Ionic working medium design method and related device

Abstract

The invention discloses a design method of an ionic working medium and a related device. The design method comprises the following steps: acquiring a design target of a target ionic working medium, and constructing and training a prediction model, a property-structure pre-training embedding model and an isotropic graph diffusion generation model; combining the trained prediction model, the property-structure pre-training embedding model and an isovariant graph diffusion generation model to construct an embedded graph diffusion generation model; and designing by utilizing an embedded graph diffusion generation model to obtain a target ionic working medium. According to the method, the design period can be greatly shortened, the design cost is reduced, the design success rate is improved, and powerful technical support is provided for research, development and application of the ionic working medium.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided molecular design technology, and relates to a design method and related apparatus for ionic working fluids. Background Technology

[0002] Working fluids, as carriers of energy conversion and transfer, are a core component of energy conservation and emission reduction technology innovation. Developing high-performance working fluids for specific application scenarios has significant scientific and engineering value for optimizing system efficiency, reducing energy consumption, and minimizing environmental pollution. Currently, working fluid design methods mainly include trial and error, predictive screening, and generative design. Trial and error methods consider the relationship between structure and properties based on existing data, explore influencing factors, and then predict and statistically analyze novel candidate working fluids with potentially superior performance. Trial and error is a traditional method for working fluid design, but it suffers from drawbacks such as slow speed, high cost, and limited screening scope. Predictive screening establishes property prediction models to calculate and predict the properties of collected candidates, and selects working fluids based on the predicted values. Predictive models mainly include quantum chemical algorithms (represented by COSMO-RS), equations of state, group contribution methods, molecular dynamics simulations, and neural networks. This method solves the problems of slow speed and high cost associated with trial and error, but the final result can only be derived from the collected candidates, and the problem of a limited screening scope remains unresolved. Generative design models based on artificial intelligence can completely solve the problems of slow speed, high cost, and limited screening range. Generative design models can directly output molecular structures that meet given design requirements, with a screening range approaching infinite.

[0003] Ionic working fluids have enormous application potential in the energy field, but their numerous thermophysical properties determine their energy consumption, efficiency, and applicability. The design of ionic working fluids requires consideration of the complex nonlinear relationships between various thermophysical properties and the performance of corresponding energy systems. This not only poses a greater challenge to generative models but also introduces the problem of increasingly sparse data. To achieve efficient and energy-saving operation of ionic working fluids, in addition to ensuring that the generated results simultaneously meet multiple objectives, the generative model also needs to implement property-conditional generation. Existing multi-objective property-optimized generation cannot incorporate these nonlinear relationships into the optimization objectives, leading to several problems: because property-conditional generation models require simultaneous input of molecular structure information and properties during training, they cannot be trained on datasets with missing values; existing technologies use SMILES as molecular descriptors, which is a compromise to the insufficient learning ability of variational autoencoder models on small datasets. SMILES's overly abstract representation of molecular structure is not suitable for training artificial intelligence models. To achieve efficient and high-quality generation, it is necessary to use descriptors with comprehensive and direct information, such as molecular graphs. The introduction of molecular diagrams brings new problems. Because the distance between the cations and anions in ionic working fluids is much greater than the bond length, and there is significant relative movement, ionic working fluids cannot be viewed as a single molecule. The generation model for ionic working fluids also requires assigning separate model structures to anions and cations. However, this leads to a serious conflict between the structural representation of the ionic working fluid and the conditional generation process: the properties of ionic working fluids depend on the combined contributions of their cations and anions. If the property conditions are directly used as input to the ion generation model, during training, the anion generation model may receive inconsistent inputs from ionic working fluids containing the same anion but different cations, yet the loss function requires it to output the same ionic structure, resulting in an input / output conflict. The model cannot be trained and cannot output reliable design results. Summary of the Invention

[0004] The purpose of this invention is to provide a design method and related apparatus for ionic working fluids, thereby solving the problems in the prior art.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A design method for ionic working fluids includes:

[0007] The design objectives of the target ionic working fluid are obtained, and a prediction model, a property-structure pre-trained embedding model, and an isomorphic graph diffusion generation model are constructed and trained.

[0008] The trained prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model are combined to construct the embedding graph diffusion generation model;

[0009] The target ionic working fluid was designed using an embedded graph diffusion generation model.

[0010] Furthermore, the prediction model accepts ionic structures as input and outputs predicted property values, including molecular layers, embedding layers, AEGNN layers, averaging layers, and FC layers.

[0011] Furthermore, the prediction model uses the following method to predict property values:

[0012] The ionic structure is input into the molecular layer, and the embedding layer embeds node features based on the molecular graph input from the molecular layer, and unifies the dimensions of the node features with those of the AEGNN layer.

[0013] The AEGNN layer extracts structural information from the ion layer by layer, and the averaging layer averages each dimension of the node features and reshapes the matrix format of the node features.

[0014] The node features after averaging in the average layer are input into the FC layer, and the FC layer summarizes the information extracted from the cation and anion layers to output the calculated property values.

[0015] Furthermore, the property-structure pre-trained embedding model includes a property encoder and a structure encoder; the property encoder takes a vector composed of multiple properties of the ionic working fluid as input and outputs the property embedding; the structure encoder includes anion encoder and cation encoder, which take the ionic structure as input and output the ionic structure embedding.

[0016] Furthermore, the isomorphic graph diffusion generation model gradually denoises and restores the sample by removing prior distribution noise that has been added up to the maximum step size, based on the generation conditions; the input includes atom type, charge, and conditional dimension, and the output is an atom type vector and atom coordinates with charge information.

[0017] Furthermore, the prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model are trained sequentially:

[0018] The prediction model was trained using data from the ILThermo dataset.

[0019] After predicting missing values ​​using a predictive model and completing the ILThermo dataset, the property-structure pre-trained embedding model is trained by combining experimental and predicted values.

[0020] The property embeddings of the collected ionic working fluids are obtained by using a property-structure pre-trained embedding model, and the property embeddings and the ionic working fluid structure are used to train an isomorphic graph diffusion generation model.

[0021] Furthermore, the method for interpreting the output of the embedding graph diffusion generation model is as follows:

[0022] Examine the one-hot encoding of the atom type in the node features, find the position with the largest value, and the corresponding atom type is the interpretation result;

[0023] When interpreting chemical bonds, the type of bond between all atoms is determined by the distance between them.

[0024] A design system for an ionic working fluid includes:

[0025] The first modeling module is used to obtain the design target of the target ionic working fluid, construct a prediction model, a property-structure pre-trained embedding model and an isomorphic graph diffusion generation model and train them.

[0026] The second modeling module is used to combine the trained prediction model, the property-structure pre-trained embedding model and the equivariant graph diffusion generation model to construct the embedding graph diffusion generation model.

[0027] The design module is used to design using an embedded graph diffusion generation model to obtain the target ionic working fluid.

[0028] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method.

[0029] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] This invention provides a design method for ionic working fluids. By obtaining the design target of the target ionic working fluid, a prediction model, a property-structure pre-trained embedding model, and an isomorphic graph diffusion generation model are constructed and trained. A high-precision prediction model is used, and this model is used to complete missing data in the dataset, solving the problem of training on datasets with missing values, enriching the training dataset, and improving data completeness and quality. The property-structure pre-trained embedding model based on contrastive learning provides embedding vectors for ions derived from the properties of the ionic working fluid, establishing a correlation between properties and ionic structure, and providing generation conditions for the isomorphic graph diffusion generation model. The isomorphic graph diffusion generation model, with stronger generation capabilities, replaces the variational autoencoder-type model, addressing the challenges of insufficient structural data and feature extraction, and can more accurately capture the structural information of the ionic working fluid, thereby generating a more reasonable and practically suitable ionic working fluid structure. By combining a trained predictive model, a property-structure pre-trained embedding model, and an isovariant graph diffusion generation model, an embedding graph diffusion generation model is constructed, forming a collaborative whole. The predictive model provides data support, the property-structure pre-trained embedding model establishes the correlation between properties and structures and provides generation conditions, and the isovariant graph diffusion generation model is responsible for generating reasonable structures. The close cooperation among these stages improves the efficiency and accuracy of the entire design process. Using the embedding graph diffusion generation model, the target ionic working fluid is obtained. This invention comprehensively considers factors such as data quality, the correlation between properties and structures, and the performance of the generation model, making the design process more scientific, rational, and efficient. Compared with traditional trial-and-error or predictive screening design methods, this invention can significantly shorten the design cycle, reduce design costs, and increase the design success rate, providing strong technical support for the research and application of ionic working fluids. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the design method for the ionic working fluid of the present invention.

[0034] Figure 2 This is a diagram illustrating the structure and workflow of the prediction model of this invention.

[0035] Figure 3 This is a structural diagram of the property-structure pre-trained embedding model of the present invention.

[0036] Figure 4This is a flowchart illustrating the workflow of the isovariant graph diffusion generation model of the present invention.

[0037] Figure 5 This is a schematic diagram of the structure and process of the carbon dioxide capture device in Embodiment 1 of the present invention.

[0038] Figure 6 A comparison chart of energy consumption of 4-ethyl-1-propylpyridine tetrafluoroborate and [EMIM][TRIFLATE] designed for Example 1 of the present invention.

[0039] Figure 7 This is a schematic diagram of the design system structure for an ionic working fluid according to a preferred embodiment of the present invention.

[0040] Figure 8 This is a schematic diagram of the electronic device structure according to a preferred embodiment of the present invention. Detailed Implementation

[0041] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0042] Obviously, the described embodiments are only some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0043] It should be noted that the terminals involved in the embodiments of this application may include, but are not limited to, mobile phones, personal digital assistants (PDAs), wireless handheld devices, tablet computers, personal computers (PCs), MP3 players, MP4 players, wearable devices (e.g., smart glasses, smartwatches, smart bracelets, etc.), smart home devices, and other smart devices.

[0044] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0045] The present invention will now be described in further detail with reference to the accompanying drawings:

[0046] See Figure 1 This invention provides a design method for ionic working fluids, specifically including the following steps:

[0047] Step 1: Obtain the design target of the target ionic working fluid, construct the prediction model, the property-structure pre-trained embedding model and the isomorphic graph diffusion generation model and train them.

[0048] 1. Prediction Model

[0049] The prediction model is responsible for filling in missing values ​​in the dataset with predicted values. It takes the ion structure as input and outputs predicted property values. For example... Figure 2 As shown, the model includes a molecular layer, an embedding layer, an AEGNN layer, an averaging layer, and an FC layer. The ionic structure is input into the molecular layer. The embedding layer embeds node features based on the molecular diagram input from the molecular layer and unifies the dimensions of the node features with those of the AEGNN layer. The AEGNN layer extracts structural information from the ion layer by layer. The averaging layer averages each dimension of the node features and reshapes the matrix format of the node features. The averaged node features are then input into the FC layer, which summarizes the information extracted from the anions and cations to output the calculated property values. The prediction model uses two sets of models to process the structural information of anions and cations respectively, and then integrates them at the end. The AEGNN layer can fully learn the structural information of ionic liquids. The smallest unit of the AEGNN model is an atom and a chemical bond; therefore, the prediction model achieves high accuracy while maintaining strong versatility.

[0050] 2. Property-Structure Pre-trained Embedded Model

[0051] like Figure 3 As shown, the property-structure pre-trained embedding model includes a property encoder and a structure encoder. The property encoder takes a vector composed of multiple properties of the ionic liquid as input and outputs the property embedding. The structure encoder is divided into anion encoder and cation encoder. Its structure and the input it receives are similar to the prediction model, but its output is the embedding of the ionic structure. By concatenating the embeddings of anions and cations, the embedding of the ionic working fluid can be obtained. Conversely, by segmenting the embedding of the ionic working fluid according to the dimensions of anions and cations, the embeddings of anions and cations can be obtained.

[0052] Since the properties of the ionic working fluid are continuous values ​​rather than the 0 / 1 variables faced by CLIP, the cross-entropy loss used by CLIP for cosine similarity cannot be applied for training. Therefore, to ensure the uniqueness between the structural embeddings and the attribute embeddings, an additional constraint needs to be added to the loss function to prevent mode collapse, similar to that in adversarial generative networks. When mode collapse occurs, the property-structure pre-trained embedding model will generate the same embeddings for all properties and structures. The loss function of the property-structure pre-trained embedding model is shown in Equation (2-1). The first part of the loss function is the similarity loss, which minimizes the embedding e of the ionic working fluid. IL and attribute embedding e p The difference between them. The second part is the distribution loss, which defines μ as the variance between multiple ions within a training batch, and constrains the variance of multiple embeddings in a batch to 0.3 by setting the hyperparameter a = 0.3, thus preserving the difference between embeddings.

[0053] loss=∑(e IL -e p ) 2 +∑(e p -μ-a) 2 (2-1)

[0054] 3. Isovariant graph diffusion generation model

[0055] 3.1: Equivariant Graph Diffusion Generation Model

[0056] The isovariant graph diffusion generation model learns the sample distribution by modeling the inverse process of diffusion. Specifically, such as... Figure 4 As shown, a maximum step size T and a sample x are defined. The diffusion process gradually adds noise to the data points, and at the Tth step, it becomes a prior distribution noise that does not retain any sample information.

[0057] The isomorphic map diffusion generation model, based on given conditions, progressively denoises and restores the sample from the prior distribution noise, from the added noise to the maximum step size T. Taking the standard normal distribution as the prior distribution as an example, for a data point x, its diffusion result in step t is z. t Add to z t The noise is also Gaussian noise:

[0058]

[0059] In the formula: α t The remaining signal strength; σ t Let α be the noise intensity. Based on the principle of the isotropic diffusion generation model, α0 = 1, and it monotonically decreases to α as t increases. T ≈0. This model is defined. Define signal-to-noise ratio The diffusion process of the original data points can be represented as:

[0060]

[0061] Because the diffusion process is a Markov (stochastic) process, for the diffusion process from step s to step t (t>s), we have α t|s =α t / α s ,and The complete diffusion process can be represented as:

[0062]

[0063] Calculate z under the condition of x s Probabilistic sampling is the inverse process of diffusion, i.e., the denoising process, z s The conditional probability also follows a normal distribution:

[0064]

[0065] In formula (3-4), μ t→s (x,z t ) and σ t→s It is obtained through the following calculation:

[0066]

[0067] During the generation process, the original data x is unknown, therefore an estimate is needed. The substitution is performed, and this estimate is calculated by the model. Through After replacing x, p(z) s |z t ) was approximately rewritten as Formula (3-4) can then be rewritten as:

[0068]

[0069] Focusing only on a single denoising process, i.e., when s = t-1, the lower bound of the log-likelihood variation of x is as follows:

[0070]

[0071] in, It is the likelihood of the data points given z0, i.e., the initial likelihood. The final noise z is formed by adding noise at step T. T The difference from the standard normal distribution, measured by the KL divergence, i.e., the final likelihood, for the portion between step 0 and step T, is as follows, i.e., the process likelihood:

[0072]

[0073] Similar to variational autoencoders, which estimate the ELBO based on likelihood to calculate the loss function, diffusion-type models also use the lower bound provided by likelihood to guide the optimization direction; therefore, these likelihoods are also called likelihood losses. In practice, given a noisy result z... t The outputs of model φ, t, and φ are the noise it predicts. And prediction Need to be It is calculated. Specifically, the noise-added result can be expressed as z. t =α t x+σ t ∈, where ∈ is the actual noise, then:

[0074]

[0075] Therefore, it can be calculated using formula (3-11).

[0076]

[0077] The data used in the isomorphic graph diffusion generation model mainly consists of node features h and coordinates x. The processing of edge features is similar to that of node features, therefore edge features are omitted. Sometimes, node features and coordinates are concatenated in the model's calculation, represented as [x, h]. Formula (3-12) is rewritten when applied to the isomorphic graph diffusion generation model as follows:

[0078]

[0079] In the formula: It is the product of two distributions. noise distribution for coordinates, The noise distribution is a feature of the node.

[0080] Calculated using the model and Replacing x and h with the original data points, the version of the isovariate map diffusion generation model for the denoising process represented by formula (3-7) is shown below:

[0081]

[0082] Similar to formula (3-10), the direct output of the isovariant graph diffusion generation model is... It is the predicted noise, not the denoised data:

[0083]

[0084] The likelihood calculation formula for the diffusion generation model of the isomorphic graph is rewritten as formula (3-15):

[0085]

[0086] In the specific implementation of this model, w(t) is set to 1 to maintain training stability. However, based on the principle of the diffusion model, the actual form of w(t) should be as shown in formula (3-16):

[0087] w(t)=1-SNR(t-1) / SNR(t) (3-16)

[0088] 3.2: Input, output, and noise intensity of the isomorphic graph diffusion generation model during noise reduction

[0089] The input and output details of the isomorphic graph diffusion generation model, i.e. The isomorphic graph diffusion generation model part. First, for nodal coordinates, a non-zero distribution cannot be invariant to displacement because it cannot be integrated to 1. However, in a linear subspace, if the centroid of this distribution always lies at the origin, then this distribution can be applied to the isomorphic graph diffusion generation model (the centroid of the point group is determined by calculating the arithmetic mean of the coordinate components). Therefore, Defined as a ∑ i x i =0 is a normal distribution on a linear subspace. For nodal features with displacement invariance, similar to ordinary diffusion generation models, the noise can follow a conventional normal distribution. However, different treatments are needed for continuous and discrete variables in the initial likelihood. The input and output forms are shown in Equation (3-17):

[0090]

[0091] In formula (3-17), since the maximum time step T is generally hundreds to thousands, the time step t is normalized to t / T, and the normalized time step is concatenated with the node features. The coordinate noise predicted by the isomorphic graph diffusion generation model... It is obtained by subtracting the coordinates of the input isomorphic graph diffusion generation model from the coordinates output by the isomorphic graph diffusion generation model. In order to maintain the rotation and flip isomorphism of the isomorphic graph diffusion generation model, We also need to subtract the center of gravity to keep it at the original point.

[0092] Throughout the diffusion and noise reduction process, the α value of each step t is... t With σ t Between Therefore, α is defined as shown in formula (3-18). t That's all.

[0093]

[0094] In the formula, s = 1 × 10 -5 This is used to avoid "division by zero" errors in the program.

[0095] During the denoising process, let α be... t|t-1 =α t / α t-1 and define α -1 =1, so α t|t-1 The value is limited to between 0.001 and 1, so that 1 / α t|t-1 The value of α will not be too large. At this point, the value of α at a specific step... t It can then pass stably. Calculate. Simultaneously define a monotonically increasing function. To improve the calculation accuracy of other process quantities, for example:

[0096]

[0097] SNR(t)=exp(-γt)) (3-21)

[0098] 3.3: Training and Generation Process

[0099] Similar to other neural network models, the isomorphic graph diffusion generation model, during training, first forward propagates to output its current output value, then calculates the error between the output value and the theoretical output value, and backpropagates the error to update the isomorphic graph diffusion generation model parameters. In application, the output result is obtained through forward propagation. Specifically, during training, for a data point x, it is first sampled from a uniform distribution. Sampling from a normal distribution And In The center of gravity is placed at the origin in space. Then the noisy result z is calculated. t =α t +σ t ∈ t The goal of training is to minimize the difference between the predicted noise and the actual added noise, i.e.:

[0100]

[0101] The loss function of the isomorphic graph diffusion generation model is defined by ELBO, but unlike variational autoencoder models, the likelihood of the isomorphic graph diffusion generation model is divided into three parts: initial likelihood L0, process likelihood L1, and process likelihood L2. t and final likelihood L T First, forward propagation is performed, and the likelihood is calculated using the following formula:

[0102]

[0103] Sampling from normal distribution The centroid of its coordinate part is placed at the origin of space. Noise is added to the original data points until z0=α0+σ0∈0. The initial likelihood and final likelihood are calculated by formulas (3-24) and (3-25).

[0104]

[0105] The total likelihood loss is:

[0106]

[0107] Backpropagation of the likelihood loss completes the training of the model generated by the diffusion of the equivalent graph from the data points.

[0108] During the generation process, we first set t = T and sample the noisy result. Let s = t-1, and sample... And placing the centroid of the coordinate system at the origin, calculate:

[0109]

[0110] Then let t = s and continue the above process until s = 0. At this point, sampling x,h ~ p(x,h|z0) will yield the final generated result.

[0111] 3.4: Initial Likelihood of the Equivariant Graph Diffusion Generation Model

[0112] Taking node features as an example, we derive the initial likelihood of the isomorphic graph diffusion generation model. The initial likelihood of node coordinates has the same form as that of node features. The form of the initial likelihood is: Distribution of raw data When a small noise perturbation is applied (where α0≈1, σ0≈0), the conditional distribution of the original data can be expressed as:

[0113]

[0114] in It is a sampling of the original data h with added noise, because It is highly concentrated; its probability at point h approaches 1, while its probability at points other than h is almost 0.

[0115] For the one-hot vector used for discrete features, the probability of its value being 1 can be represented by integrating the sampled value over a period between 0.5 and 1.5:

[0116]

[0117] In the formula, C is a multinomial distribution, and n is the normalization operation on the integral result. Continuous features (and node coordinates) can be obtained by... Assuming it is a constant, it can be approximated as:

[0118]

[0119] The integral in this formula is obtained by passing through And introduce predicted noise Then, we can obtain:

[0120]

[0121] The initial likelihood can then be expressed as:

[0122]

[0123] Define w(0) = -1, which is also the form of the process likelihood at t = 0. Z is the normalization constant, expressed as:

[0124]

[0125] 3.5: Conditional Generation

[0126] To achieve conditional generation, an additional initial likelihood term needs to be provided to the isomorphic graph diffusion generation model during training. Process likelihood term and final likelihood term During generation, generation conditions need to be added to the isomorphic map diffusion generation model. The initial likelihood terms for node features and coordinates are shown in formulas (3-34) and (3-35), respectively:

[0127]

[0128] The process likelihood term is shown in formula (3-36):

[0129]

[0130] The final likelihood term is shown below:

[0131]

[0132] because and Approximately equal to 0, these two likelihood terms can be omitted. In summary, after sampling from c,M to p(c,M), ​​a trained conditional generation equivariant graph diffusion generation model can provide p(x) , The distribution of h|c,M), ​​and the process of sampling from this distribution x,h~p(x,h|c,M) is the conditional generation process.

[0133] 3.6: Training of Predictive Models, Property-Structure Pre-trained Embedded Models, and Equivariant Graph Diffusion Generation Models

[0134] The training process is to train the prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model in sequence. Specifically, first, the prediction model is trained using the data in the ILThermo dataset. Then, after predicting the missing values using the prediction model and completing the dataset supplementation, the property-structure pre-trained embedding model is trained by combining the experimental values and the predicted values. Finally, the property embeddings of each collected ionic working fluid are obtained using the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model is trained using the property embeddings and the physical structure of the ionic working fluid.

[0135] Step 2: Combine the trained prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model to construct an embedded graph diffusion generation model. The application process is as follows:

[0136] In the generation process of the traditional embedded graph diffusion generation model, its neural network only needs to perform a forward propagation once to achieve the output of the sample. The same is true for the decoder of the variational autoencoder and the generator of the generative adversarial network. Different from the traditional generation model, the embedded graph diffusion generation model completes the output of a sample by repeating its denoising process hundreds of times. This feature makes its generation process more flexible. In addition to the "from scratch" generation starting from the T-th step, this flexibility also allows for substructure completion and structure optimization, similar to inpainting and style transfer in image generation techniques.

[0137] Substructure completion means locking the defined substructure in the noisy graph and randomly sampling other atoms nearby. To guide the generation process to specific structural features (such as adding a carbon chain to an imidazole ring), the sampling process can be restricted to a specific direction. Subsequently, the graph is denoised to generate a complete molecule that meets the specified requirements.

[0138] Structure optimization means adding noise to the optimization object to the t-th step (0 < t < T) and letting the model start the denoising process from the intermediate step t instead of starting from the completely noisy state (the T-th step). This allows the generated molecule to retain some features of the original structure during optimization. The hyperparameter t is used to control the degree of change. A larger t can make the molecule deviate more significantly from the initial structure, thus promoting a deep exploration of the chemical space; in contrast, a smaller t can make the output of the embedded graph diffusion generation model more likely to be easily synthesized by retaining a certain similarity. Substructure completion and structure optimization are more likely to generate optimized ionic working fluid structures that can be synthesized and meet the required performance criteria, while the from-scratch generation is more original and has a wider generation range.

[0139] Step 3: Use the embedded graph diffusion generation model for design to obtain the target ionic working fluid.

[0140] Node coordinates and node features have different units and orders of magnitude, so these quantities need to be scaled to facilitate the training of the embedding graph diffusion generation model. During interpretation, these physical quantities will also be processed back to their initial values ​​according to the weights.

[0141] The output of the embedding graph diffusion generation model is the denoised features of each node and edge of the graph, which needs to be interpreted back to the specific molecular structure. Due to the powerful generation capability of the embedding graph diffusion generation model, no manual constraints need to be imposed during output interpretation; therefore, the interpretation method is approximately simultaneous. During interpretation, the atoms are first interpreted by checking the one-thermal encoding of the atom type in the node's features and finding the position with the largest value; the atom type corresponding to this position is the interpretation result. Then, the chemical bonds are interpreted; the bond type between all atoms is determined by their distance.

[0142] The present invention will be further described in detail below through specific embodiments:

[0143] Example 1:

[0144] Ionic working fluids, as next-generation carbon dioxide capture and absorbents, help improve capture efficiency, reduce equipment size, and adapt to the long-cycle operation requirements of industrial plants. This invention designs low-energy-consumption ionic liquid absorbents for carbon dioxide capture devices. The energy consumption of carbon dioxide capture devices is affected by several complex nonlinear properties of the absorbent. Lower viscosity leads to reduced pump energy consumption; specific heat capacity under absorption and desorption conditions affects heat consumption in the capture process; higher solubility differences in rich and dilute solvents allow for a smaller ionic liquid flow rate required to capture a unit of carbon dioxide, thus reducing overall energy consumption. For carbon dioxide capture devices, this invention uses an embedded graph diffusion generation model, comprehensively considering the key properties affecting system energy consumption, to design a low-energy-consumption ionic liquid, and the design results were synthesized and verified.

[0145] 1.1 Carbon Dioxide Capture Device and Energy Consumption Model

[0146] An ionic liquid-based carbon dioxide capture system comprises three main modules: an energy recovery module, a compression module, and a carbon dioxide capture module, such as... Figure 5 As shown.

[0147] The energy recovery module includes a hot tank, a cold tank, a primary preheater, a primary turbine, a secondary preheater, and a secondary turbine; the compression module includes a cyclone separator, a primary compressor, a primary intercooler, a secondary compressor, and a secondary intercooler; the carbon dioxide capture module includes an absorption tower and a desorption tower, a rich-dilute solution heat exchanger, a rich solution pump, and a dilute solution pump. After processing by the system, the high-purity carbon dioxide extracted from the desorption tower is guided to the post-processing module for further application.

[0148] In the compression module, the flue gas is first purified by a cyclone separator to remove solid impurities such as ash. Then, the gas is subjected to proportional compression by a two-stage compressor. During compression, an intercooler is used to cool the compressed gas, and the heat absorbed from the compressed gas can be recovered to reduce energy consumption. Furthermore, the second intercooler increases the solubility of carbon dioxide in the ionic liquid by cooling. After compression and cooling, the gas is transported to the carbon dioxide capture module, while the heat absorbed by the intercooler from the carbon dioxide is transferred to the energy recovery module via a heat storage liquid (dodecylcyclohexasiloxane, D6).

[0149] In the absorption tower of the carbon dioxide capture module, gas from the compression module is mixed with an ionic liquid under high pressure and low temperature conditions. The carbon dioxide in the gas is absorbed by the ionic liquid, forming a rich solution. The rich solution is then pumped to a rich-dilute solution heat exchanger for preheating. After preheating, the solvent is sent to a desorption tower for desorption under high temperature and low pressure conditions. The heat used to heat the rich solution in the desorption tower is provided by waste heat from the power plant's boiler, releasing high-purity carbon dioxide. After desorption, the dilute solution is pumped to a rich-dilute solution heat exchanger for precooling, and then returned to the absorber to continue the absorption-desorption cycle.

[0150] This system differs from traditional ionic liquid-based carbon dioxide capture systems in that it incorporates an energy recovery module that recovers residual kinetic energy from the high-pressure N2-O2 mixture emitted from the absorber to further reduce energy consumption. In the energy recovery module, a preheater uses a regenerated liquid from the intercooler to heat the gas mixture, which is then fed into a turbine for isotropic expansion. The cooled regenerated fluid then flows through a cold tank, cooling the compressed gas in the intercooler and completing the regenerated fluid cycle. The exhaust gas mixture is then released into the atmosphere.

[0151] In carbon dioxide capture systems, the temperature of the ionic liquid ranges from 335.64 K to 353.15 K. The ionic liquid properties affecting carbon dioxide capture energy consumption include: molecular weight; dynamic viscosity, density, and specific heat capacity at isobaric pressures at 335.64 K, 1 MPa and 353.15 K, 0.101 MPa; and solubility of carbon dioxide at 335.64 K, 1 MPa, 353.15 K, 0.101 MPa, and 339.14 K, 0.101 MPa.

[0152] The ILThermo dataset contains 23 ionic liquids with sufficient data on dynamic viscosity, density, specific heat capacity, and carbon dioxide solubility for energy consumption analysis. The molecular weights of these ionic liquids can be calculated using RDKit. The energy consumption range of these ionic liquids is 1.84 GJ·tCO2. -1 Up to 7.27 GJ·tCO2 -1 .

[0153] 1.2 Algorithm Implementation Details

[0154] 1.2.1 Dataset and Parameter Settings

[0155] The embedding graph diffusion generation model was trained using the ILThermo dataset. Since using molecular graphs as molecular identifiers is not limited by the maximum number of branches in the spanning tree, a total of ionic liquids consisting of 322 anions and 860 cations participated in the training of the property-structure pre-trained embedding model and the isovariant graph diffusion generation model. For the prediction model, within the range of 100–573 K and 70.28–4999 kPa, 20,426 data points from 936 ionic liquids were used to train the viscosity model; 44,645 data points from 1,158 ionic liquids were used to train the density model; 15,935 data points from 265 ionic liquids were used to train the specific heat capacity model; and 14,007 data points from 208 ionic liquids were used to train the solubility model.

[0156] During training of each model, the dataset was randomly divided into training, validation, and test sets at a ratio of 70%, 20%, and 10%, respectively. The training set was used to train the model, and the iteration with the highest prediction accuracy on the validation set was taken as the training result. For the prediction model, a total of 15 iterations were performed with a batch size of 10. The ADAM optimization algorithm was used, with a learning rate of 0.0001, exponential decay factors for the first and second moments of 0.9 and 0.999, respectively, and a weight decay of 0.001. The learning rate was reduced by 20% in each iteration until the minimum learning rate of 0.00001 was reached. For the property-structure pre-trained embedding model, a total of 50 iterations were performed with a batch size of 10. The ADAM optimization algorithm was used with a learning rate of 0.0001, and the exponential decay factors for the first and second moments and the weight decay were the same as above. The learning rate was reduced by 60% every two iterations until the minimum learning rate of 0.00001 was reached. For the isomorphic graph diffusion generation model, a total of 600 iterations were performed with a batch size of 10. The ADAMW optimization algorithm was used with a learning rate of 1×10⁻⁶. -5 The exponential decay factors for the first and second moments are the same as above, and the weight decay is 1×10. -12 .

[0157] 1.2.2 Network Structure

[0158] The LeakyReLU and SiLU activation functions used are shown below:

[0159]

[0160] 1) Prediction Model

[0161] For a given ion, the prediction model first uses an embedding layer to embed node features and unifies the dimensions of the node features with those of the AEGNN layer. Then, several AEGNN layers extract structural information from the ion layer by layer. After all AEGNN layers have passed, an averaging layer averages each dimension of the node features and reshapes the node feature matrix from [number of nodes × number of features] to [number of features], ensuring consistent input size for the final FC layer. Finally, the outputs of the cation and anion models are concatenated with temperature and pressure conditions and input into an FC layer. This layer then summarizes the information extracted from the cation and anion to output the calculated attribute values. The number of features, layers, etc., of the prediction models vary slightly depending on their properties; their detailed structures are shown in Table 1-1.

[0162] Table 1-1 Structure of the prediction model

[0163]

[0164]

[0165] Note: a 'b' represents the anion model, and 'b' represents the cation model.

[0166] 2) Property-structure pre-trained embedding model

[0167] As shown in Table 1-2, the structure and input of the ion encoder are generally similar to those of the prediction model, but differ in the last layer: to ensure that the ion embedding is unique and constant, the anion and cation parts are not connected through the FC layer, and the output dimension is equal to the embedding dimension. The structure of the attribute encoder is a multi-layer FC network.

[0168] Table 1-2 Structure of the Embedded Model

[0169]

[0170]

[0171] 3) Isomorphic graph diffusion generation model

[0172] As shown in Table 1-3, in the isomorphic graph diffusion generation model, the input dimension = atom type + charge + conditional dimension (9+1+10 for anions, 9+1+20 for cations). In the isomorphic graph diffusion generation model, the atom type includes an additional type, "non-atomic," therefore, the atom type is 9-dimensional instead of the 8-dimensional one in the prediction model. The output of the isomorphic graph diffusion generation model contains an atom type vector (h) with charge information. c and h a ) and atomic coordinates (x c and x aThe atom type vector is output by the final embedding layer. The atom coordinates are not processed by the final embedding layer and are directly output by the AEGNN layer.

[0173] Table 1-3 Structures generated by diffusion in isomorphic diagrams

[0174]

[0175] 1.2.3 Training and Application Process

[0176] The training process involves sequentially training the prediction model, the property-structure pre-trained embedding model, and the isomorphic graph diffusion generation model. Specifically, the prediction model is first trained using data from the ILThermo dataset. Then, after the prediction model predicts missing values ​​and completes the dataset, the property-structure pre-trained embedding model is trained using both experimental and predicted values. Finally, the property-structure pre-trained embedding model is used to obtain the property embeddings for each collected ionic working fluid, and the isomorphic graph diffusion generation model is trained using these property embeddings and the actual structures of the ionic working fluids.

[0177] 1.2.4 Interpretation Method of Embedded Graph Diffusion Generation Model Output

[0178] Node coordinates and node features have different units and orders of magnitude. In the final embedded graph diffusion generation model, the one-hot encoding of the atom type in the structural features is multiplied by 0.4, and the relative atomic mass is multiplied by 0.05.

[0179] During the interpretation process, the atoms are first interpreted by checking the unique thermal encoding of the atom type in the node features. The position with the largest value is found, and the atom type corresponding to that position is the interpretation result. Then, the chemical bonds are interpreted. The bond type between all atoms is determined by their distance: if the distance is within the bond length ± tolerance, the atoms are considered to be connected by a bond (the tolerance specified in this invention is 8 pm). Typical lengths of single, double, and triple bonds are shown in Tables 1-4, 1-5, and 1-6.

[0180] Table 1-4 Typical values ​​of single bond length

[0181] Bond length / pm C O N P S F I B C 154 143 147 184 182 135 214 157 O — 148 140 163 151 142 194 138 N — — 145 177 168 136 222 145 P — — — 221 210 156 241 190 S — — — — 204 158 234 181 F — — — — — 142 187 145 I — — — — — — 266 213 B — — — — — — — 177

[0182] Table 1-5 Typical values ​​of double bond length

[0183]

[0184]

[0185] Table 1-6 Typical values ​​of triple bond length

[0186] Bond length / pm C O N C 120 113 116 O — — — N — — 110

[0187] 1.3 Design Result Analysis

[0188] In the design results, the ionic liquid with the lowest energy consumption—4-ethyl-1-propylpyridinium tetrafluoroborate—was selected for further experimental evaluation. This ionic liquid was derived from the structural optimization of 1-ethyl-3-methylimidazolium trifluoromethanesulfonate ([EMIM][TRIFLATE]). [EMIM][TRIFLATE] is currently the most advanced low-energy ionic liquid, with an energy consumption of 1.84 GJ·tCO2. -1 The optimization process involves diffusing the molecular map of [EMIM][TR IFLATE] to step 400, and then denoising it back to step 0 under defined generation conditions.

[0189] Table 1-7 summarizes the properties and experimental measurements of the embedding graph diffusion generation model used as the generation condition input for 4-ethyl-1-propylpyridine tetrafluoroborate. The energy consumption for generating 4-ethyl-1-propylpyridine tetrafluoroborate is 1.52 GJ·tCO2. -1 It reduces energy consumption by 17.6% compared to [EMIM][TRIFLATE], the ionic liquid with the lowest known energy consumption.

[0190] Table 1-7 Formation conditions and experimental measurement results of 4-ethyl-1-propylpyridine-1-tetrafluoroborate ammonium

[0191]

[0192]

[0193] As shown in Tables 1-7, the average relative deviation between the experimental results and the formation conditions for various properties of 4-ethyl-1-propylpyridine tetrafluoroborate was 9.2%, and the energy consumption deviation was 3.6%. Specifically, the molar mass deviation was 1.23%; the dynamic viscosity deviation was less than 13%; the density deviation was less than 2.8%; the specific heat capacity deviation was less than 2.7%; and the carbon dioxide solubility deviations at 1 MPa and 0.101 MPa were less than 4.75%, 48.15%, and 5.9%, respectively. This demonstrates that the embedded graph diffusion generation model of this invention, in addition to significantly reducing the energy consumption for carbon dioxide capture, also exhibits a high degree of consistency between the designed ionic liquid and the formation conditions and experimental measurements. This proves that this invention can accurately perform multi-objective property conditional generation.

[0194] A detailed energy consumption comparison between [EMIM][TRIFLATE] and 4-ethyl-1-propylpyridine tetrafluoroborate is as follows: Figure 6As shown in the figure, the energy consumption for compression of carbon dioxide per unit is only related to the operating conditions and therefore has no difference. Due to the lower specific heat capacity and viscosity of 4-ethyl-1-propylpyridine tetrafluoroborate, both the heat consumption for carbon dioxide desorption and the pump energy consumption are reduced. Furthermore, the significant difference in solubility between absorption and desorption conditions results in less ionic liquid required to capture a unit mass of carbon dioxide, which further reduces overall energy consumption, ultimately reducing the desorption heat consumption by 0.029 GJ·tCO2. -1 The pump's energy consumption was reduced by 0.30 GJ·tCO2 -1 This makes 4-ethyl-1-propylpyridine tetrafluoroborate superior to all currently known ionic liquids in terms of properties.

[0195] Example 2:

[0196] This invention also provides a design system for ionic working fluids, such as... Figure 7 As shown, the system includes: a first modeling module, a second modeling module, and a design module.

[0197] The first modeling module is used to obtain the design target of the target ionic working fluid, construct a prediction model, a property-structure pre-trained embedding model and an isomorphic graph diffusion generation model and train them.

[0198] The second modeling module is used to combine the trained prediction model, the property-structure pre-trained embedding model and the equivariant graph diffusion generation model to construct the embedding graph diffusion generation model.

[0199] The design module is used to design using an embedded graph diffusion generation model to obtain the target ionic working fluid.

[0200] It is understood that the design system for ionic working fluids provided by the present invention corresponds to the design methods for ionic working fluids provided in the foregoing embodiments. The relevant technical features of the design system for ionic working fluids can be referred to the relevant technical features of the design methods for ionic working fluids, and will not be repeated here.

[0201] Another object of the present invention is to provide an electronic device, such as... Figure 8 As shown, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor performing the steps of the design method for the ionic working fluid.

[0202] The design method for the ionic working fluid includes the following steps:

[0203] The design objectives of the target ionic working fluid are obtained, and a prediction model, a property-structure pre-trained embedding model, and an isomorphic graph diffusion generation model are constructed and trained.

[0204] The trained prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model are combined to construct the embedding graph diffusion generation model;

[0205] The target ionic working fluid was designed using an embedded graph diffusion generation model.

[0206] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the design method for the ionic working fluid.

[0207] The design method for the ionic working fluid includes the following steps:

[0208] The design objectives of the target ionic working fluid are obtained, and a prediction model, a property-structure pre-trained embedding model, and an isomorphic graph diffusion generation model are constructed and trained.

[0209] The trained prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model are combined to construct the embedding graph diffusion generation model;

[0210] The target ionic working fluid was designed using an embedded graph diffusion generation model.

[0211] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0212] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0213] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0214] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0215] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A design method for an ionic working fluid, characterized in that, include: The design objectives of the target ionic working fluid are obtained, and a prediction model, a property-structure pre-trained embedding model, and an isomorphic graph diffusion generation model are constructed and trained. The trained prediction model, the property-structure pre-trained embedding model, and the equivariant graph diffusion generation model are combined to construct the embedding graph diffusion generation model; The target ionic working fluid was obtained by designing using an embedded graph diffusion generation model.

2. The design method for an ionic working fluid according to claim 1, characterized in that, The prediction model takes the ionic structure as input and outputs predicted property values, including molecular layers, embedding layers, AEGNN layers, averaging layers, and FC layers.

3. The design method for an ionic working fluid according to claim 2, characterized in that, The prediction model uses the following method to predict property values: The ionic structure is input into the molecular layer, and the embedding layer embeds node features based on the molecular graph input from the molecular layer, and unifies the dimensions of the node features with those of the AEGNN layer. The AEGNN layer extracts structural information from the ion layer by layer, and the averaging layer averages each dimension of the node features and reshapes the matrix format of the node features. The node features after averaging in the average layer are input into the FC layer, and the FC layer summarizes the information extracted from the cation and anion layers to output the calculated property values.

4. The design method for an ionic working fluid according to claim 1, characterized in that, The property-structure pre-trained embedding model includes a property encoder and a structure encoder. The property encoder takes a vector composed of multiple properties of the ionic working fluid as input and outputs the property embedding. The structure encoder includes anion encoder and cation encoder, which take the ionic structure as input and output the ionic structure embedding.

5. The design method for an ionic working fluid according to claim 1, characterized in that, The isomorphic graph diffusion generation model gradually denoises and restores samples by removing prior distribution noise that has been added up to the maximum step size, based on the generation conditions. The input includes atom type, charge, and conditional dimension, and the output is an atom type vector and atom coordinates with charge information.

6. The design method for an ionic working fluid according to claim 1, characterized in that, The prediction model, the property-structure pre-trained embedding model, and the isomorphic graph diffusion generation model are trained sequentially: The prediction model was trained using data from the ILThermo dataset. After predicting missing values ​​using a predictive model and completing the ILThermo dataset, the property-structure pre-trained embedding model is trained by combining experimental and predicted values. The property embeddings of the collected ionic working fluids are obtained by using a property-structure pre-trained embedding model, and the property embeddings and the ionic working fluid structure are used to train an isomorphic graph diffusion generation model.

7. The design method for an ionic working fluid according to claim 1, characterized in that, The method for interpreting the output of the embedding graph diffusion generation model is as follows: Examine the one-hot encoding of the atom type in the node features, find the position with the largest value, and the corresponding atom type is the interpretation result; When interpreting chemical bonds, the type of bond between all atoms is determined by the distance between them.

8. A design system for an ionic working fluid, characterized in that, include: The first modeling module is used to obtain the design target of the target ionic working fluid, construct a prediction model, a property-structure pre-trained embedding model and an isomorphic graph diffusion generation model and train them. The second modeling module is used to combine the trained prediction model, the property-structure pre-trained embedding model and the equivariant graph diffusion generation model to construct the embedding graph diffusion generation model. The design module is used to design using an embedded graph diffusion generation model to obtain the target ionic working fluid.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images