000 records), N2 (approximately 380,000 records), CH4 (approximately 850,000 records), H2 (approximately 280,000 records), n / i-butane (approximately 24,000 records), C6H... 14 (Approximately 8,000 entries) The types of materials stored include metal-organic frameworks (MOFs, approximately 150,000), covalent organic frameworks (811), zeolites (216), and amorphous porous carbon materials (614).

[0064] like Figure 5 As shown, this embodiment of the invention provides a method for rapidly predicting the adsorption performance of materials using the physical constraint artificial intelligence service system provided in any embodiment of the invention, comprising: step 501, collecting material description data using the data acquisition module in the user terminal and determining prediction labels; step 502, determining whether the data meets the conditions using the judgment processing module in the user terminal; step 503, deciding whether the user needs to supplement data based on the judgment result, and feeding back to the user or transmitting to the data storage and processing layer through the feedback confirmation module; step 504, storing the description data collected from the user terminal using the data storage module in the data storage and processing layer; step 505, extracting descriptor data and adsorption performance labels of similar materials under the same conditions from the database using the data extraction module in the data storage and processing layer; step 506, utilizing the data storage and processing layer... The data standardization module in the AI ​​training and prediction layer performs data standardization to meet the model input requirements; step 507: using the training set-test set partitioning module, the training set and test set are partitioned based on a custom sampling strategy or randomly, and then fed into the AI ​​training and prediction layer; step 508: using the model training module in the AI ​​training and prediction layer, the model is trained using a preset physical constraint neural network based on the input description data and label data; step 509: using the model evaluation module in the AI ​​training and prediction layer, the training results are evaluated, and if the requirements are met, they are fed into the performance prediction module; step 510: using the performance prediction module in the AI ​​training and prediction layer, the various adsorption and separation properties of the target material are predicted; step 511: using the data output module in the AI ​​training and prediction layer, the prediction results are packaged and transmitted to the user terminal for display and output.

[0065] This invention defines a novel physically constrained artificial intelligence model, deployed in the model training module of the AI ​​training and prediction layers. Training this physically constrained AI model requires only a small amount of sample data (less than 1%) to obtain a prediction model with high accuracy, good extrapolation performance, and strong generalization ability. The key innovation of this invention lies in its novel definition of an adsorption amount prediction method with physical meaning. The physical constraint is achieved through the following formula:

[0066]

[0067] The embodiments of this invention have at least the following beneficial effects: 1. It does not limit the source of input data; it can come from experiments or simulations. Users only need to input descriptor data for one material to obtain one or more adsorption properties, eliminating the need for experimental testing or theoretical simulation of multiple adsorption properties of the material, greatly saving time and cost. 2. This invention innovatively defines a novel adsorption prediction method with physical constraints. This method can quickly obtain adsorption performance data of a target material within a given temperature and pressure range based on the material's structural characteristics, thereby assisting researchers in quickly assessing the material's application potential. 3. This invention innovatively supports multiple descriptors as input, including but not limited to traditional structure-chemical descriptors and the recently emerging integrated descriptors, for training artificial intelligence models. The data requirement of this invention is extremely small (less than 1%), reducing the R&D cost of adsorption materials and improving R&D efficiency.

[0068] Example 1

[0069] Taking ionic porous materials as an example, the user inputs the adsorption isotherm data of CH4 at room temperature and the calculated Henry's constant for CO2 as descriptors. Simultaneously, the user is required to predict the CO2 adsorption isotherm of the material at room temperature (i.e., the predicted operating condition parameter) (i.e., the adsorption performance label to be predicted). First, the user inputs the adsorption isotherms of an ionic porous material (SIFSIX-2-Zn-i, i.e., the target material) at 298K at the following temperatures: 1e+00, 1e+01, 1e+02, 1e+03, 5e+03, 1e+04, 3e+04, 5e+04, 7e+04, 1e+05, 3e+05, 5e+05, 7e+05, 1e+06, 3e+06, 5e+06, 7e+06, 1e+07. The CH4 adsorption amounts at the Pa pressure point were 2.92e-05, 0.000236738, 0.002382246, 0.023955691, 0.118843062, 0.228694481, 0.642318282, 1.008761837, 1.314171014, 1.715710713, 3.293993875, 3.969012841, 4.345035479, 4.686648514, 5.437284372, 5.655576693, 5.787276834, and 5.877590205 mmol / g, respectively, along with the Henry's constant for CO2 (0.00126862). The input CH4 adsorption isotherm data (mol / kg / Pa), i.e., material descriptor data, aims to predict the CO2 adsorption isotherm of the material at room temperature. The input CH4 adsorption isotherm data has no missing or outlier values, and the goodness of fit is greater than 0.8, meeting the requirements of subsequent processes. Therefore, the data is transferred to the data storage and processing layer. The structure of ionic porous materials consists of three parts: anionic clusters, organic chains, and a metal center. All ionic porous materials (i.e., sample materials) in the database were extracted.

[0070] A total of 800 types were obtained by arranging 12 anionic clusters, 13 organic chains, and 6 metal centers. The metal centers included: Zn²⁺, Cu²⁺, Co²⁺, Ni²⁺, Al³⁺, and Zr. 4 ⁺. Anionic clusters include: SiF6²⁻ (SIFSIX), TiF6²⁻ (TIFSIX), GeF6²⁻ (GEFSIX), NbF6⁻, TaF6⁻, PF6⁻, BF4⁻, SO4²⁻, NO3⁻, Cl⁻, [PW 12 O 40 ]³⁻、[Fe(CN)6] 4⁻. Organic chains / ligands include: pyrazine, 4,4-bipyridine, 1,2-bis(4-pyridyl)ethylene (bpee), DABCO, imidazole, benzimidazole, terephthalic acid (H2BDC), trimesic acid (H3BTC), fumaric acid (H2fum), adipic acid (H2adip), isonicotinic acid (Hina), 4-hydroxybenzoic acid, and glycine.

[0071] The data includes isothermal adsorption data of CH4 at room temperature, the Henry's constant for CO2 (descriptor data for the sample material), the amount of CO2 adsorbed by the material at room temperature (adsorption performance label data under the same predicted operating conditions), and the corresponding pressure data. It is worth noting that the descriptors should be obtained under the same reference, such as ensuring that the temperature and pressure points of all descriptive data are consistent.

[0072] In this embodiment, the specific form of physical constraint is as follows:

[0073]

[0074] CH4 adsorption isotherm data, CO2 Henry's constant, CO2 adsorption amount and corresponding pressure are input into the artificial intelligence training and prediction layer. Among them, the CH4 adsorption isotherm data (adsorption amount at pressure points of 1e+00, 1e+01, 1e+02, 1e+03, 5e+03, 1e+04, 3e+04, 5e+04, 7e+04, 1e+05, 3e+05, 5e+05, 7e+05, 1e+06, 3e+06, 5e+06, 7e+06, 1e+07 Pa at 298K) and the CO2 Henry's constant are used as inputs to the f function, i.e., X1, X2, X3, X4...Xn in the formula;

[0075] The f function is a linear function:

[0076]

[0077] The amount of CO2 adsorbed is used as a label, and the corresponding CO2 pressure data (P) and CO2 Henry's constant (K) are also included. HThe values ​​were directly substituted into the physical constraint formula. Four different sized physical constraint neural networks were trained based on the input data. Each neural network consisted of three fully connected layers, with ReLU selected as the activation function between each layer. The number of neurons in the middle layers was 16, 32, 64, and 128, respectively. MSE was chosen as the loss function, and the optimal model was selected based on the principle of minimum coefficient of determination. In this example, the optimal model was the physical constraint neural network model with 32 neurons in the middle layers. This model was then fed into the adsorption performance prediction unit to predict the CO2 adsorption isotherms of the material at room temperature. The isotherms can be obtained by connecting discrete points. Therefore, the model predicted 1e+00, 1e+01, 1e+02, 1e+03, 5e+03, 1e+04, 3e+04, 5e+04, 7e+04, 1e+05, and 3e+05. The CO2 adsorption amounts at pressure points of 5e+05, 7e+05, 1e+06, 3e+06, 5e+06, and 7e+06 Pa are calculated, with output results of 0.001268401, 0.012664281, 0.12470359, 1.0814406, 3.4003654, 4.6455393, 6.145914, 6.5703187, 6.7706966, 6.9291887, 7.191038, 7.245801, 7.269527, 7.287424, 7.3154383, 7.3210673, and 7.323483 mmol / g. These prediction results are packaged and transmitted to the user terminal, where discrete points are connected, and the data is displayed and output. It is worth noting that the evaluation criteria for the training effect of artificial intelligence models are not fixed and are not limited to the principle of minimizing the coefficient of determination in the demonstration examples. Figure 6 This paper demonstrates the analysis of the prediction results of a physically constrained artificial intelligence model for CO2 at room temperature. It can be seen that the physically constrained neural network model can provide relatively accurate prediction results, with a determination coefficient reaching 0.907. Figure 6 a). Compared to a model without physical constraints (where the rest of the model remains unchanged, i.e., three fully connected layers), under the same input and output variables, the physically constrained neural network model will achieve higher numerical accuracy ( Figure 6 b) and the trend rate of isotherms Figure 6 c) Increased from 18.3% and 91.8% to 100% and 100%, respectively. It is worth noting that there are various methods for obtaining the Henry's constant, and it is not limited to those listed in the examples; the evaluation criteria for the training effect of artificial intelligence models are also not fixed, and are not limited to the principle of the minimum determination coefficient in the examples.

[0078] Example 2

[0079] Taking zeolite materials as an example, the user inputs the material's structural parameters and the calculated propylene / propane Henry's constant as a descriptor, and requests a prediction of the material's adsorption capacity for propylene / propane at room temperature (40 bar). First, the user inputs the structural parameters of a zeolite material (ABW_0) through the user terminal, including PLD: 2.8804, LCD: 3.60282, VF: 0.208281, GASA: 0, VASA: 0, and the propylene / propane Henry's constant (propylene: 22.0838, propane: 15.4254 mol / kg / Pa), hoping to predict the material's propylene / propane adsorption capacity at room temperature (40 bar). The above input data contains no missing or outliers, meeting the requirements of subsequent processes. Therefore, the data is transferred to the data storage and processing layer. Simultaneously, the structural parameters of all zeolite materials (170 in total, each with a different topology) in the database are extracted, along with the Henry's constant for propylene gas adsorption, the Henry's constant for propylene / propane adsorption, the corresponding adsorption amount label, the adsorption amount, and the pressure data at a temperature (T=298K). It is worth noting that the descriptors should be obtained under the same reference, such as ensuring that all descriptive data have consistent temperature and pressure points.

[0080] In this embodiment, the specific form of physical constraint is as follows:

[0081]

[0082] The structural parameters, propylene / propane Henry's constant, adsorption amount and corresponding pressure are input into the artificial intelligence training and prediction layer. The structural parameters (PLD, LCD, VF, GASA, VASA) and propylene / propane Henry's constant are used as inputs to the f function, i.e. X1, X2, X3, X4...Xn in the formula.

[0083] The F function is a linear function:

[0084] .

[0085] Propylene / propane adsorption capacity is used as a label, and the corresponding pressure data and the Henry's constant for propylene / propane are directly substituted into the physical constraint formula. Four different sized physical constraint neural networks are trained based on the input data. Each neural network consists of three fully connected layers, with ReLU selected as the activation function between each layer. The number of neurons in the middle layers is 16, 32, 64, and 128, respectively. MSE is selected as the loss function, and the optimal model is selected based on the principle of minimum coefficient of determination. In this example, the optimal model is the physical constraint neural network model with 16 neurons in the middle layer, which is then fed into the adsorption performance prediction unit. For the propylene / propane adsorption capacity of the material at room temperature (40 bar), the model outputs propylene: 3.0350053 mmol / g and propane: 2.7896636 mmol / g. The prediction results are packaged and transmitted to the user terminal for display and output. It is worth noting that the evaluation criteria for the training effect of the artificial intelligence model are not fixed and are not limited to the principle of minimum coefficient of determination in the example. Figure 7 The image shows a comparison between the predicted and actual values ​​of the propylene / propane selectivity of zeolite at room temperature (40 bar) using a physically constrained artificial intelligence model. It can be seen that the physically constrained neural network model can provide relatively accurate predictions, with a comprehensive coefficient of determination of 0.904 for propylene / propane.

[0086] Example 3

[0087] Taking metal-organic framework materials as an example, users input the adsorption isotherm data of Ar at 77K and the calculated Henry's constant of hexane isomers as descriptors, and at the same time, they are required to predict the adsorption amount of various hexane isomers of the material at room temperature 3E+06 Pa. First, the user inputs the following values ​​for a metal-organic framework material (HOFSOL) at 77K via their user terminal: 0.001, 0.01, 0.1, 1, 10, 100, 1000, 2000, 4000, 6000, 8000, 10000, 12000, 14000, 16000, 18000, 20000, 24000, 28000, 30000, 32000, 36000, 40000, 50000, 60000, 70000, 80000, 90000, 100000. The Ar adsorption amounts at the Pa pressure points were 0.019759, 0.20798, 2.08258, 22.9854, 358.272, 425.621, 466.899, 482.071, 497.764, 499.399, 511.313, 517.403, 518.539, 515.439, 519.754, 526.402, 525.768, 529.095, 530.789, 525.644, 530.65, 536.45, 536.432, 539.853, 547.02, 545.128, 549.415, 549.748, and 547.261 cm⁻¹, respectively. 3 The data, along with the Henry's constants for hexane isomers (22DMB: 0.169159, 23DMB: 0.547803, 2MP: 0.78191, 3MP: 0.830454, nHEX: 1.0239 mol / kg / Pa), were used to predict the adsorption capacity of the hexane isomers of this material at room temperature (3E+06 Pa). The input Ar adsorption isotherms did not contain missing or outlier values, and the goodness of fit was greater than 0.8, meeting the requirements of subsequent procedures. Therefore, the data was transferred to the data storage and processing layer. Simultaneously, isothermal adsorption data for Ar at 77 K was extracted from the database for all metal-organic framework materials (a total of 737), along with the Henry's constants, adsorption capacities, and corresponding pressure data for the hexane isomers at room temperature. It is worth noting that the descriptors should be obtained under the same baseline, ensuring that all descriptive data have consistent temperature and pressure points.

[0088] In this embodiment, the specific form of physical constraint is as follows:

[0089]

[0090] The f function is a linear function:

[0091]

[0092] Ar adsorption isotherm data, the Henry's constant for hexane isomers, adsorption capacity, and corresponding pressure are input into the AI ​​training and prediction layer. The Ar adsorption isotherm data (at 77 K) are 0.001, 0.01, 0.1, 1, 10, 100, 1000, 2000, 4000, 6000, 8000, 10000, 12000, 14000, 16000, 18000, 20000, 24000, 28000, 3000... The adsorption amounts at pressure points of 0, 32000, 36000, 40000, 50000, 60000, 70000, 80000, 90000, and 100000 Pa are used as inputs to the f-function, i.e., X1, X2, X3, X4...Xn in the formula; the adsorption amounts of hexane isomers are used as labels, and the pressure data corresponding to the labels and the hexane isomer Henry's constant are directly substituted into the physical constraint formula. Four physically constrained neural networks of different sizes were trained on the input data. Each neural network consisted of three fully connected layers, with ReLU chosen as the activation function between each layer. The number of neurons in the middle layers was 16, 32, 64, and 128, respectively. MSE was chosen as the loss function, and the optimal model was selected based on the principle of minimizing the coefficient of determination. In this example, the optimal model was the physically constrained neural network model with 64 neurons in the middle layers. This model was then fed into the adsorption performance prediction unit to calculate the adsorption capacity of hexane isomers at room temperature (3E+06 Pa). The model outputs were 22 DMB: 3.5579917, 23 DMB: 4.4073415, 2 MP: 4.722728, 3 MP: 4.8816385, and nHEX: 4.8795385 mmol / g, respectively. The prediction results were packaged and transmitted to the user terminal for display and output. It is worth noting that the evaluation criteria for the training effect of the artificial intelligence model are not fixed and are not limited to the principle of minimizing the coefficient of determination in the example. Figure 8 The image shows a comparison between the predicted and actual values ​​of hexane isomer adsorption in metal-organic framework materials at room temperature (3E+06Pa) using a physically constrained artificial intelligence model. It can be seen that the physically constrained neural network model can provide relatively accurate prediction results, with determination coefficients of 0.977, 0.980, 0.965, 0.990, and 0.990 for 22DMB, 23DMB, 2MP, 3MP, and nHEX, respectively.

Claims

1. A method for predicting the adsorption performance of porous materials based on physical constraint artificial intelligence, characterized in that, include: 1) Obtain the material descriptor data of the target material, and determine the adsorption performance label to be predicted and the prediction operating condition parameters, wherein the prediction operating condition parameters include at least one of temperature, pressure point and type of adsorbent material; 2) Determine whether the material description data meets the preset input conditions; if not, output a data supplement or modification prompt and receive the supplemented material description data; 3) Extract sample materials of the same type as the target material from the database, obtain the descriptor data of the sample materials, and the adsorption performance label data of the sample materials under the same conditions as the predicted operating parameters; 4) Perform consistency processing on the data obtained in step 3) to ensure that the descriptor data of each sample material meets the model input requirements; 5) Construct training and testing sets based on the descriptor data and adsorption performance label data of the sample materials; 6) Training is performed using a physical constraint artificial intelligence model, which is trained through parameterizable functions. Predicted adsorption amount by combining physical constraint relationships and according to Compared with the actual adsorption amount The error between them is used to construct a loss function to optimize the loss function. The parameters, wherein the physical constraint relationship is: in, As a pressure point, Henry's constant, For descriptor vectors; 7) Evaluate the training results to obtain a physical constraint artificial intelligence model that meets the preset evaluation criteria; 8) Use the physical constraint artificial intelligence model to predict the adsorption performance of the target material under the predicted operating parameters, and output the prediction results.

2. The method according to claim 1, characterized in that, The parameterizable function It can be a linear function or a nonlinear function; The linear functions include at least: The The parameters to be trained, For descriptor components.

3. The method according to claim 1, characterized in that, The physical constraint artificial intelligence model is a physical constraint neural network; the physical constraint neural network uses descriptor vectors. Input and output for calculation core parameters Or output used for characterization The function parameters.

4. The method according to claim 1, characterized in that, The loss function in step 6) is Compared with the actual adsorption amount The error function includes one or more of the following: mean square error, mean absolute error, and root mean square error.

5. The method according to claim 1, characterized in that, The similar materials mentioned in step 3) refer to the set of materials in the parent material category to which the target material belongs.

6. The method according to claim 1, characterized in that, The material description data includes one or more of the following: structural descriptors, chemical descriptors, interaction descriptors, integration descriptors, adsorption isotherm data, and Henry's constant.

7. The method according to claim 1, characterized in that, The standardization process in step 4) includes at least one or more of the following: temperature condition alignment, pressure point alignment, unit unification, normalization, standardization, and maximum / minimum scaling.

8. The method according to claim 1, characterized in that, The preset evaluation criteria in step 7) include one or more of the following: coefficient of determination, mean square error, root mean square error, mean absolute error, and mean absolute percentage error.

9. The method according to claim 1, characterized in that, The prediction results include one or more of the following: adsorption isotherm, adsorption capacity at a single pressure point, working capacity within a given pressure range, and separation performance indicators; wherein, when an adsorption isotherm is output, the predicted operating parameters include multiple pressure points. And calculate them respectively ; Henry constant The adsorption Henry's constant of the target material at the temperature corresponding to the predicted operating conditions is used as an input parameter in the adsorption amount prediction.

10. A system for predicting the adsorption performance of porous materials based on physical constraint artificial intelligence, characterized in that, This includes user terminals, data processing and storage layers, and artificial intelligence training and prediction layers; The data processing and storage layer is configured to perform steps 3) to 5) of the method of claim 1. The artificial intelligence training and prediction layer is configured to perform steps 6) to 8) of the method of claim 1.

Images