Method and device for training and optimizing an analytical model, electronic device and storage medium
By combining graph neural networks and multilayer perceptrons, the feature representation and loss function of the analysis model are optimized, which solves the problem of the realism and accuracy of electrolyte simulation and realizes efficient and accurate prediction of electrolyte properties in large systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- DOUYIN VISION CO LTD
- Filing Date
- 2023-10-12
- Publication Date
- 2026-07-07
Smart Images
Figure CN117316316B_ABST
Abstract
Description
Technical Field
[0001] The exemplary embodiments disclosed herein generally relate to the field of computers, and particularly to methods, apparatus, electronic devices, and computer-readable storage media for training and optimizing analytical models. Background Technology
[0002] Electrolytes play an indispensable role in wearable devices, electric vehicles, and static energy storage devices. Furthermore, electrolyte research is also significant for non-energy storage applications, such as in the food, environmental, and pharmaceutical fields.
[0003] With advancements in computing power, the design and iteration of electrolytes are increasingly trending towards digitalization. However, current tools cannot simultaneously satisfy the requirements of realism and accuracy in electrolyte simulations. Summary of the Invention
[0004] In a first aspect of this disclosure, a method for optimizing an analytical model is provided. The method includes: fine-tuning the analytical model using a first set of values regarding a first property of a target substance to determine a second set of values for the target substance regarding a second property; determining a correlation between the first and second properties of the target substance based on the first and second sets of values; determining a target value for the target substance regarding the first property using the correlation based on a reference value for the target substance regarding the second property, the reference value being determined experimentally based on the target substance; and optimizing the analytical model using the target value for the target substance regarding the first property.
[0005] In a second aspect of this disclosure, a method for training an analytical model is provided. The method includes: determining a set of feature representations of a set of atoms in an electrolyte sample using the analytical model; determining charge information of the electrolyte sample using the analytical model based on the set of feature representations; determining energy information of the electrolyte sample using the analytical model based on the charge information and the set of feature representations; and training the analytical model based on the charge information and the energy information, wherein a loss function used to train the analytical model includes at least a first part and / or a second part, the first part being associated with a multipole moment determined based on the charge information, and the second part being associated with stress information determined based on the energy information.
[0006] In a third aspect of this disclosure, a method for training an analytical model is provided. The method includes: acquiring multiple analytical models for analyzing a target substance, the multiple analytical models being trained based on different initial values of model parameters; using the multiple analytical models to determine multiple predicted results regarding the forces acting on a group of atoms in the target substance; based on the multiple predicted results, determining target data regarding the forces acting on the group of atoms; and using the target data to train a target analytical model for analyzing the target substance.
[0007] In a fourth aspect of this disclosure, an apparatus for optimizing an analytical model is provided. The apparatus includes: a fine-tuning module configured to fine-tune the analytical model using a first set of values relating to a first property of a target substance to determine a second set of values relating to a second property of the target substance; a first determining module configured to determine a correlation between the first and second properties of the target substance based on the first and second sets of values; a second determining module configured to determine a target value of the target substance relating to the first property using the correlation based on a reference value of the target substance relating to the second property, the reference value being determined based on experiments with the target substance; and an optimization module configured to optimize the analytical model using the target value of the target substance relating to the first property.
[0008] In a fifth aspect of this disclosure, an apparatus for training an analytical model is provided. The apparatus includes: a feature determination module configured to determine a set of feature representations of a set of atoms in an electrolyte sample using the analytical model; a charge determination module configured to determine charge information of the electrolyte sample based on the set of feature representations using the analytical model; an energy determination module configured to determine energy information of the electrolyte sample based on the charge information and the set of feature representations using the analytical model; and a first training module configured to train the analytical model based on the charge information and the energy information, wherein a loss function for training the analytical model includes at least a first part and / or a second part, the first part being associated with a multipole moment determined based on the charge information, and the second part being associated with stress information determined based on the energy information.
[0009] In a sixth aspect of this disclosure, an apparatus for training an analytical model is provided. The apparatus includes: a model acquisition module configured to acquire multiple analytical models for analyzing a target substance, the multiple analytical models being trained based on different initial values of model parameters; a result prediction module configured to use the multiple analytical models to determine multiple predicted results regarding the forces acting on a group of atoms in the target substance; a third determination module configured to determine target data regarding the forces acting on the group of atoms based on the multiple prediction results; and a second training module configured to use the target data to train a target analytical model for analyzing the target substance.
[0010] In a seventh aspect of this disclosure, an electronic device is provided. The device includes at least one processing unit; and at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit. When executed by the at least one processing unit, the instructions cause the electronic device to perform the methods of the first aspect, the second aspect, or the third aspect.
[0011] In an eighth aspect of this disclosure, a computer-readable storage medium is provided. A computer program is stored on the medium, which, when executed by a processor, implements the methods of the first, second, or third aspect.
[0012] It should be understood that the description in this section is not intended to limit the key or essential features of the embodiments of this disclosure, nor is it intended to restrict the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description
[0013] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:
[0014] Figure 1 A schematic diagram of an example environment in which embodiments of the present disclosure can be implemented is shown;
[0015] Figure 2 A flowchart illustrating an example process for training an analysis model according to some embodiments of this disclosure is shown;
[0016] Figure 3 An example architecture of an analysis model according to some embodiments of this disclosure is shown;
[0017] Figure 4 A flowchart illustrating an example process of an optimization analysis model according to some embodiments of this disclosure is shown;
[0018] Figure 5 A schematic diagram of a fine-tuning analysis model according to some embodiments of the present disclosure is shown;
[0019] Figure 6 A flowchart illustrating an example process for training an analysis model according to other embodiments of this disclosure is shown;
[0020] Figures 7A to 7I Performance comparison charts of models according to some embodiments of this disclosure are provided;
[0021] Figure 8A A schematic structural block diagram of an apparatus for optimizing an analysis model according to some embodiments of the present disclosure is shown;
[0022] Figure 8B A schematic structural block diagram of an apparatus for training an analysis model according to some embodiments of the present disclosure is shown;
[0023] Figure 8C A schematic structural block diagram of an apparatus for training an analysis model according to some embodiments of the present disclosure is shown; and
[0024] Figure 9 A block diagram of an electronic device capable of implementing several embodiments of the present disclosure is shown. Detailed Implementation
[0025] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.
[0026] It should be noted that the headings of any section / subsection provided herein are not limiting. Various embodiments are described throughout this document, and embodiments of any type may be included under any section / subsection. Furthermore, embodiments described in any section / subsection may be combined in any way with any other embodiments described in the same section / subsection and / or different sections / subsections.
[0027] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The term "some embodiments" should be understood as "at least some embodiments". Other explicit and implicit definitions may also be included below. The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.
[0028] The embodiments of this disclosure may involve user data, data acquisition, and / or use. All of these aspects comply with applicable laws, regulations, and relevant provisions. In the embodiments of this disclosure, all data collection, acquisition, processing, manipulation, forwarding, and use are conducted with the user's knowledge and confirmation. Accordingly, in implementing the embodiments of this disclosure, the type, scope of use, and usage scenarios of any data or information that may be involved should be communicated to the user and their authorization obtained in accordance with relevant laws and regulations through appropriate means. The specific methods of notification and / or authorization may vary depending on the actual situation and application scenario, and the scope of this disclosure is not limited in this respect.
[0029] In this specification and the embodiments, any processing of personal information will be carried out only under the premise of legality (such as obtaining the consent of the personal information subject, or being necessary for the performance of a contract), and will only be carried out within the scope stipulated or agreed upon. A user's refusal to process personal information other than that necessary for basic functions will not affect the user's use of basic functions.
[0030] As used herein, the term "solution" refers to a system formed by mixing two or more substances. For example, a solution may include a matrix and a solute salt added to the matrix. In embodiments of this disclosure, solutions may have any suitable physical state, including but not limited to liquid, solid, gas, etc. Electrolytes are formed by adding a metal salt to a solvent. An electrolyte is an example of a solution.
[0031] As briefly mentioned earlier, with the advancement of computing power, the design and iteration of electrolytes are gradually becoming digitalized. However, current tools cannot simultaneously meet the requirements of realism and accuracy in electrolyte simulation.
[0032] High-precision, first-principles-based methods, such as density functional theory (DFT) and AIMD (Abstract Molecular Dynamics) simulations, are only suitable for simulations at the nanometer and nanosecond scales. They cannot provide high-throughput, effective predictions of key properties such as density, viscosity, dielectric constant, and ionic conductivity of electrolyte materials. While empirical molecular dynamics (MD) simulations can cover scales up to the micrometer and microsecond levels, their low precision and reliance on predefined parameter forms for energy curve fitting to describe intramolecular and intermolecular interactions limit their extrapolation capabilities and applicability, making them only suitable for preliminary qualitative exploration of materials.
[0033] In addition to the previously mentioned materials simulation methods, with the continuous advancement of artificial intelligence (AI) technology, methods that integrate AI with materials simulation are gradually emerging. This integration brings about the concept of Machine Learning Force Fields (MLFF). MLFF aims to achieve a scale similar to molecular dynamics (MD) simulations on the simulation scale, and approach the accuracy of first-principles calculations, thereby striving to achieve reliable predictions of materials properties consistent with experimental results.
[0034] Training MLFFs first requires generating training data using AIMD and DFT. A crucial step is selecting a suitable and reliable functional to describe the system's electron density. However, functional selection is often empirical, leading to different material predictions from different functionals trained on MLFFs. Finding a suitable functional through trial and error is time-consuming and resource-intensive, and lacks transferability; for example, a functional suitable for liquids / polymers may not be suitable for inorganic amorphous materials.
[0035] Furthermore, most machine learning model frameworks assume that the potential energy surface can be represented by the summation of many-body interaction terms of a series of atoms within a short, fixed cutoff distance, often neglecting or underestimating long-range electrostatic interactions. This makes it challenging to accurately capture the properties such as density, viscosity, dielectric constant, and ionic conductivity affected by long-range electrostatic interactions in liquid and other amorphous states. These two limitations result in a significant gap between most machine learning models and experimental predictions of the properties of amorphous electrolyte materials.
[0036] The following section provides a detailed description of various example implementations of this scheme, with reference to the accompanying drawings.
[0037] Example Environment
[0038] Figure 1 A schematic diagram of an example environment 100 in which embodiments of the present disclosure can be implemented is shown. For example... Figure 1 As shown, example environment 100 may include electronic device 110.
[0039] In this example environment 100, electronic device 110 can be used to train and optimize analysis model 120. Analysis model 120 can be used to predict the properties of a substance, such as density, viscosity, conductivity, etc. Therefore, analysis model 120 can also be called a property prediction model. For example, such analysis model 120 can be used to simulate electrolytes to predict their properties. As will be described in detail below, analysis model 120 can, for example, receive a sample 130 to be simulated (e.g., an electrolyte or solution) and generate prediction results 140 regarding specific properties of sample 130.
[0040] The following section will detail the training and optimization process of Model 120.
[0041] Joint training of analytical models
[0042] Figure 2 A flowchart illustrating an example process 200 for constructing a ligand generation model according to some embodiments of the present disclosure is shown. Process 200 can be implemented at an electronic device 110. Reference is made below. Figure 1 To describe process 200.
[0043] In box 210, electronic device 110 uses an analytical model to determine a set of characteristic representations of a group of atoms in an electrolyte sample.
[0044] The following will be referenced Figure 3 To describe process 200, Figure 3Example architectures according to some embodiments of this disclosure are shown. Figure 3 As shown, the analysis model 120 can receive electrolyte information 310 from the electrolyte sample.
[0045] In some embodiments, such electrolyte information 310 may include, for example, atomic positions, atomic types, and edge information.
[0046] In some embodiments, in order to ensure rotational and translational invariance / isotropy during the MLFF process, the analysis model 120 may, for example, utilize graph neural networks (GNNs) to determine the characteristic representation of each atom in the electrolyte.
[0047] Specifically, the analysis model 120 can determine a graph representation corresponding to a set of atoms in the electrolyte sample, wherein multiple nodes in the graph representation correspond to a set of atoms, and the edges in the graph representation indicate that the distance between two corresponding atoms is less than a second threshold.
[0048] Furthermore, the analysis model 120 can utilize the graph neural network of the analysis model to process the graph representation and determine a set of feature representations of a set of atoms in the electrolyte sample.
[0049] Therefore, GNNs can model each atom in the system as a point in a graph, connecting pairs of atoms with a distance less than a certain threshold with an edge, forming a graph. Atoms exchange information through multiple rounds via edges to update their features. Finally, the system's energy, the forces acting on atoms, and other information are predicted using the features on atoms and edges. Strict rotation and translation invariance / isotropy can be achieved in graph neural networks.
[0050] The GNN process updates the feature representations of each atom through an attention mechanism. However, when based on global attention, its computational complexity reaches O(N). 2 In electrolyte simulations, common scenarios involve systems with more than 10,000 atoms, therefore O(N) 2 The complexity of this will make the simulation very slow.
[0051] In some embodiments, the analysis model 120 may employ a local attention mechanism to update the feature representations of atoms. Specifically, the analysis model 120 may determine a first feature representation for a first atom in a set of atoms, and may further determine at least one second atom whose distance to the first atom is less than a third threshold, based on the distances of the set of atoms to the first atom. Exemplarily, the analysis model 120 may consider only the atom closest to the first atom.
[0052] Additionally, the analysis model 120 can update the first feature representation of the first atom using the feature representation of at least one second atom. Furthermore, the analysis model 120 can determine a set of feature representations for a group of atoms based on the updated first feature representation.
[0053] Therefore, during feature propagation, the attention mechanism can operate only between atoms that are connected by an edge. This reduces the complexity to O(N), thereby significantly improving simulation efficiency.
[0054] In box 220, electronic device 110 uses an analytical model to determine the charge information of an electrolyte sample based on a set of feature representations.
[0055] Specifically, after determining the feature representation (i.e., embedding) of each atom, the electronic device 110 can use the multilayer perceptron (MLP) included in the analysis model 120 to determine the charge information 325 of each atom.
[0056] In some embodiments, the analysis model 120 can also ensure that the charge is bounded by a Tanh layer and that the total charge of the system is constant by subtracting the predicted mean.
[0057] In box 230, electronic device 110 uses an analytical model to determine the energy information of an electrolyte sample based on charge information and a set of feature representations.
[0058] Specifically, the characteristic representation and predicted charge information 325 of each atom can be processed by a multilayer perceptron (MLP) in the analysis model 120 to determine the energy of each atom. Further, the energy information, i.e., the energy prediction 315, can be obtained through summation, which can be expressed, for example, as:
[0059]
[0060] Among them, emb i This represents the characteristic representation of the i-th atom.
[0061] In box 240, electronic device 110 trains an analysis model based on charge information and energy information, wherein the loss function used to train the analysis model includes at least a first part and / or a second part, the first part being associated with the multipole moment determined based on charge information and the second part being associated with stress information determined based on energy information.
[0062] Specifically, the electronic device 110 can train the analysis model 120 by jointly training multiple loss functions. In some embodiments, the electronic device 110 can determine the system's multipole moments based on charge information 325, such multipole moments may correspond to second-order or higher-order terms in a multi-level expansion. For example, such multipole moments may include the system's quadrupole moments, octupole moments, etc.
[0063] For example, electronic device 110 may use the mean-square error (MSE) of the system's multipole moments as part of the loss function for model training to train the model.
[0064] In physics, the multipole expansion method is used to approximate the electric potential generated by a charge distribution. The charge and dipole moment correspond to the 0th and 1st order terms of the multipole expansion, respectively. Introducing the multipole moment as a training objective helps the model better predict higher-order terms of the multipole expansion, thus providing a better approximation of the electric potential.
[0065] Additionally, the loss function includes a third part, which is associated with charge information and / or the dipole moment determined based on the charge information. For example, electronic device 110 can jointly train the model using the charge information, dipole moment, and mean square error of the multipole moment as the training loss function.
[0066] In some embodiments, the electronic device 110 may also determine the system's stress information based on energy prediction 315, such as a virial tensor, which may be determined, for example, based on the following process:
[0067]
[0068]
[0069] As shown in formula (2), the negative derivative of the total energy of the system with respect to each atomic position is the force on the atom; as shown in formula (3), the virial tensor of the system is defined as the sum of the product of the atomic force and the atomic position.
[0070] The pressure of the system is calculated based on the diagonal terms of the virial tensor, while the viscosity is calculated based on the off-diagonal terms. Therefore, better prediction of the virial tensor of the system is beneficial for better prediction of the system properties.
[0071] Furthermore, considering that GNNs are graphs constructed based on nearest neighbors, they cannot characterize long-range interactions between atoms (i.e., electrostatic interactions and dispersion interactions). Figure 3 As shown, the total energy 345 predicted by the analysis model 120 can be based on the energy predicted by the neural network (i.e., energy prediction 315), the electrostatic energy 335, and the dispersive energy ( Figure 3(Not shown in the figure) combination. The electrostatic energy 335 can be obtained, for example, based on the Coulomb formula, while the dispersion energy is determined by dispersion correction of the DFT.
[0072] Accordingly, the total force 340 could also be based on the force predicted by that network (i.e., force prediction 320), the electrostatic force 330, and the dispersion force ( Figure 3 (Not shown in the figure). In some quantities, to prevent electrostatic and dispersion forces from becoming too large when the atomic distance is too close, which could lead to simulation instability, a damping function can be introduced to ensure that these two long-range effects decay to 0 at short ranges.
[0073] Therefore, the electronic device 110 can utilize loss functions of multiple dimensions to jointly train the analysis model. In some embodiments, the value of the loss function is determined based on a comparison with reference data for the electrolyte sample, the reference data being generated based on density functional theory (DFT).
[0074] Optimization of the analysis model
[0075] In machine learning force fields, density functional theory (DFT) is typically used to generate data. In electrolyte simulations, the number of atoms in the system is usually over 10,000. However, DFT, due to its high time complexity, cannot calculate systems with tens of thousands of atoms. Therefore, the common practice is to extract a smaller cluster (usually around 100 atoms) from the simulated system and use DFT for calculation. This introduces another problem: in smaller clusters, the magnitude of intermolecular forces is much smaller than that of intramolecular forces. This causes MLFF to focus on learning intramolecular forces, resulting in insufficient precision in characterizing intermolecular forces and inaccurate predictions in larger systems. The following will refer to... Figure 4 This describes the process of optimizing an analytical model using limited experimental data according to embodiments of the present disclosure.
[0076] Figure 4 A flowchart illustrating an example process 400 of an optimization analysis model according to some embodiments of the present disclosure is shown. Process 400 can be implemented at electronic device 110. Reference is made below. Figure 1 To describe process 400.
[0077] In box 410, electronic device 110 uses a first set of values for a first property of the target substance to fine-tune the analysis model to determine a second set of values for a second property of the target substance.
[0078] In some embodiments, the target substance may include, for example, a solution. In some cases, experimental data for solutions are relatively scarce compared to experimental data for pure solvents. The inventors of this application have experimentally discovered that, using a small amount of experimental data for pure solvents, predictions can be extrapolated to solutions with multiple species and ratios.
[0079] In some embodiments, the first property and the second property can be properties of the target substance that have a predetermined correlation, wherein the first property is a property that can be predicted by MLFF, and the second property is a property that is relatively easy to obtain and has high measurement accuracy.
[0080] The following explanation will use solution pressure as the first property and density as the second property. It should be understood that such first and second properties can also include other suitable examples, such as the dipole moment of the solution as the first property and the conductivity of the solution as the second property.
[0081] There exists a linear relationship between the density and pressure of a liquid. As an example, this linear relationship can be expressed by the following equation:
[0082]
[0083] Where β represents the compressibility of the liquid, ρ represents the density, and P represents the pressure. Therefore, for a solution, its correct density is ρ0.
[0084] Therefore, the electronic device 110 can first use an unoptimized analysis model and determine the density ρ1 of the first round based on the NPT system, the corresponding pressure of which can be, for example, P1. Here, P1 can be, for example, a preset pressure value, that is, a preset value for the first property (pressure).
[0085] Furthermore, the electronic device 110 can determine a first set of values regarding the first property. For example, the electronic device 110 can determine the first set of values based on a preset value (e.g., P1) regarding the first property (i.e., pressure) and at least one change (e.g., at least one dP).
[0086] Figure 5 A schematic diagram 500 is shown illustrating a fine-tuning analysis model according to some embodiments of the present disclosure. For example... Figure 5 As shown, multiple dP values can be 2000, 4000, and 6000. Correspondingly, the first set of values can be represented as P1+2000, P1+4000, and P1+6000.
[0087] Accordingly, the electronic device 110 can use the determined first set of values (P1+2000, P1+4000 and P1+6000) for the first property to fine-tune the model, thereby determining a set of fine-tuned intermediate models.
[0088] Furthermore, the electronic device 110 can utilize a set of intermediate models to determine a second set of values for the target substance with respect to the second property. For example, the electronic device 110 can use a fine-tuned MLFF and determine the density of the second round, e.g., ρ2, according to the NPT system.
[0089] In box 420, electronic device 110 determines the correlation between the first and second properties of the target substance based on the first set of values and the second set of values.
[0090] Taking density and pressure as an example, electronic device 110 can perform linear fitting on at least the first set of values and the second set of values to determine the correlation between the first and second properties of the target substance. It should be understood that the method for determining the correlation may differ depending on the properties. Figure 5 As an example, electronic device 110 can perform linear fitting on dP and d(lnρ).
[0091] In box 430, electronic device 110 determines the target value of the target substance with respect to the first property based on a reference value of the target substance with respect to the second property, using correlation. The reference value is determined based on experiments on the target substance.
[0092] Furthermore, after determining the correlation, the electronic device 110 can determine the target value of the pressure, or the target dP, based on a reference value determined experimentally (e.g., the correct density ρ0).
[0093] In box 440, electronic device 110 optimizes the analysis model using the target value of the target substance with respect to the first property. For example, electronic device 110 can use the target pressure value to further fine-tune the analysis model.
[0094] In some embodiments, after two rounds of fine-tuning, the electronic device 110 can use the fine-tuned analytical model to determine a first value of the target material with respect to the second property, namely a new density prediction.
[0095] In some embodiments, if the difference between a first value (e.g., a new density prediction) and a reference value (e.g., the correct density ρ0) is less than or equal to a first threshold, the electronic device 110 determines the fine-tuned analysis model as the optimized analysis model.
[0096] Conversely, if the difference between the first value and the reference value is greater than a first threshold, the electronic device 110 can continue to fine-tune the model. Specifically, the electronic device 110 can determine a second correlation between the first and second properties of the target substance based on a first set of values, a second set of values, a preset value for the first property (e.g., P1), and a second value for the second property (e.g., ρ1). For example, the electronic device 110 can use the two rounds of prediction results to perform linear fitting to improve the accuracy of linear fitting.
[0097] Similarly, electronic device 110 can determine a second target value of the target substance with respect to the first property, such as a new target pressure value, based on a reference value of the target substance with respect to the second property and utilizing a second correlation. Accordingly, electronic device 110 can optimize the analysis model using the second target value of the target substance with respect to the first property (e.g., the new target pressure value).
[0098] Based on the above methods, the embodiments of this disclosure can optimize the analysis model using limited experimental data.
[0099] Knowledge distillation
[0100] Figure 6 A flowchart illustrating an example process 600 for constructing a ligand generation model according to some embodiments of the present disclosure is shown. Process 600 can be implemented at an electronic device 110. Reference is made below. Figure 1 To describe process 600.
[0101] In box 610, electronic device 110 acquires multiple analytical models for analyzing the target substance, the multiple analytical models being trained based on different initial values of model parameters.
[0102] In some embodiments, the electronic device 110 can train multiple analysis models based on different random seeds.
[0103] In box 620, electronic device 110 uses multiple analytical models to determine multiple predicted results regarding the forces acting on a group of atoms in the target material.
[0104] In box 630, electronic device 110 determines target data on the forces acting on a set of atoms based on multiple predictions.
[0105] Specifically, the electronic device 110 can determine the average and / or standard deviation of multiple prediction results, and can determine target data of the force applied to a set of atoms based on the average and / or standard deviation of multiple prediction results.
[0106] In box 640, electronic device 110 uses target data to train a target analysis model for analyzing target substances.
[0107] In some embodiments, the electronic device 110 can train a target analysis model such that: the target analysis model's target prediction result for the forces acting on a set of atoms is close to the average of multiple prediction results; and / or the target analysis model's standard deviation prediction result for the forces acting on a set of atoms is close to the standard deviation of multiple prediction results.
[0108] Therefore, by utilizing the mean and standard deviation, embodiments of this disclosure can achieve knowledge distillation, enabling a single model to possess predictive power and uncertainty estimation capabilities approaching those of multiple models as a whole.
[0109] In some embodiments, considering that the standard deviation indicates the uncertainty of multiple models for the prediction results, if the standard deviation is greater than a fourth threshold, the electronic device 110 can also acquire reference data on the forces acting on the target material and can train a target analysis model based on the reference data.
[0110] For example, if the standard deviation of predictions for a new sample from multiple models is large, it indicates that the analysis results of the multiple models have high uncertainty. Therefore, the electronic device 110 can further acquire reference data about the sample to continue training the multiple models. In some embodiments, such reference data may be generated, for example, based on density functional theory (DFT).
[0111] In some embodiments, the target analysis model to be trained may be, for example, a new model, or at least one of a plurality of analysis models trained based on different initial values of model parameters.
[0112] Experimental results
[0113] The following will combine Figures 7A to 7I This describes a comparison of analytical models (both optimized and unoptimized) and experimental data based on this disclosure. In the figure, DMC, EC, EMC, and MA represent the solvent names of dimethyl carbonate, enthylenecarbonate, etheyl methyl carbonate, and methyl acetate, respectively.
[0114] Figures 7A to 7C A comparison of the viscosity, density, and diffusion coefficient of pure DMC solvent is shown. Figures 7A to 7C The study presents the effect of optimized and unoptimized models on the viscosity of DMC as a pure solvent at different temperatures. Figure 7A ),density( Figure 7B ) and diffusion coefficient ( Figure 7CThe prediction results are shown in the figure. A significant difference can be clearly observed between the optimized and unoptimized results. Although the model optimization is based solely on the density of DMC, a significant improvement in the prediction accuracy for the viscosity and diffusion coefficient of DMC can also be observed in the figure.
[0115] Figures 7D to 7F Viscosity predictions for EC-EMC mixed solvents are shown. Figures 7D to 7F The results show the viscosity predictions of different ratios of the binary mixed solvent EC-EMC at different temperatures using optimized and unoptimized models. Figure 7D This indicates that the weight ratio of EC-EMC is 30:70; Figure 7E It is 20:80; Figure 7F The ratio is 10:90. It is evident that the optimized model better matches the experimental results across different solvent ratios and temperature ranges, whereas this accuracy is not significant in the unoptimized model. Therefore, based on model optimization using only a limited number of pure solvent physical property data, the disclosed scheme demonstrates excellent generalization ability across mixed solvents and temperature ranges.
[0116] Figures 7G to 7I The model shows predictions of solvent MA properties that were not seen before. Figures 7G to 7I This demonstrates the model's predictive ability on the unseen solvent MA. This model for predicting MA properties has not been optimized for the experimental physical properties of MA. Here, optimization refers to optimization on solvents with structures similar to MA. Figure 7G The viscosity of MA is shown. Figure 7H For density, Figure 7I The diffusion coefficient is denoted as . It can be seen that although the model has never encountered MA molecules and has not been optimized for MA, it has been able to grasp the intermolecular and intramolecular interactions of MA by learning from molecules with similar structures. This enables the model to make accurate predictions regarding density and exhibit excellent generalization ability regarding viscosity and diffusion coefficient.
[0117] Example devices and equipment
[0118] Embodiments of this disclosure also provide corresponding apparatus for implementing the above methods or processes. Figure 8A A schematic structural block diagram of an apparatus 800A for optimizing an analysis model according to some embodiments of the present disclosure is shown. The apparatus 800A may be implemented as or included in an electronic device 110. The various modules / components in the apparatus 800A may be implemented by hardware, software, firmware, or any combination thereof.
[0119] like Figure 8AAs shown, the apparatus 800A includes: a fine-tuning module 802 configured to fine-tune an analytical model using a first set of values relating to a first property of a target substance to determine a second set of values relating to a second property of the target substance; a first determination module 804 configured to determine the correlation between the first property and the second property of the target substance based on the first set of values and the second set of values; a second determination module 806 configured to determine a target value relating to the first property of the target substance using the correlation based on a reference value relating to the second property of the target substance, the reference value being determined based on experiments on the target substance; and an optimization module 808 configured to optimize the analytical model using the target value relating to the first property of the target substance.
[0120] Embodiments of this disclosure also provide corresponding apparatus for implementing the above methods or processes. Figure 8A A schematic structural block diagram of an apparatus 800A for optimizing an analysis model according to some embodiments of the present disclosure is shown. The apparatus 800A may be implemented as or included in an electronic device 110. The various modules / components in the apparatus 800A may be implemented by hardware, software, firmware, or any combination thereof.
[0121] like Figure 8A As shown, the apparatus 800A includes: a fine-tuning module 802 configured to fine-tune an analytical model using a first set of values relating to a first property of a target substance to determine a second set of values relating to a second property of the target substance; a first determination module 804 configured to determine the correlation between the first property and the second property of the target substance based on the first set of values and the second set of values; a second determination module 806 configured to determine a target value relating to the first property of the target substance using the correlation based on a reference value relating to the second property of the target substance, the reference value being determined based on experiments on the target substance; and an optimization module 808 configured to optimize the analytical model using the target value relating to the first property of the target substance.
[0122] In some embodiments, the fine-tuning module 802 is further configured to: fine-tune the analysis model using a first set of values for the first property to determine a set of intermediate models; and use the set of intermediate models to determine a second set of values for the target substance with respect to the second property.
[0123] In some embodiments, the first determining module 804 is further configured to perform linear fitting on at least the first set of values and the second set of values to determine the correlation between the first property and the second property of the target substance.
[0124] In some embodiments, the optimization module 808 is further configured to: fine-tune the analysis model using the target value; determine a first value of the target substance with respect to the second property using the fine-tuned analysis model; and determine the fine-tuned analysis model as the optimized analysis model if the difference between the first value and the reference value is less than or equal to a first threshold.
[0125] In some embodiments, the correlation is a first correlation, the target value is a first target value, and the optimization module 808 is further configured to: if the difference between the first value and the reference value is greater than a first threshold, determine a second correlation between the first property and the second property of the target substance based on a first set of values, a second set of values, a preset value for the first property, and a second value for the second property, wherein the second value corresponds to the preset value of the first property; determine a second target value of the target substance for the first property based on the reference value of the target substance for the second property using the second correlation; and optimize the analysis model using the second target value of the target substance for the first property.
[0126] In some embodiments, the first set of values is determined based on a preset value for the first property and at least one variable.
[0127] In some embodiments, the second value is determined by the analysis model based on the NPT ensemble.
[0128] In some embodiments, the target substance includes a solution, the first property being the pressure of the solution and the second property being the density of the solution.
[0129] In some embodiments, the target substance includes a solution, the first property being the dipole moment of the solution, and the second property being the conductivity of the solution.
[0130] Embodiments of this disclosure also provide corresponding apparatus for implementing the above methods or processes. Figure 8B A schematic structural block diagram of an apparatus 800B for training an analysis model according to some embodiments of the present disclosure is shown. The apparatus 800B may be implemented as or included in an electronic device 110. The various modules / components in the apparatus 800B may be implemented by hardware, software, firmware, or any combination thereof.
[0131] like Figure 8B As shown, the device 800B includes: a feature determination module 812 configured to determine a set of feature representations of a set of atoms in an electrolyte sample using an analysis model; a charge determination module 814 configured to determine charge information of the electrolyte sample based on the set of feature representations using an analysis model; an energy determination module 816 configured to determine energy information of the electrolyte sample based on the charge information and the set of feature representations using an analysis model; and a first training module 818 configured to train the analysis model based on the charge information and the energy information, wherein the loss function used to train the analysis model includes at least a first part and / or a second part, the first part being associated with the multipole moment determined based on the charge information, and the second part being associated with the stress information determined based on the energy information.
[0132] In some embodiments, the loss function further includes a third portion, which is associated with charge information and / or a dipole moment determined based on the charge information.
[0133] In some embodiments, the stress information is a virial tensor determined based on energy information.
[0134] In some embodiments, the value of the loss function is determined based on a comparison with reference data for the electrolyte sample, which is generated based on density functional theory (DFT).
[0135] In some embodiments, the feature determination module 812 is further configured to: determine a graph representation corresponding to a set of atoms in an electrolyte sample, wherein a plurality of nodes in the graph representation correspond to a set of atoms, and an edge in the graph representation indicates that the distance between two corresponding atoms is less than a second threshold; and determine a set of feature representations of a set of atoms in the electrolyte sample by processing the graph representation using a graph neural network of an analysis model.
[0136] In some embodiments, the feature determination module 812 is further configured to: determine a first feature representation for a first atom in a set of atoms; determine at least one second atom whose distance to the first atom is less than a third threshold based on the distance from the set of atoms to the first atom; update the first feature representation of the first atom using the feature representation of the at least one second atom; and determine a set of feature representations for the set of atoms based on the updated first feature representation.
[0137] Embodiments of this disclosure also provide corresponding apparatus for implementing the above methods or processes. Figure 8C A schematic structural block diagram of an apparatus 800C for training an analysis model according to some embodiments of the present disclosure is shown. The apparatus 800C may be implemented as or included in an electronic device 110. Various modules / components in the apparatus 800C may be implemented by hardware, software, firmware, or any combination thereof.
[0138] like Figure 8C As shown, the device 800C includes: a model acquisition module 822 configured to acquire multiple analytical models for analyzing a target substance, the multiple analytical models being trained based on different initial values of model parameters; a result prediction module 824 configured to use the multiple analytical models to determine multiple predicted results regarding the forces acting on a group of atoms in the target substance; a third determination module 826 configured to determine target data for the forces acting on a group of atoms based on the multiple predicted results; and a second training module 828 configured to use the target data to train a target analytical model for analyzing the target substance.
[0139] In some embodiments, the third determining module 826 is further configured to: determine the average and / or standard deviation of a plurality of prediction results; and determine target data of the forces acting on a set of atoms based on the average and / or standard deviation of the plurality of prediction results.
[0140] In some embodiments, the second training module 828 is further configured to: train the target analysis model such that: the target analysis model's target prediction result for the force on a set of atoms is close to the average of multiple prediction results; and / or the target analysis model's standard deviation prediction result for the force on a set of atoms is close to the standard deviation of multiple prediction results.
[0141] In some embodiments, the device 800C further includes a third training module configured to: if the standard deviation is greater than a fourth threshold, acquire reference data on the forces acting on the target material, and train a target analysis model based on the reference data.
[0142] In some embodiments, the reference data is generated based on density functional theory (DFT).
[0143] In some embodiments, the target analysis model includes at least one of a plurality of analysis models.
[0144] Figure 9 A block diagram of an electronic device 900 that may implement one or more embodiments of the present disclosure is shown. It should be understood that... Figure 9 The electronic device 900 shown is merely exemplary and should not be construed as limiting the functionality and scope of the embodiments described herein. Figure 9 The electronic device 900 shown can be used to achieve Figure 1 Electronic devices 110.
[0145] like Figure 9 As shown, electronic device 900 is in the form of a general-purpose computing device. Components of electronic device 900 may include, but are not limited to, one or more processors or processing units 910, memory 920, storage device 930, one or more communication units 940, one or more input devices 950, and one or more output devices 960. Processing unit 910 may be a physical or virtual processor and is capable of performing various processes according to programs stored in memory 920. In a multiprocessor system, multiple processing units execute computer-executable instructions in parallel to improve the parallel processing capability of electronic device 900.
[0146] Electronic device 900 typically includes multiple computer storage media. Such media can be any available media accessible to electronic device 900, including but not limited to volatile and non-volatile media, removable and non-removable media. Memory 920 can be volatile memory (e.g., registers, cache, random access memory (RAM)), non-volatile memory (e.g., read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory), or some combination thereof. Storage device 930 can be removable or non-removable media and can include machine-readable media, such as flash drives, disks, or any other media that can be used to store information and / or data (e.g., training data for training) and can be accessed within electronic device 900.
[0147] Electronic device 900 may further include additional removable / non-removable, volatile / non-volatile storage media. Although not explicitly stated... Figure 9 As shown, disk drives for reading from or writing to removable, non-volatile disks (e.g., "floppy disks") and optical disk drives for reading from or writing to removable, non-volatile optical disks can be provided. In these cases, each drive can be connected to a bus (not shown) via one or more data media interfaces. Memory 920 may include computer program product 925 having one or more program modules configured to perform various methods or actions of various embodiments of this disclosure.
[0148] The communication unit 940 enables communication with other electronic devices via a communication medium. Additionally, the functionality of the components of the electronic device 900 can be implemented using a single computing cluster or multiple computing machines capable of communicating via communication connections. Therefore, the electronic device 900 can operate in a networked environment using logical connections to one or more other servers, network personal computers (PCs), or another network node.
[0149] Input device 950 can be one or more input devices, such as a mouse, keyboard, trackball, etc. Output device 960 can be one or more output devices, such as a monitor, speaker, printer, etc. Electronic device 900 can also communicate with one or more external devices (not shown) via communication unit 940 as needed. These external devices include storage devices, display devices, etc., and can communicate with one or more devices that enable user interaction with electronic device 900, or with any device that enables electronic device 900 to communicate with one or more other electronic devices (e.g., network card, modem, etc.). Such communication can be performed via input / output (I / O) interface (not shown).
[0150] According to exemplary embodiments of the present disclosure, a computer-readable storage medium is provided that stores computer-executable instructions thereon, wherein the computer-executable instructions are executed by a processor to implement the methods described above. According to exemplary embodiments of the present disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, which are executed by a processor to implement the methods described above.
[0151] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, devices, and computer program products implemented according to this disclosure. It should 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-readable program instructions.
[0152] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.
[0153] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.
[0154] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0155] Various implementations of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed implementations. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described implementations. The terminology used herein is chosen to best explain the principles, practical applications, or improvements to technology in the market, or to enable others skilled in the art to understand the various embodiments disclosed herein.
Claims
1. A method for optimizing an analysis model, comprising: The analysis model is fine-tuned using a first set of values for a first property of a target substance to determine a second set of values for a second property of the target substance, the target substance including a solution, the first property and the second property being properties of the solution with a predetermined correlation, and the value of the first property being predictable by a machine learning force field, and the value of the second property being obtainable by measurement; Based on the first set of values and the second set of values, the correlation between the first property and the second property of the target substance is determined; Based on a reference value of the target substance with respect to the second property, a target value of the target substance with respect to the first property is determined using the correlation, wherein the reference value is determined based on experiments on the target substance; and The analytical model is optimized using the target value of the target substance with respect to the first property.
2. The method of claim 1, wherein fine-tuning the analysis model using a first set of values related to the first property comprises: The analysis model is fine-tuned using a first set of values related to the first property to determine a set of intermediate models; as well as The second set of values for the target substance with respect to the second property is determined using the aforementioned set of intermediate models.
3. The method according to claim 1, wherein determining the correlation between the first property and the second property of the target substance comprises: At least a linear fit is performed on the first set of values and the second set of values to determine the correlation between the first property and the second property of the target substance.
4. The method according to claim 1, wherein optimizing the analytical model using the target value of the target substance with respect to the first property comprises: The analysis model is fine-tuned using the target value; The first value of the target substance with respect to the second property is determined using the finely tuned analytical model; as well as If the difference between the first value and the reference value is less than or equal to the first threshold, the fine-tuned analysis model is determined as the optimized analysis model.
5. The method according to claim 4, wherein the correlation is a first correlation, the target value is a first target value, and optimizing the analytical model using the target value of the target substance with respect to the first property further comprises: If the difference between the first value and the reference value is greater than the first threshold, a second correlation between the first property and the second property of the target substance is determined based on the first set of values, the second set of values, a preset value for the first property, and a second value for the second property, wherein the second value corresponds to the preset value for the first property; Based on the reference value of the target substance with respect to the second property, the second target value of the target substance with respect to the first property is determined using the second correlation; as well as The analytical model is optimized using the second target value of the target substance with respect to the first property.
6. The method of claim 5, wherein the first set of values is determined based on the preset value and at least one variable related to the first property.
7. The method according to claim 6, wherein the second value is determined by the analysis model based on the NPT ensemble.
8. The method according to claim 1, wherein the target substance comprises a solution, the first property being the pressure of the solution, and the second property being the density of the solution.
9. The method according to claim 1, wherein the target substance comprises a solution, the first property being the dipole moment of the solution, and the second property being the conductivity of the solution.
10. An apparatus for optimizing an analysis model, comprising: The fine-tuning module is configured to fine-tune the analysis model using a first set of values for a first property of a target substance to determine a second set of values for the target substance with respect to a second property, the target substance including a solution, the first property and the second property being properties of the solution having a predetermined correlation, and the value of the first property being predictable by a machine learning force field, and the value of the second property being obtainable by measurement; The first determining module is configured to determine the correlation between the first property and the second property of the target substance based on the first set of values and the second set of values; The second determining module is configured to determine a target value of the target substance with respect to the first property based on a reference value of the target substance with respect to the second property, using the correlation, wherein the reference value is determined based on experiments on the target substance; as well as The optimization module is configured to optimize the analysis model using the target value of the target substance with respect to the first property.
11. An electronic device, comprising: At least one processing unit; as well as At least one memory, coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, which, when executed by the at least one processing unit, cause the electronic device to perform the method according to any one of claims 1 to 9.
12. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method according to any one of claims 1 to 9.