Method and system for predicting catalyst performance and method and system for training network

Abstract

The embodiment of the invention provides a catalyst performance prediction method and system and a network training method and system. In the prediction method, a prediction system generates a plurality of core descriptors and a plurality of non-core descriptors for describing a target catalyst based on proportion information corresponding to each element in the target catalyst and attribute values of a plurality of attributes, and inputting at least part of the descriptors into the target network to obtain the performance index of the target performance of the target catalyst. The target network is a network obtained by training a plurality of sample catalysts by a training system, and comprises a plurality of sub-networks which have the same structure and are arranged according to levels, each sub-network comprises N input nodes and one output node, the output node of the ith level sub-network is used as the input node of the (i + 1) th level sub-network, and the output node of the (i + 1) th level sub-network is used as the input node of the (i + 1) th level sub-network; n core descriptors input into each sub-network in the first hierarchy are the same, N-N1 non-core descriptors are not completely the same, both N and N1 are greater than 1, and i is greater than or equal to 1.

Description

[0001] This application claims priority to Chinese patent application No. 2024111718620, filed on August 23, 2024, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This specification relates to the field of chemical materials, and in particular to a method and system for predicting catalyst performance, and a method and system for training networks. Background Technology

[0003] The oxygen evolution reaction (OER) involves a complex four-electron transfer process and is relatively time-consuming. Therefore, catalysts are needed to improve OER efficiency in reactions such as water splitting and metal-air batteries. Traditional noble metal catalysts, such as IrO2 and RuO2, exhibit good catalytic activity, but their scarcity and high cost limit their widespread application. Therefore, researchers are dedicated to developing non-noble metal OER electrocatalysts to reduce costs and improve catalyst sustainability.

[0004] Non-precious metal OER electrocatalysts have attracted much attention due to their multiple active sites and associated synergistic effects, such as multi-element materials (e.g., ternary materials, quaternary materials, and even high-entropy materials). As the number of constituent elements in the catalyst increases, the difficulty of studying its catalytic performance also gradually increases. How to accurately predict the catalytic performance of catalysts has become a problem that needs to be solved in catalyst research.

[0005] The information in the background section is merely information known only to the inventor and does not imply that such information had entered the public domain before the date of this application, nor does it imply that it can be considered prior art in this disclosure. Summary of the Invention

[0006] This specification provides a method and system for predicting catalyst performance, as well as a method and system for training a network, which improves the accuracy of predicting catalyst performance and solves the problem that the network training cannot achieve the expected results due to insufficient training samples but a large number of catalyst descriptors.

[0007] Firstly, this specification provides a method for predicting catalyst performance, configured to predict the target performance of a target catalyst, the target catalyst comprising at least one element. The method includes: generating K descriptors to describe the target catalyst based on the proportion information of the at least one element in the target catalyst and attribute values ​​of X attributes corresponding to each of the at least one element; determining N1 core descriptors and K-N1 non-core descriptors from the K descriptors, wherein the core descriptors have a greater influence on the target performance than the non-core descriptors; and inputting at least a portion of the K descriptors into a target network for prediction to obtain performance indicators corresponding to the target performance, wherein the target network... The target network is a network trained using Y sample catalysts. The target network includes multiple sub-networks with the same structure and arranged hierarchically. Each sub-network includes N input nodes and 1 output node. The output node of the sub-network in the i-th level serves as the input node of the sub-network in the (i+1)-th level. Each sub-network in the first level is input with the N1 core descriptors and N-N1 non-core descriptors. The N-N1 non-core descriptors are a subset of the K-N1 non-core descriptors, and the non-core descriptors input to different sub-networks are not completely identical. The ratio of N to Y is less than 1, and the ratio of K to Y is greater than 1. Wherein, X, K, N1, Y, and N are all integers greater than 1, and i is an integer greater than or equal to 1.

[0008] In some embodiments, the X properties include at least two of the following dimensions: periodic table dimension; thermodynamic dimension; physical dimension; or crystallographic dimension.

[0009] In some embodiments, the K descriptors include at least one of the following: a first set of descriptors for describing the statistical characteristics of the target catalyst in the X properties; a second set of descriptors for describing the valence electron occupation state of the target catalyst; a third set of descriptors for describing the ionicity of the target catalyst; and a fourth set of descriptors for describing the configuration entropy of the target catalyst.

[0010] In some embodiments, the first set of descriptors is generated in the following manner:

[0011] For each of the X attributes, based on the proportion information of the at least one element in the target catalyst and the attribute value of the at least one element in the attribute, P statistical values ​​are obtained from multiple statistical dimensions to statistically analyze the characteristics of the target catalyst in the attribute; and based on the P statistical values ​​of the target catalyst in each of the X attributes, a first set of descriptors is generated, the number of the first set of descriptors being equal to X*P.

[0012] In some embodiments, the second set of descriptors includes: occupancy information of Q valence electrons in the target catalyst, wherein the occupancy information of the z-th valence electron is obtained in the following manner:

[0013] For each of the at least one element, the number of each valence electron and the total number of valence electrons in the element are determined based on the attribute values ​​of at least some of the X attributes corresponding to the element; and for the z-th valence electron, the occupation information of the z-th valence electron is determined based on the proportion information of the at least one element in the target catalyst, the number of the z-th valence electrons corresponding to each of the at least one element, and the total number of valence electrons corresponding to each of the at least one element.

[0014] In some embodiments, the third set of descriptors is obtained based on the proportion information of the at least one element in the target catalyst, the electronegativity of each of the at least one element, and any of the following: the maximum value of the electronegativity of each of the at least one element; or the average value of the electronegativity of each of the at least one element.

[0015] In some embodiments, the fourth set of descriptors is obtained based on the proportion of the at least one element in the target catalyst, the Boltzmann constant, and room temperature.

[0016] In some embodiments, the target performance is the Tafel slope, and the N1 core descriptors include multiples of the following: a descriptor representing the average deviation of the electronegativity of the different elements in the at least one element; a descriptor representing the maximum value of the weighted attribute values ​​of the at least one element over the BCC atomic energy difference; a descriptor representing the minimum value of the weighted attribute values ​​of the at least one element over the row in the periodic table; a descriptor representing the maximum value of the deviation of the different elements in the BCC atomic energy difference; and a descriptor representing the average deviation of the different elements in the at least one element over the work function.

[0017] In some embodiments, the target performance is the initial overpotential, and the N1 core descriptors include multiples of the following: a descriptor representing the maximum value of the weighted attribute values ​​of the at least one element in terms of the number of unfilled valence electrons (d); a descriptor representing the average value of the weighted attribute value deviations of different elements in terms of mass density; a descriptor representing the minimum value of the weighted attribute value deviations of different elements in terms of work function; a descriptor representing the average value of the deviations of different elements in terms of the number of filled valence electrons (d); and a descriptor representing the maximum value of the weighted attribute value deviations of different elements in terms of covalent radius.

[0018] In some embodiments, the target network includes F layers, and the step of inputting at least a portion of the K descriptors into the target network for prediction to obtain the performance metric corresponding to the target performance includes:

[0019] Based on the K descriptors, generate m1 sets of descriptors, each set including the N1 core descriptors and N-N1 non-core descriptors; input the m1 sets of descriptors into the m1 sub-networks of the first layer of the target network to obtain m1 intermediate performance indicators of the target performance; according to the order of i from 2 to F, the m1 sub-networks of the (i-1)th layer of the target network... 1-1 The target performance m obtained from the subnetwork i-1 Each intermediate performance metric is divided into m i The m-th layer of the target network is then input after the group. i A sub-network is used to obtain m of the target performance. i An intermediate performance metric; and the intermediate performance metric of the target performance output by the sub-network of layer F as the final performance metric of the target performance; wherein, m1, F, and m i-1 All are integers greater than 1, and m i It is an integer greater than or equal to 1.

[0020] Secondly, this specification provides a training method for a target network used to predict the target performance of a catalyst. The method includes: obtaining experimental values ​​of Y sample catalysts for the target performance, each sample catalyst including at least one element; for each sample catalyst, generating K descriptors to describe the sample catalyst based on the proportion of the at least one element in the sample catalyst and the attribute values ​​of X attributes corresponding to the at least one element, and determining N1 core descriptors and K-N1 non-core descriptors from the K descriptors, wherein the core descriptors have a greater influence on the target performance than the non-core descriptors; inputting a portion of the K descriptors corresponding to each sample catalyst into the target network for prediction, to obtain the predicted performance corresponding to the target performance. The measurement values ​​are as follows: the target network includes multiple sub-networks with the same structure and arranged hierarchically. Each sub-network includes N input nodes and 1 output node. The output node of the sub-network in the i-th level serves as the input node of the sub-network in the (i+1)-th level. Each sub-network in the first level is input with the N1 core descriptors and N-N1 non-core descriptors. The N-N1 non-core descriptors are a subset of the K-N1 non-core descriptors, and the non-core descriptors input to different sub-networks are not completely identical. The ratio of N to Y is less than 1, and the ratio of K to Y is greater than 1. The parameters of the target network are updated with the training objective of minimizing the difference between the predicted value and the experimental value. Wherein, X, K, N1, Y, and N are all integers greater than 1, and i is an integer greater than or equal to 1.

[0021] In some embodiments, the X properties include at least two of the following dimensions: periodic table dimension; thermodynamic dimension; physical dimension; or crystallographic dimension.

[0022] In some embodiments, the K descriptors include at least one of the following: a first set of descriptors for describing the statistical characteristics of the sample catalyst in the X properties; a second set of descriptors for describing the valence electron occupation state of the sample catalyst; a third set of descriptors for describing the ionicity of the sample catalyst; and a fourth set of descriptors for describing the configuration entropy of the sample catalyst.

[0023] In some embodiments, the first set of descriptors is generated in the following manner:

[0024] For each of the X attributes, based on the proportion information of the at least one element in the sample catalyst and the attribute value of the at least one element in the attribute, P statistical values ​​are obtained from multiple statistical dimensions to statistically analyze the characteristics of the sample catalyst in the attribute; and based on the P statistical values ​​of the sample catalyst in each of the X attributes, a first set of descriptors is generated, the number of the first set of descriptors being equal to X*P.

[0025] In some embodiments, the second set of descriptors includes: occupancy information of Q valence electrons in the sample catalyst, wherein the occupancy information of the z-th valence electron is obtained in the following manner:

[0026] For each of the at least one element, based on the attribute values ​​of at least some of the X attributes corresponding to the element, the number of each valence electron and the total number of valence electrons in the element are determined; and for the z-th valence electron, based on the proportion information of the at least one element in the sample catalyst, the number of the z-th valence electrons corresponding to each of the at least one element, and the total number of valence electrons corresponding to each of the at least one element, the occupancy information of the z-th valence electron is determined.

[0027] In some embodiments, the third set of descriptors is obtained based on the proportion information of the at least one element in the sample catalyst, the electronegativity of each of the at least one element, and any of the following: the maximum value of the electronegativity of the at least one element; or the average value of the electronegativity of the at least one element.

[0028] In some embodiments, the fourth set of descriptors is obtained based on the proportion of the at least one element in the target catalyst, the Boltzmann constant, and room temperature.

[0029] In some embodiments, the target performance is the Tafel slope, and the N1 core descriptors include multiples of the following: a descriptor representing the average deviation of the electronegativity of the different elements in the at least one element; a descriptor representing the maximum value of the weighted attribute values ​​of the at least one element over the BCC atomic energy difference; a descriptor representing the minimum value of the weighted attribute values ​​of the at least one element over the row in the periodic table; a descriptor representing the maximum value of the deviation of the different elements in the BCC atomic energy difference; and a descriptor representing the average deviation of the different elements in the at least one element over the work function.

[0030] In some embodiments, the target performance is the initial overpotential, and the N1 core descriptors include multiples of the following: a descriptor representing the maximum value of the weighted attribute values ​​of the at least one element in terms of the number of unfilled valence electrons (d); a descriptor representing the average value of the weighted attribute value deviations of different elements in terms of mass density; a descriptor representing the minimum value of the weighted attribute value deviations of different elements in terms of work function; a descriptor representing the average value of the deviations of different elements in terms of the number of filled valence electrons (d); and a descriptor representing the maximum value of the weighted attribute value deviations of different elements in terms of covalent radius.

[0031] In some embodiments, the target network includes F levels, and the step of inputting a portion of the K descriptors corresponding to each sample catalyst into the target network for prediction to obtain the predicted value corresponding to the target performance includes:

[0032] Based on the K descriptors, generate m1 sets of descriptors, each set including the N1 core descriptors and N-N1 non-core descriptors; input the m1 sets of descriptors into the m1 subnetworks of the first layer of the target network to obtain m1 intermediate prediction values ​​of the target performance; according to the order of i from 2 to F, the m1 subnetworks of the (i-1)th layer of the target network... 1-1 The target performance m obtained from the subnetwork i-1 The intermediate predicted values ​​are divided into m i After grouping, input the m of the i-th layer of the target network. i A sub-network is used to obtain m of the target performance. i An intermediate predicted value; and the intermediate predicted value of the target performance output by the sub-network of layer F as the final predicted value of the target performance; wherein, m1, F, and m i-1 All are integers greater than 1, and m i It is an integer greater than or equal to 1.

[0033] Thirdly, this specification also provides a prediction system configured to predict the target performance of a target catalyst, comprising: at least one storage medium storing at least one instruction set; and at least one processor communicatively connected to the at least one storage medium, wherein, when the prediction system is running, the at least one processor reads the at least one instruction set and executes the prediction method described in the first aspect according to the instructions of the at least one instruction set.

[0034] Fourthly, this specification also provides a training system configured to train a target network for predicting the target performance of a catalyst, comprising: at least one storage medium storing at least one instruction set; and at least one processor communicatively connected to the at least one storage medium, wherein, when the training system is running, the at least one processor reads the at least one instruction set and executes the training method described in the second aspect according to the instructions of the at least one instruction set.

[0035] The methods and systems for predicting catalyst performance, the methods and systems for training networks, and other functionalities of the systems provided in this specification will be partially listed in the following description. The inventive aspects of the methods and systems for predicting catalyst performance, the methods and systems for training networks, and the systems provided in this specification can be fully explained through practice or by using the methods, systems, and combinations described in the detailed examples below. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A schematic diagram illustrating an application scenario for predicting catalyst performance according to embodiments of this specification is shown.

[0038] Figure 2 A schematic diagram of the hierarchical structure of a target network provided according to an embodiment of this specification is shown;

[0039] Figure 3 A schematic diagram of the structure of a subnetwork provided according to an embodiment of this specification is shown;

[0040] Figure 4 A hardware structure diagram of a system provided according to an embodiment of this specification is shown;

[0041] Figure 5 A flowchart of a training method for a target network according to an embodiment of this specification is shown;

[0042] Figure 6 A schematic diagram of the GA iteration process provided according to an embodiment of this specification is shown;

[0043] Figure 7 A schematic diagram showing the relationship between the core descriptor corresponding to the Tafel slope and the Tafel slope provided in the embodiments of this specification is shown.

[0044] Figure 8A schematic diagram showing the relationship between the core descriptor and the initial overpotential provided in the embodiments of this specification is shown.

[0045] Figure 9 A schematic diagram illustrating the inputs and outputs between the hierarchical structures of the target network provided according to an embodiment of this specification is shown.

[0046] Figure 10 A schematic diagram illustrating the performance of target networks T1 and T2 during the training process according to embodiments of this specification is shown.

[0047] Figure 11 A schematic diagram illustrating the performance of target networks T3 and T4 during the training process according to embodiments of this specification is shown.

[0048] Figure 12 Experimental and predicted values ​​of Tafel slopes for sample catalysts with different elemental compositions provided in the embodiments of this specification are shown.

[0049] Figure 13 A schematic diagram illustrating the performance of target networks O1 and O2 during the training process according to embodiments of this specification is shown.

[0050] Figure 14 A schematic diagram illustrating the performance of target networks O3 and O4 during the training process according to embodiments of this specification is shown.

[0051] Figure 15 Experimental and predicted values ​​of the initial overpotential of sample catalysts with different elemental compositions provided in the embodiments of this specification are shown.

[0052] Figure 16 A schematic diagram illustrating the relationship between the prediction performance of the target network T on the Tafel slope and the number of layers provided in the embodiments of this specification is shown.

[0053] Figure 17 A schematic diagram illustrating the relationship between the predictive performance of the target network O for the initial overpotential and the number of layers provided in the embodiments of this specification is shown.

[0054] Figure 18 A schematic diagram comparing the prediction performance of different networks for Tafel slope and onset overpotential according to embodiments of this specification is shown; and

[0055] Figure 19 A flowchart is shown of a method for predicting catalyst performance according to embodiments of this specification. Detailed Implementation

[0056] The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to make and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0057] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0058] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0059] The flowcharts used in this specification illustrate operations implemented according to some embodiments of this specification. It should be clearly understood that the operations in the flowcharts may not be implemented in a sequential order. Instead, the operations may be implemented in reverse order or simultaneously. Furthermore, one or more additional operations may be added to the flowcharts. One or more operations may be removed from the flowcharts.

[0060] With advancements in computing power, artificial intelligence (AI) has been applied to various scientific fields. For example, in the study of multi-element catalysts and even high-entropy alloy (HEAs) catalysts, multiple catalytic active sites are used as features, and the binding energies of various reaction intermediates at these sites are predicted using DFT combined with machine learning (ML). As another example, researchers have identified catalytically active HEAs by combining ML with first-principles calculations. These studies demonstrate the powerful role of AI in advancing catalyst research.

[0061] Therefore, we can also use AI neural networks to predict the catalytic performance of catalysts to meet the urgent needs of catalyst research. However, the performance of AI neural networks is highly dependent on the quantity and quality of training samples. For sample catalysts, the more types of descriptors describing the sample catalyst, the better the target network will perform. For ordinary neural networks, the more types of descriptors there are, the more sample catalysts are needed. However, obtaining sample catalysts is difficult and requires a lot of time. Therefore, how to overcome the difficulty of training the target network to achieve the expected results due to insufficient samples but a large number of catalyst descriptors, while ensuring the accuracy of the prediction results, is a difficult problem in the catalyst research process.

[0062] A descriptor is information that describes a catalyst or its catalytic performance, such as information about catalyst components that affect catalytic performance, or information about the properties of the catalyst itself.

[0063] The following is combined Figure 1 This manual introduces the application scenarios for predicting catalyst performance.

[0064] Figure 1 A schematic diagram illustrating an application scenario for predicting catalyst performance according to embodiments of this specification is shown. For example... Figure 1 As shown, scenario 001 may include a training system 100 and a prediction system 200.

[0065] Scenario 001 can be divided into two stages: the training stage and the prediction stage.

[0066] Training phase

[0067] The training system 100 can train a target network based on multiple sample catalysts to predict the target performance of the catalyst, such as Tafel slope or onset overpotential.

[0068] After training is completed in training system 100, the target network can be deployed in prediction system 200.

[0069] Prediction phase

[0070] The prediction system 200 can call the target network to predict the performance indicators corresponding to the target performance of the target catalyst. For distinction, the target network for predicting the Tafel slope of the catalyst will be denoted as target network T, and the target network for predicting the initial overpotential of the catalyst will be denoted as target network O.

[0071] In some embodiments, the training system 100 may store data and instructions for implementing a training method for a target network, and may execute or be used to execute the data and instructions. In some embodiments, the training system 100 may include a hardware device with data processing capabilities and the necessary programs required to drive the hardware device to operate.

[0072] In some embodiments, the prediction system 200 may store data and instructions for predicting catalyst performance, and may execute or be used to execute said data and instructions. In some embodiments, the prediction system 200 may include hardware devices with data processing capabilities and the necessary programs required to drive said hardware devices.

[0073] It is understood that the training system 100 and the prediction system 200 may correspond to the same system or different systems, and this specification does not impose any restrictions on this.

[0074] It should be noted that the training system 100 can correspond to a single device or a cluster of devices; this specification does not impose any restrictions on this. When the training system 100 corresponds to a single device, the training method for the target network can be executed entirely on that device. When the training system 100 corresponds to a cluster of devices, the training method for the target network can be executed collaboratively on multiple devices corresponding to the cluster, or the training method for the target network can have other execution methods; this specification does not impose any restrictions on this.

[0075] The prediction system 200 can correspond to a single device or a cluster of devices; this specification does not impose any limitations on this. When the prediction system 200 corresponds to a single device, the catalyst performance prediction method can be executed entirely on that device. When the prediction system 200 corresponds to a cluster of devices, the catalyst performance prediction method can be executed collaboratively on multiple devices corresponding to the cluster, or the catalyst performance prediction method can have other execution methods; this specification does not impose any limitations on this.

[0076] The network structure of the target network (target network T or target network O) is as follows: Figure 2 As shown, there are F levels, representing the network from level 1 to level F. Each level can include one or more subnetworks, and the number of subnetworks decreases from level 1 to level F. For example, the level 1 network includes m1 subnetworks, the level 2 network includes m2 subnetworks, where m2 < m1, and so on. This represents the first subnetwork in the first level. This represents the first subnetwork in the second level, and so on. The F-level network consists of one subnetwork. The outputs of each subnetwork in the first level serve as the inputs of each subnetwork in the second level, and so on. The network structure of each subnetwork can be the same; this specification does not impose any restrictions on this.

[0077] This specification does not specify the method for decreasing the number of subnetworks at each level. For example, the number of subnetworks at each level may decrease in a regular manner, such as decreasing by a fixed number sequentially, meaning the difference between the number of subnetworks in level 1 and level 2 is the same as the difference between the number of subnetworks in level 2 and level 3. Alternatively, the number of subnetworks at each level may decrease irregularly, such as decreasing by a non-fixed number.

[0078] It should be noted that the number of network layers in target network T and target network O can be different, and the number of sub-networks included in each layer can also be different. This specification does not impose any restrictions on this.

[0079] For example, the sub-network is an Artificial Neural Network (ANN). The network structure of an ANN is as follows: Figure 3 As shown, it includes an input layer, hidden layers, and an output layout. Each neuron in an ANN layer is connected to the neuron in the previous layer, for example... Figure 3 shown Until The input to each neuron in the process includes Until Therefore, ANN can also be called a fully connected neural network. Here, j, j1, j2, and j3 are all integers greater than 0.

[0080] Figure 4 A hardware structure diagram of a system 400 provided according to an embodiment of this specification is shown.

[0081] like Figure 4 As shown, the system 400 can be Figure 1 The training system 100 or the prediction system 200 are used.

[0082] The system 400 includes at least one storage medium 430 and at least one processor 420. In some embodiments, the system 400 may further include a communication port 450 and an internal communication bus 410. Furthermore, the system 400 may also include I / O components 460.

[0083] The internal communication bus 410 can connect to different system components. For example, the internal communication bus 410 can connect to storage medium 430, processor 420, communication port 450, and I / O component 460.

[0084] I / O component 460 supports input / output between the system 400 and other components.

[0085] Communication port 450 is used for data communication between the system 400 and the outside world. For example, communication port 450 can be used for data communication between the system 400 and a network. Communication port 450 can be a wired communication port or a wireless communication port.

[0086] In some embodiments, the network can be any type of wired or wireless network, or a combination thereof. For example, the network may include a cable network, a wired network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a public switched telephone network (PSTN), a Bluetooth network™, a ZigBee™ short-range wireless network, a near field communication (NFC) network, or a similar network.

[0087] In some embodiments, the network may include one or more network access points. For example, the network may include wired or wireless network access points, such as base stations or internet switching points. Through these access points, one or more components of various devices corresponding to the system 400 can connect to the network to exchange data or information.

[0088] Storage medium 430 may include a data storage device. The data storage device may be a non-transitory storage medium or a temporary storage medium. For example, the data storage device may include one or more of a disk 432, a read-only storage medium (ROM) 434, or a random access storage medium (RAM) 436. Storage medium 430 also includes at least one instruction set stored in the data storage device. The instruction set may include computer program code, which may include programs, routines, objects, components, data structures, procedures, modules, etc., that execute the catalyst performance prediction method or the target network training method provided in this specification.

[0089] Processor 420 can be communicatively connected to storage medium 430. Processor 420 is used to execute at least one of the above-described instruction sets. When the system 400 is running, processor 420 reads the at least one instruction set and, according to the instructions of the at least one instruction set, executes the catalyst performance prediction method or the target network training method provided in this specification.

[0090] Processor 420 may be in the form of one or more processors. In some embodiments, processor 420 may include one or more hardware processors, such as microcontrollers, microprocessors, reduced instruction set computers (RISC), application-specific integrated circuits (ASICs), application-specific instruction set processors (ASIPs), central processing units (CPUs), graphics processing units (GPUs), physical processing units (PPUs), microcontroller units, digital signal processors (DSPs), field-programmable gate arrays (FPGAs), advanced RISC machines (ARMs), programmable logic devices (PLDs), any circuit or processor capable of performing one or more functions, or any combination thereof.

[0091] For illustrative purposes only, only one processor 420 is shown in the accompanying drawings of system 400. However, it should be noted that system 400 described herein may also include multiple processors. Therefore, the operational and / or method steps disclosed herein may be executed by a single processor as described herein, or they may be executed jointly by multiple processors. For example, if processor 420 of system 400 described herein executes steps A and B, it should be understood that steps A and B may also be executed jointly or separately by two different processors 420 (e.g., a first processor executes step A, a second processor executes step B, or the first and second processors jointly execute steps A and B).

[0092] Figure 5 A flowchart of a target network training method P500 according to an embodiment of this specification is shown. The training system 100 can execute the target network training method P500, which is used to predict the target performance of a catalyst. Figure 5 As shown, the training method P500 for the target network includes the following steps.

[0093] S510: Obtain experimental values ​​of the target performance for each of the Y sample catalysts, each sample catalyst including at least one element.

[0094] Where Y is an integer greater than 1. The target performance can be understood as a parameter characterizing catalytic properties, such as the Tafel slope or the initial overpotential.

[0095] Among the Y sample catalysts, some sample catalysts contain the same elements, but the proportion of each element in the sample catalyst differs. For example, Co... 80 Ce 20 and Co 60 Ce 40 Some sample catalysts contain different elements, and the proportions of these elements in the sample catalysts also differ. For example, Co... 20 Ni 80 and Fe 37.5 Co 62.5 The number of elements included in some of the catalyst samples also varied; for example, Ce... 100 Co 80 Ni 20 and Fe 37.5 Co 37.5 Ni 25 .

[0096] For example, compared to other materials, the performance indicators of the target properties of amorphous materials are easier to measure experimentally. Therefore, this specification uses amorphous material catalysts to obtain data for training the target network, that is, the experimental values ​​of the target properties of each of the Y sample catalysts.

[0097] For example, this specification uses high-throughput aerogel synthesis technology and electrochemical characterization to obtain experimental values ​​of the Tafel slope and onset overpotential for each of the Y sample catalysts. Specifically, a multi-channel feed system equipped with an online mixer allows each channel to independently deliver different elemental components. By accurately controlling the proportion of each element, a metal precursor solution is prepared. After sol-gel transformation and supercritical drying of the metal precursor solutions prepared with different elements and different proportions, each sample catalyst is obtained. The experimental values ​​of the Tafel slope and onset overpotential for each sample catalyst are measured using linear sweep voltammetry (LSV). In wet chemical synthesis, synthesis parameters (including temperature, pressure, etc.) have a significant impact on the phase, morphology, and structure of catalysts. Therefore, high-throughput synthesis technology ensures that the synthesis and measurement conditions of all sample catalysts are the same. For example, amorphous aerogels have similar morphologies and mass-to-surface-area ratios, thus ensuring that the experimental values ​​obtained are mainly determined by the elements in the sample catalyst and the corresponding proportions of those elements. This provides reliable sample data for the subsequent training of the target network and eliminates the uncertainty caused by differences in catalyst synthesis and measurement conditions.

[0098] The experimental values ​​of the Tafel slope and initial overpotential for each of the Y sample catalysts can be found in Tables 1, 2 and 3.

[0099] Table 1

[0100]

[0101]

[0102] Table 2

[0103]

[0104]

[0105] Table 3

[0106] Sample catalyst Tafel slope (mv / dec) Initial overpotential (mV) <![CDATA[La1Co1Al1]]> 44 310 <![CDATA[La1Co1Cu1]]> 43 251 <![CDATA[La1Co1Ni1]]> 49 301 <![CDATA[La1Ni1Fe1]]> 58 250 <![CDATA[La1Co1Fe1]]> 56 248 <![CDATA[La1Co1]]> 55 300 <![CDATA[Ce1Co1Fe1]]> 67 242 <![CDATA[Al1Co1]]> 64 490 <![CDATA[Ni1Co1Fe1]]> 75 220 <![CDATA[Al1Co1Ni1]]> 81 360 <![CDATA[La1Ni1Mn1]]> 90 294 <![CDATA[Ce1Co1Ni1]]> 94 250 <![CDATA[La1Ni1]]> 100 286 <![CDATA[La1Ni1Li1]]> 103 260 <![CDATA[La1Ni1K1]]> 99 300

[0107] Table 1 shows the experimental values ​​of Tafel slope and onset overpotential for two types of multi-element catalysts. The first type of multi-element catalyst consists of Fe, Co, and Ce, with varying proportions of different elements; some elements may have a proportion of 0. The second type of multi-element catalyst consists of Fe, Co, and Ni, with varying proportions of different elements; some elements may also have a proportion of 0. Table 2 shows the experimental values ​​of Tafel slope and onset overpotential for a third type of multi-element catalyst (constituting Fe, Co, and La), with varying proportions of different elements; some elements may have a proportion of 0. Table 3 also shows the experimental values ​​of Tafel slope and onset overpotential for multiple multi-element catalysts. The difference is that some of the multi-element catalysts shown in Table 3, in addition to containing at least one of Fe, Co, Ce, Ni, and La, also contain non-transition metal elements or rare earth elements (such as Al, Li, and K). It is evident that the sample catalysts in this specification cover a wide range of elements, which is beneficial for improving the prediction accuracy of the target network and expanding the prediction applicability.

[0108] S520: For each sample catalyst, based on the proportion information of at least one element in the sample catalyst and the attribute values ​​of X attributes corresponding to each of the at least one element, generate K descriptors to describe the sample catalyst, and determine N1 core descriptors and K1 non-core descriptors from the K descriptors. The core descriptors have a greater influence on the target performance than the non-core descriptors, and K1 = K - N1.

[0109] In some embodiments, the X attributes include at least two of the following dimensions: periodic table dimension; thermodynamic dimension; physical dimension; and crystallographic dimension. In this embodiment, multiple dimensions of elemental attributes are considered, resulting in K descriptors that can more comprehensively characterize the sample catalyst, avoiding the omission of important descriptors and providing a good foundation for improving the prediction accuracy and generalization ability of the target network.

[0110] For example, see at least some of the X attributes in Table 4.

[0111] Table 4

[0112]

[0113]

[0114] In Table 4, BCC refers to Body Centered Cubic, FCC refers to Face Centered Cubic, ICSD refers to Inorganic Crystal Structure Database, and GS refers to the ground state of an object.

[0115] For example, the sample catalyst is Fe 40 Co 30 La 30 Describing Fe 40 Co 30 La 30 The K descriptors can be generated based on the following: the proportion of Fe element. The proportion of Co element The proportion of La element And the attribute values ​​of at least some of the attributes listed in Table 4 for the Fe element, the attribute values ​​of at least some of the attributes for the Co element, and the attribute values ​​of at least some of the attributes for the La element.

[0116] The following section uses sample catalyst #1 as an example to introduce the generation process of the K descriptors describing sample catalyst #1. The K descriptors describing other sample catalysts can be derived in the same way.

[0117] In some embodiments, the K descriptors describing sample catalyst #1 include at least one of the following:

[0118] The first set of descriptors is used to describe the statistical characteristics of sample catalyst #1 on the X properties;

[0119] The second set of descriptors is used to describe the valence electron occupation state of sample catalyst #1.

[0120] The third set of descriptors is used to describe the ionicity of sample catalyst #1; and

[0121] The fourth set of descriptors is used to describe the configurational entropy (CE) of sample catalyst #1.

[0122] In this embodiment, K descriptors can describe different aspects of the sample catalyst, providing a comprehensive description of the sample catalyst. This supports the prediction of complex multi-element catalysts, improves the prediction accuracy of the target network, and enhances the generalization ability of the target network.

[0123] For example, the first set of descriptors is generated as follows: For each of the X attributes, the training system 100, based on the proportion of each element in sample catalyst #1 and the attribute value of each element on the attribute, performs statistical analysis on the characteristics of sample catalyst #1 on the attribute from multiple statistical dimensions to obtain P statistical values. Then, based on the P statistical values ​​corresponding to sample catalyst #1 on each of the X attributes, the first set of descriptors is generated. The number of descriptors in the first set is equal to X*P. In this example, the first set of descriptors involves multiple statistical dimensions, and both the number of descriptors and the statistical dimensions involved are relatively large. This approach uncovers the basic statistical characteristics of the sample catalyst, enabling accurate and comprehensive characterization of the sample catalyst, which is beneficial for training the target network.

[0124] Taking attribute #1 as an example, the training system 100 can statistically analyze the characteristics of sample catalyst #1 in attribute #1 from multiple statistical dimensions to obtain P statistical values. These P statistical values ​​are all the descriptors corresponding to attribute #1 in the first set of descriptors. The following mainly introduces the generation process of all descriptors corresponding to attribute #1 in the first set of descriptors. The descriptors corresponding to other attributes in the first set of descriptors are generated in the same way and will not be described in detail.

[0125] For example, the multiple statistical dimensions include statistical dimension #1, statistical dimension #2, statistical dimension #3, statistical dimension #4, and statistical dimension #5. Statistical dimension #1 is used to calculate the weighted average of the attribute values ​​of each element included in the sample catalyst #1 on attribute #1. The formula corresponding to statistical dimension #1 is as follows:

[0126]

[0127] Where, x i f represents the percentage of the i-th element in the chemical formula of sample catalyst #1. i This represents the attribute value of the i-th element on attribute #1. This represents the descriptor corresponding to attribute #1 obtained from statistical dimension #1. Assume the sample catalyst #1 is Fe. 40 Co 30 La 30The training system 100 can be determined based on formula (1).

[0128] Statistical dimension #2 is used to calculate the weighted deviation of the attribute values ​​of each element included in sample catalyst #1 on attribute #1. The formula for statistical dimension #2 is as follows:

[0129]

[0130] in, This represents the descriptor corresponding to attribute #1 obtained from statistical dimension #2.

[0131] Statistical dimension #3 is used to calculate the standard deviation of the attribute values ​​of each element included in sample catalyst #1 on attribute #1. The formula for statistical dimension #3 is as follows:

[0132]

[0133] Where f represents the descriptor corresponding to attribute #1 obtained from statistical dimension #3.

[0134] Statistical dimension #4 is used to determine the maximum, minimum, and range of attribute values ​​or weighted attribute values ​​for each element included in sample catalyst #1. The formula for statistical dimension #4 is as follows:

[0135] (f i )_min,max,range (4)

[0136] (x i *f i )_min,max,range (5)

[0137] Training system 100 can obtain 6 descriptors corresponding to attribute #1 from statistical dimension #4, as follows:

[0138] (f i )_min represents the minimum value of the attribute values ​​of all elements included in sample catalyst #1 in attribute #1; (f i )_max represents the maximum value of the attribute values ​​of all elements included in sample catalyst #1 in attribute #1; (f i )_range represents the range of values ​​for each element in sample catalyst #1 within the attribute values ​​of attribute #1, which can be (f i )_max and (f i )_min difference. (x i *f i)_min represents the minimum value of the weighted attribute value of all elements included in sample catalyst #1 on attribute #1; (x i *f i )_max represents the maximum value among the weighted attribute values ​​of all elements included in sample catalyst #1 on attribute #1; (x i *f i )_range represents the range of values ​​for each element included in sample catalyst #1 within the weighted attribute values ​​of attribute #1, which can be (x i *f i )_max and (x i *f i The difference between )_min.

[0139] Statistical dimension #5 is used to statistically analyze the degree of deviation of different elements in all elements included in sample catalyst #1 regarding attribute #1. The formula corresponding to statistical dimension #5 is as follows:

[0140]

[0141] Where, x i1 This represents the proportion of the i1th element in the chemical formula of sample catalyst #1, x. i2 f represents the percentage of the i2th element in the chemical formula of sample catalyst #1. i1 f represents the attribute value of the i1th element on attribute #1. i2 AP represents the attribute value of the i2th element on attribute #1, and AP represents the degree of deviation between different elements on attribute #1. 加权 This indicates the degree of deviation of the weighted attribute value of different elements on attribute #1.

[0142] Training system 100 can obtain 8 descriptors corresponding to attribute #1 from statistical dimension #5, as follows:

[0143] AP_ave represents the average deviation of different elements in attribute #1 among all elements included in sample catalyst #1; AP_min represents the minimum deviation of different elements in attribute #1 among all elements included in sample catalyst #1; AP_max represents the maximum deviation of different elements in attribute #1 among all elements included in sample catalyst #1; AP_range represents the range of values ​​for the deviation of different elements in attribute #1 among all elements included in sample catalyst #1, which can be the difference between AP_max and AP_min; AP 加权 _ave represents the average weighted attribute value deviation of different elements in attribute #1 among all elements included in sample catalyst #1; AP 加权_min represents the minimum weighted attribute value deviation of different elements in attribute #1 among all elements included in sample catalyst #1; AP 加权 _max represents the maximum weighted attribute value deviation of different elements in attribute #1 among all elements included in sample catalyst #1; AP 加权 _range represents the range of weighted attribute value deviations for different elements in attribute #1 among all elements included in sample catalyst #1, and can be AP. 加权 _max and AP 加权 The difference between _min and AP. 加权 The variations in atomic size and electronegativity among different elements have a significant impact on the structure and performance of catalyst materials. Training system 100 considered AP and AP when generating the first set of descriptors. 加权 Comprehensive descriptions of catalysts are crucial. Additionally, if sample catalyst #1 consists of a single element, all eight descriptors mentioned above will be zero.

[0144] Therefore, the training system 100 can obtain 17 descriptors corresponding to attribute #1 through the above-mentioned multiple statistical dimensions. If the X attributes are the 53 types exemplified in Table 4, for sample catalyst #1, the first set of descriptors obtained by the training system 100 can include 17 * 53 = 901 descriptors, which fully covers the basic statistical attributes of sample catalyst #1.

[0145] In some embodiments, the second set of descriptors includes: occupancy information of Q valence electrons in sample catalyst #1, wherein the occupancy information of the z-th valence electron is obtained in the following manner:

[0146] For each element in sample catalyst #1, the training system 100 determines the number of each valence electron and the total number of valence electrons in the element based on the attribute values ​​of at least some of the X attributes corresponding to the element; for the z-th valence electron, the system determines the occupation information of the z-th valence electron based on the proportion information of each element in sample catalyst #1, the number of the z-th valence electron, and the total number of valence electrons corresponding to each element.

[0147] In other words, each descriptor in the second set of descriptors describes the valence electron occupation state. The second set of descriptors obtained by training system 100 includes Q descriptors, with one descriptor corresponding to each type of valence electron. Valence electrons have a significant impact on the catalytic reaction process; therefore, considering the occupation states of multiple valence electrons in the descriptors can characterize the key properties of the sample catalyst and improve the accuracy of the target network's prediction of catalytic performance.

[0148] For example, the training system 100 obtains the descriptor corresponding to the z-th valence electron based on the following formula:

[0149]

[0150] Among them, F z E represents the descriptor corresponding to the z-th valence electron. z E represents the number of z-th valence electrons in the i-th element of the chemical formula of sample catalyst #1. n This represents the total number of Q types of valence electrons in the i-th element, where the Q types of valence electrons include p, s, d, and f. Specifically, the descriptor corresponding to valence electron p is... The descriptors for other valence electrons follow the same pattern, so training system 100 can obtain 4 descriptors.

[0151] In some embodiments, the third set of descriptors is obtained based on the proportion information of each element in sample catalyst #1, the electronegativity of each element, and any of the following: the maximum value of the electronegativity of each element included in sample catalyst #1; the average value of the electronegativity of each element included in sample catalyst #1. Electronegativity is a measure of the ability of an element's atoms to attract electrons in a compound. Therefore, considering electronegativity in the descriptors can characterize the key properties of the sample catalyst and improve the accuracy of the target network's prediction of catalytic performance.

[0152] For example, training system 100 obtains the third set of descriptors based on the following formula:

[0153]

[0154] Where I represents the third set of descriptors, This represents the maximum electronegativity among the elements included in sample catalyst #1, or the average electronegativity among the elements included in sample catalyst #1; when When representing the maximum value of electronegativity, training system 100 can obtain the first descriptor in the third set of descriptors based on formula (9); when When representing the average value of electronegativity, training system 100 can obtain the second descriptor in the third set of descriptors based on formula (9); when When representing the average value of electronegativity, the training system 100 can also obtain the third descriptor in the third set of descriptors. Specifically, when the I calculated by the training system 100 based on formula (9) exceeds 1.7, the third descriptor is 1, and when the calculated I does not exceed 1.7, the third descriptor is 0 (this is also the standard for determining the formation of ionic bonds).

[0155] In some embodiments, the fourth set of descriptors is obtained based on the proportion of each element in sample catalyst #1, the Boltzmann constant, and room temperature. Configuration entropy can indicate the degree of disorder in atomic distribution. Considering configuration entropy in the descriptors can characterize the key properties of the sample catalyst and improve the accuracy of the target network's prediction of catalytic performance.

[0156] For example, the training system 100 obtains the fourth set of descriptors based on the following formula:

[0157] ΔS con =-k B T∑x i *lnx i (10)

[0158] Where, ΔS con k represents configuration entropy. B represents the Boltzmann constant; T represents room temperature, such as 298K. Training system 100 can obtain a descriptor based on formula (10).

[0159] In summary, with 53 attributes, for each sample catalyst, the training system 100 can generate 901 descriptors in the first group, 4 in the second group, 3 in the third group, and 1 in the fourth group based on formulas (1) to (10), that is, a total of 909 descriptors in the K groups.

[0160] In the field of materials science, obtaining experimental data on sample catalysts is challenging and time-consuming. Therefore, when the number of sample catalysts is limited, training a neural network using only 909 descriptors as features of a single sample catalyst is prone to overfitting, especially when the number of descriptors far exceeds the number of sample catalysts. For instance, in an ANN with two hidden layers, inputting 909 descriptors would require training to determine over a billion hyperparameters, making it difficult to achieve the desired results with a limited number of sample catalysts. Therefore, this specification employs a Genetic Algorithm (GA) to reduce the number of input descriptors in individual subnetworks of the target network.

[0161] For example, the training system 100 groups the K descriptors based on the competitive swarm optimizer (CSO) algorithm, with descriptors within each group competing in pairs. After each competition, the winning descriptor proceeds to the next iteration, while the losing descriptors can learn from the winning descriptor to update their position and velocity. The basic principle of the CSO algorithm can be seen in the following formula:

[0162]

[0163] Where t represents the number of iterations; and It is in [0,1] n Three vectors randomly generated within the range; for each descriptor, '1' indicates that the descriptor was selected after the competition, and '0' indicates that the descriptor was not selected after the competition; 'n' represents the number of descriptors, for example, 'n' is 909; variables and These represent the winning and losing descriptors, respectively. φ represents the average position of the descriptor population in the t-th iteration, and φ determines the position of the population. The influence This represents the velocity of the i-th descriptor in the t-th iteration.

[0164] In this example, training system 100 uses 1-R from the test set (at least a portion of the catalysts from Y sample catalysts). 2 As the fitness function, the fitness function is minimized through strategic selection of descriptors. Specifically, in the early stage of iteration, the training system 100 randomly selects multiple combinations of descriptors and their speeds, and then randomly selects two combinations of descriptors to evaluate the fitness function. The descriptor combination with the lower fitness function value is the winner, and the descriptor combination with the higher fitness function value is the loser. During the iteration process, updates and iterations are performed according to formulas (11) and (12). The iteration process terminates when the fitness function value no longer decreases.

[0165] Figure 6 The GA iterative process is illustrated. a) An example of the GA iterative process with the Tafel slope as the prediction target is illustrated. Figure 6 b) illustrates the GA iterative process with the learning of the initial overpotential as the prediction objective.

[0166] Figure 6 In the diagram, the x-axis represents the number of GA iterations, the left y-axis represents the number of descriptors (d), and the right y-axis represents the test R. 2 Regarding a), during the GA iteration process, multiple preferred descriptor combinations #1 will appear. The difference between the predicted value and the experimental value of the Tafel slope of the sample catalyst obtained by the preferred descriptor combination #1 is less than or equal to a preset threshold 1. Regarding b), during the GA iteration process, multiple preferred descriptor combinations #2 will appear. The difference between the predicted value and the experimental value of the onset overpotential of the sample catalyst obtained by the preferred descriptor combination #2 is less than or equal to a preset threshold 2. The preset threshold 1 and preset threshold 2 can be the same or different; this specification does not impose any restrictions on this.

[0167] See also Figure 6 After iterative selection, the training system 100 determines that the number of descriptors corresponding to the Tafel slope and the initial overpotential is N (e.g., 15). The test R obtained in the test set is determined by a) 2 The value is 0.682, determined by b) in the test set. 2 The value is 0.651, which basically meets the test R... 2 The upper limit. That is, when the number of descriptors for each subnetwork input is N, it is possible to predict the Tafel slope and the initial overpotential with relatively high accuracy.

[0168] It should be noted that the determination of N may be related to the amount of sample catalyst, therefore this specification does not limit the value of N.

[0169] Furthermore, during the iteration process, descriptors from multiple preferred descriptor combinations #1 whose frequency of occurrence is greater than or equal to a preset threshold 3 are used as core descriptors required for predicting the Tafel slope, while descriptors whose frequency of occurrence is less than the preset threshold 3 are used as non-core descriptors required for predicting the Tafel slope. Similarly, descriptors from multiple preferred descriptor combinations #2 whose frequency of occurrence is greater than or equal to a preset threshold 4 are used as core descriptors required for predicting the initial overpotential, while descriptors whose frequency of occurrence is less than the preset threshold 4 are used as non-core descriptors required for predicting the initial overpotential.

[0170] Table 5 provides the core descriptors required to predict the Tafel slope and the core descriptors required to predict the initial overpotential.

[0171] Table 5

[0172]

[0173] For example, AP-ave-Electronegativity, representing the average deviation of different elements in electronegativity in the sample catalyst, can be obtained based on formula (6). max-weight-BCCenergy-diff, representing the maximum value of the weighted attribute values ​​of each element in the sample catalyst along the BCC atomic energy difference, can be obtained based on formula (5). min-weight-Row, representing the minimum value of the weighted attribute values ​​of each element in the sample catalyst along the row in the periodic table, can be obtained based on formula (5). AP-max-BCCenergy-diff, representing the maximum value of the deviation of different elements in the sample catalyst along the BCC atomic energy difference, can be obtained based on formula (6). AP-ave-phi, representing the average deviation of different elements in the sample catalyst along the work function, can be obtained based on formula (6).

[0174] This example provides several descriptors that are highly correlated with the Tafel slope, which are used as core descriptors in the training process of the target network T. During the training process, the target network T will assign appropriate weights to them to reflect their importance in predicting the Tafel slope.

[0175] Figure 7 Examples illustrate the relationship between the core descriptor corresponding to the Tafel slope and the Tafel slope. Specifically, a) shows the relationship between MDT1 and the Tafel slope; b) shows the relationship between MDT2 and the Tafel slope; c) shows the relationship between MDT3 and the Tafel slope; and d) shows the relationship between MDT4 and the Tafel slope.

[0176] a) shows two distinct regions. The left region exhibits smaller electronegativity differences and a weaker correlation between MDT1 and the Tafel slope. Conversely, the right region shows larger electronegativity differences and a stronger correlation. Specifically, the Tafel slope is higher when MDT1 values ​​are between 0.4 and 0.5, and decreases when MDT1 values ​​are at the extremes of this range.

[0177] See b) and c). It can be seen that the two descriptors MDT2 and MDT3 are positively correlated with the Tafel slope. These findings are consistent with the results of GA iteration and provide a theoretical basis for training the target network corresponding to the Tafel slope.

[0178] For example, the maximum value of the weighted attribute value of each element in the sample catalyst in terms of the number of unfilled valence electrons (d) can be obtained based on formula (5). The average value of the deviation of the weighted attribute value of different elements in the sample catalyst in terms of mass density (AP-ave-weighted-Desity) can be obtained based on formula (7). The minimum value of the deviation of the weighted attribute value of different elements in the sample catalyst in terms of work function (AP-min-weighted-phi) can be obtained based on formula (7). The average value of the deviation of different elements in the sample catalyst in terms of the number of filled valence electrons (d) can be obtained based on formula (6). The maximum value of the deviation of the weighted attribute value of different elements in the sample catalyst in terms of covalent radius (AP-maxweight-CovalenceRadius) can be obtained based on formula (7).

[0179] This example provides several descriptors that are highly correlated with the initial overpotential. These descriptors are used as core descriptors in the training process of the target network O. During the training process, the target network O will assign appropriate weights to them to reflect their importance in predicting the initial overpotential.

[0180] Figure 8 Examples illustrate the relationship between the kernel descriptor corresponding to the initial overpotential and the initial overpotential. Specifically, a) shows the relationship between MDO1 and the initial overpotential; b) shows the relationship between MDO5 and the initial overpotential; c) shows the relationship between MDO2 and the initial overpotential; and d) shows the relationship between MDO3 and the initial overpotential.

[0181] See a), MDO1 is positively correlated with the initial overpotential; see b), when the value of MDO5 is between 1 and 4, the initial overpotential value is relatively small.

[0182] See Figure 7 and Figure 8 We can observe that the core descriptors obtained through the GA algorithm in this specification conform to the principles of electrocatalyst design. For example, the degree of deviation of different elements in the sample catalyst on MDT1 affects the charge distribution and electron cloud around the catalytic active sites, thereby affecting the adsorption and desorption kinetics of reactants in the catalytic reaction. MDT5 and MDO3 are related to interfacial charge transfer and activity in the electrocatalyst. MDO1 and MDO4 are related to the characteristics of d-electron orbitals; regulating d-orbital electrons is effective in controlling reactant adsorption and the formation of intermediate active species. In other words, the core descriptors obtained in this specification can describe the key characteristics of the catalyst, which is beneficial for training a target network with high prediction accuracy.

[0183] S530: Input a portion of the K descriptors corresponding to each sample catalyst into the target network for prediction to obtain the predicted value corresponding to the target performance of the sample catalyst. Wherein: the target network includes multiple sub-networks with the same structure and arranged hierarchically. Each sub-network includes N input nodes and 1 output node. The output node of the sub-network in the i-th level serves as the input node of the sub-network in the (i+1)-th level. Each sub-network in the first level is input with the N1 core descriptors and N2 non-core descriptors. The N2 non-core descriptors are a portion of the K1 non-core descriptors, and the non-core descriptors input to different sub-networks are not completely identical. The ratio of N to Y is less than 1, the ratio of K to Y is greater than 1, and N2 = N - N1.

[0184] Where X, K, K1, N1, N2, Y, and N are all integers greater than 1, and i is an integer greater than or equal to 1. It should be noted that the ratio of K to Y can also be greater than 2, 3, 4, 5, 6, 7, or even larger integers; these are not listed here. The larger the ratio of K to Y, the more the advantages of the target network in this specification are demonstrated. Using the target network can avoid the overfitting problem during target network training when the number of descriptors far exceeds the number of sample catalysts.

[0185] In some embodiments, see Figure 9 The training system 100 generates m1 sets of descriptors based on the K descriptors, each set including the N1 core descriptors and N-N1 non-core descriptors; the m1 sets of descriptors are input into the m1 subnetworks of the first layer of the target network to obtain m1 intermediate prediction values ​​of the target performance; according to the order of i from 2 to F, the m1 subnetworks of the (i-1)th layer of the target network are... 1-1 The target performance m obtained from the subnetwork i-1 The intermediate predicted values ​​are divided into m i Group input of m of the i-th layer of the target network i A subnetwork is used to obtain the target performance m. i The intermediate predicted values ​​are then used as the final predicted values ​​of the target performance output by the sub-network of layer F. Where m1, F, and m are intermediate predicted values. i-1 All are integers greater than 1, m i It is an integer greater than or equal to 1.

[0186] In this embodiment, the K descriptors are divided into multiple groups and input into multiple sub-networks. This avoids the problem of overfitting caused by a large number of descriptors inputting into a single sub-network, which would result in an excessive number of hyperparameters required for training due to the low number of sample catalysts. Within the hierarchical structure of the target network, descriptors can be distributed across multiple sub-networks, preventing overfitting and allowing the number of descriptors to be expanded to 909, enriching the dimensions for catalyst characterization. For each sub-network, the number of input descriptors is relatively small, thus resolving the contradiction caused by the number of descriptors far exceeding the number of sample catalysts. Furthermore, testing (test numbers are described below) shows that the prediction accuracy of the target network is higher than other neural networks, which is beneficial for catalyst research.

[0187] For example, when K is 909, the training system 100 can obtain the following from 909-5=904 non-core descriptors through random combination. There are N non-core descriptor groups, and each non-core descriptor group plus N1 core descriptors can be obtained. There are 2.5 × 10^10 descriptor groups. That is, the number of subnetworks in the first level can reach 2.5 × 10^10. 23 However, this specification does not limit the number of subnetworks in the first level.

[0188] Tests have shown that, No descriptor set is completely indifferent to the Tafel slope or initial overpotential. Different descriptor sets contribute to the prediction of the Tafel slope or initial overpotential to varying degrees, a degree that can be determined during network training. Compared to traditional methods that train a network using a single descriptor combination, the training method described in this specification considers various descriptor combinations, capturing or exploring the potential influence of different descriptor combinations on the predicted Tafel slope or initial overpotential, thus training a target network with higher prediction accuracy. Traditional algorithms, such as bagging, boosting, and stacking, typically rely on simple geometric mean, failing to consider the differences between descriptor sets and thus unable to effectively capture the varying contributions of different descriptor sets to the predicted Tafel slope or initial overpotential.

[0189] For the first level of the target network, the input of each sub-network is a descriptor set, and each descriptor set includes N (e.g., 15) descriptors; that is, each input node of each sub-network in the first level corresponds to one descriptor. For example, sub-network The input is descriptor group 1, and so on. Each descriptor group includes N1 (e.g., 5) core descriptors and N2 (e.g., 10) non-core descriptors. At least one of the N2 non-core descriptors in different descriptor groups must be different. For the second level, the input to each sub-network is N intermediate predictions from the outputs of the sub-networks in the first level, such as the sub-network... The input is a subnetwork To-child network The N intermediate predicted values ​​output are respectively to For level 3, the input to each subnetwork is N intermediate predictions from the outputs of the subnetworks in level 2, and so on. For level F, the subnetwork... F The input is the intermediate prediction values ​​of the output of the N sub-networks at the F-1 level.

[0190] During training, the target network can learn iterative knowledge from different sub-networks. This knowledge can include overlapping information as well as unique knowledge from each sub-network, thereby improving prediction accuracy. When the number of sample catalysts is limited, training a neural network with only 909 descriptors avoids an excessive number of hyperparameters, which could negatively impact training accuracy. As the number of layers in the target network increases, the predicted performance of the target network gradually converges.

[0191] It should be noted that, during the training of the target network T, the target performance corresponding to the sample catalyst is the Tafel slope. The core descriptors input to the first-level sub-network are, for example, MDT1 to MDT5, and the outputs of each sub-network after the first level are predicted values ​​of the Tafel slope. During the training of the target network O, the target performance corresponding to the sample catalyst is the initial overpotential. The core descriptors input to the first-level sub-network are, for example, MDO1 to MDO5, and the outputs of each sub-network after the first level are predicted values ​​of the initial overpotential.

[0192] S540: The parameters of the target network are updated with the training objective of minimizing the difference between the predicted and experimental values ​​of the target performance of the sample catalyst.

[0193] For example, this specification uses the coefficient of determination (R²). 2 R² and mean absolute error (MAE) are used as evaluation metrics for the target network. 2 The calculation formula is as follows:

[0194]

[0195] Where n represents the number of sample catalysts in the current training round, y i This represents the experimental value of the i-th sample catalyst with respect to the target performance. F(x) represents the average experimental value of the sample catalyst with respect to the target performance. i R represents the predicted value of the i-th sample catalyst with respect to the target performance, as predicted by the target network. 2 A higher MAE value indicates better performance of the target network, meaning more accurate prediction results.

[0196] Target network T

[0197] For example, training system 100 generates training set T1 based on all sample catalysts and their corresponding Tafel slopes in Table 1, and trains target network T1 based on training set T1. Based on training set T1, training system 100 adds 30% of the sample catalysts and their corresponding Tafel slopes from Table 2, forming training set T2, and updates the parameters in target network T1 based on training set T2 to obtain target network T2. Based on training set T1, training system 100 adds all sample catalysts and their corresponding Tafel slopes from Table 2, forming training set T3, and updates the parameters in target network T2 based on training set T3 to obtain target network T3. Based on training set T1, training system 100 adds all sample catalysts and their corresponding Tafel slopes from Table 3, forming training set T4, and updates the parameters in target network T3 based on training set T4 to obtain target network T4.

[0198] Figure 10 and Figure 11 The performance of the target network T during the training process described above is illustrated.

[0199] Figure 10 (a) presents the predicted values ​​of the catalyst samples in the training set T1 with respect to the Tafel slope using the target network T1, along with a comparison with experimental values. During the training process using the training set T1, the training system 100 uses 25% of the data from the training set T1 as test data, thereby determining the test R of the target network T1. 2 The value is 0.756, and the MAE is 3.17.

[0200] Figure 10 b) presents the predicted Tafel slope of the sample catalyst composed of Fe, Co, and La using the target network T1, and it can be found that... Figure 10 b) There are two regions with relatively small Tafel slopes. Twenty-one experiments were conducted in these two regions, and the results are as follows: Figure 10 As shown in d). The sample catalysts corresponding to these 21 experiments, along with all the sample catalysts in Table 1, form the training set T2.

[0201] Figure 10 c) presents a comparison between the predicted and experimental values ​​of the Tafel slope for the sample catalyst composed of Fe, Co, and La using the target network T1, from which R can be obtained. 2 It is 0.314.

[0202] Figure 10 (e) presents the predicted values ​​and experimental values ​​of the catalyst's Tafel slope for each sample in training set T2 using the target network T2. During the optimization of the target network T1 using training set T2, the training system 100 uses 35% of the data from training set T2 as test data, thus determining the test R-value of the target network T2. 2 The value increased from 0.756 to 0.949, while the MAE decreased from 3.17 to 2.19.

[0203] Figure 10 f) presents a comparison between the predicted and experimental values ​​of the Tafel slope for the sample catalyst composed of Fe, Co, and La by the target network T2, R. 2 It increased from 0.314 to 0.952.

[0204] Figure 11 (a) presents the predicted values ​​of the catalyst samples in training set T3 with respect to the Tafel slope using the target network T3, along with a comparison with experimental values. During the training process using training set T3, the training system 100 uses 35% of the data from training set T3 as test data, which determines the test R of the target network T3. 2The value is 0.952, which is higher than the test R of the target network T2. 2 value.

[0205] Figure 11 b) presents the predicted values ​​of the Tafel slope for each sample catalyst in Table 3 using the target network T3, and compares them with the experimental values. Figure 11 c) presents the predicted values ​​of the catalyst samples in the training set T4 with respect to the Tafel slope through the target network T4, and compares them with the experimental values. Figure 11 d) presents the predicted values ​​of the Tafel slope for each sample catalyst in Table 3 using the target network T4, and compares them with the experimental values.

[0206] The experimental and predicted values ​​of the Tafel slope for each sample catalyst in Table 3, obtained by the training system 100 through the target network T3, are shown in Table 6. The sample catalysts in Table 3 contain three types of elements: transition metals, rare earth metals, and alkali metals. Transition metals (training and prediction: Fe, Co, Ni; prediction only: Cu, Mn) are ideal non-noble metal OER catalysts because their partially filled d-orbital electrons can effectively participate in multi-electron transfer processes. Rare earth metals (training and prediction: La, Ce) can modulate the electronic state of transition metals through their unique 4f electronic structure, thereby enhancing OER catalytic activity. The introduction of alkali metals (prediction only: Al, Li, K) can further modulate the electronic structure of transition metals, enhance the O 2p band, and stimulate the release of lattice oxygen, thus enhancing OER activity. Halogens are excluded from consideration as sample catalysts due to their high electronegativity and strong ionic bonding with metals, as they have no catalytic effect on OER.

[0207] Table 6

[0208]

[0209] The target network T3 was able to predict the Tafel slopes of the 15 electrocatalysts listed in Table 3 relatively well. Some of these electrocatalysts include transition metals and rare earth elements. Each of these 15 electrocatalysts contains at least one element selected from Fe, Co, Ce, Ni, and La. The target network T3 showed slightly lower prediction accuracy for the Tafel slopes of electrocatalysts containing non-transition metals and rare earth elements (such as Al, Li, and K), with an error range of 2.6% to 10.8%. For example, the prediction error for the Tafel slope of La1Co1Al1 was 10.8%. Conversely, the prediction error for catalysts containing at least one element from Fe, Co, Ce, Ni, and La was lower, ranging from 0.9% to 6%. For example, the prediction error for the Tafel slope of La1Co1Cu1 was 0.9%. This indicates that the target network T3 has effectively captured the interactions between transition metals and rare earth elements in OER electrocatalysis.

[0210] To further improve performance, we added sample catalysts containing Al, Li, and K to training set 3, forming training set 4. The performance of the target network T4 was slightly improved, with R... 2 It is 0.961 ( Figure 11 c). For all 15 catalysts, the difference between experimental and target network T4 predictions is shown in Figure 11 In d, R 2 The value is 0.981. As shown in Table 6, the maximum prediction error of the target network T4 is 5.2%, and the minimum prediction error is 0.01%. These data demonstrate the strong predictive ability of the target network T with a relatively small number of samples for catalyst training.

[0211] Figure 12 The experimental and predicted values ​​of Tafel slopes for sample catalysts with different elemental compositions are illustrated. Specifically: a) shows the experimental Tafel slopes for sample catalysts composed of Fe, Co, and Ni; b) shows the experimental Tafel slopes for sample catalysts composed of Fe, Co, and Ce; c) shows the experimental Tafel slopes for sample catalysts composed of Fe, Co, and La; d) shows the predicted Tafel slopes for sample catalysts composed of Fe, Co, and Ni using the target network T4; e) shows the predicted Tafel slopes for sample catalysts composed of Fe, Co, and Ce using the target network T4; and f) shows the predicted Tafel slopes for sample catalysts composed of Fe, Co, and La using the target network T4. Regions 1 / 2 / 3 / 4 marked in d), e), and f) represent regions with Tafel slopes greater than 74.

[0212] Referring to Table 6, the training system 100 can determine that the prediction error of the target network T4 is 2%. When the prediction error is less than or equal to a preset threshold of 5, the training system 100 can use the target network T4 as the final optimized target network T and deploy it in the prediction system 200. If the prediction error is greater than the preset threshold of 5, the training system 100 can continue training based on other sample catalysts until the prediction error is less than or equal to the preset threshold of 5, at which point the final optimized target network T can be deployed in the prediction system 200.

[0213] Target Network O

[0214] For example, training system 100 generates training set O1 based on all sample catalysts and their corresponding initial overpotentials in Table 1, and trains target network O1 based on training set O1. Based on training set O1, training system 100 adds 30% of the sample catalysts and their corresponding initial overpotentials from Table 2, forming training set O2, and updates the parameters in target network O1 based on training set O2 to obtain target network O2. Based on training set O1, training system 100 adds all sample catalysts and their corresponding initial overpotentials from Table 2, forming training set O3, and updates the parameters in target network O2 based on training set O3 to obtain target network O3. Based on training set O1, training system 100 adds all sample catalysts and their corresponding initial overpotentials from Table 3, forming training set O4, and updates the parameters in target network O3 based on training set O4 to obtain target network O4.

[0215] Figures 13 to 15 The performance of the target network O during the training process described above is illustrated.

[0216] Figure 13 (a) presents the predicted values ​​and experimental values ​​of the onset overpotential for each sample catalyst in the training set O1 using the target network O1. During the training process using the training set O1, the training system 100 uses 25% of the data in the training set O1 as test data, which determines the test R of the target network O1. 2 The value is 0.732, and the MAE is 9.31. Figure 13 b) gives the predicted values ​​of the initial overpotential for the sample catalyst composed of Fe, Co, and La through the target network O1.

[0217] against Figure 10 In region b) with smaller Tafel slopes, 21 experiments were conducted simultaneously to measure the initial overpotential. The measurement results are shown in [reference needed]. Figure 13 d).

[0218] Figure 13 c) provides the predicted and experimental values ​​of the initial overpotential for the sample catalyst composed of Fe, Co, and La using the target network O1. R can be obtained from this. 2 It is 0.550.

[0219] Figure 13 (e) presents the predicted values ​​and experimental values ​​of the catalyst samples in the training set O2 using the target network O2 regarding the initial overpotential. During the optimization process using the training set O2, the training system 100 uses 35% of the data in the training set O2 as test data, thereby determining the test R of the target network O2. 2 The value increased to 0.905, and the MAE decreased to 7.57.

[0220] Figure 13 f) provides a comparison between the predicted and experimental values ​​of the target network O2 for the sample catalyst composed of Fe, Co, and La regarding the initial overpotential, from which R can be obtained. 2 Increased from 0.550 to 0.925.

[0221] Figure 14 (a) presents the predicted values ​​of the catalyst samples in the training set O3 with respect to the initial overpotential, and compares them with experimental values, obtained through the target network O3. During the training process using the training set O3, the training system 100 randomly uses 35% of the data from the training set O3 as test data, thereby determining the test R of the target network O3. 2 The value is 0.916, which is higher than the test R of the target network O2. 2 value.

[0222] Furthermore, the predicted onset overpotentials of the target network O3 for the 15 electrocatalysts in Table 3 are shown in Table 6 and [other tables not provided in the original text]. Figure 14 (b). See Table 6. The maximum prediction error of the initial overpotential of the target network O3 for each catalyst is 11.8%, and the minimum prediction error is 0.03%. More than 5% of the error occurs in the composition of elements that are far away from Fe, Co, Ni, Ce and La on the periodic table, such as Cu in Cu1Co1La1.

[0223] Figure 14 c) presents the predicted values ​​of the catalyst samples in the training set O4 with respect to the initial overpotential through the target network O4, and compares them with the experimental values, testing R... 2 It reached 0.925. Figure 14 d) presents the predicted values ​​and experimental values ​​of the onset overpotential for each sample catalyst in Table 3 using the target network O4, R. 2 It is 0.938.

[0224] Referring to Table 6, the maximum prediction error of the target network O4 for the initial overpotential of each catalyst is 10.4%, and the minimum prediction error is 0.01%.

[0225] Figure 15The experimental and predicted values ​​of the initial overpotential of sample catalysts with different elemental compositions are illustrated. Specifically: a) shows the experimental values ​​of the initial overpotential of the sample catalyst composed of Fe, Co, and Ni; b) shows the experimental values ​​of the initial overpotential of the sample catalyst composed of Fe, Co, and Ce; c) shows the experimental values ​​of the initial overpotential of the sample catalyst composed of Fe, Co, and La; d) shows the predicted values ​​of the initial overpotential of the sample catalyst composed of Fe, Co, and Ni by the target network O4; e) shows the predicted values ​​of the initial overpotential of the sample catalyst composed of Fe, Co, and Ce by the target network O4; and f) shows the predicted values ​​of the initial overpotential of the sample catalyst composed of Fe, Co, and La by the target network O4.

[0226] It is evident that the predicted values ​​of the target network O4 for each sample catalyst are highly consistent with the experimental values. This approach, combining experimentation and prediction, significantly reduces repetitive experiments and provides timely feedback for further research on multi-element electrocatalysts. This demonstrates the strong predictive ability of the target network O with a relatively small number of sample catalysts.

[0227] Referring to Table 6, the training system 100 can determine that the prediction error of the target network O4 is 4%. When the prediction error is less than or equal to a preset threshold 6, the training system 100 can use the target network O4 as the final optimized target network O and deploy it in the prediction system 200. If the prediction error is greater than the preset threshold 6, the training system 100 can continue training based on other sample catalysts until the prediction error is less than or equal to the preset threshold 6, at which point the final optimized target network O can be deployed in the prediction system 200.

[0228] The number of levels in the target network T and the target network O can be different, and the number of subnetworks included in each level can also be different. This specification does not impose any restrictions on this.

[0229] Figure 16 A schematic diagram is given showing the relationship between the prediction performance of the target network T on the Tafel slope and the number of layers.

[0230] like Figure 16 As shown, the number of subnetworks in each layer of the target network T is represented by the numbers in parentheses. Specifically, a) shows the relationship between the prediction performance of the target network T1 on the Tafel slope and the number of layers; b) shows the relationship between the prediction performance of the target network T2 on the Tafel slope and the number of layers; c) shows the relationship between the prediction performance of the target network T3 on the Tafel slope and the number of layers; and d) shows the relationship between the prediction performance of the target network T4 on the Tafel slope and the number of layers. Figure 16 We can see that testing R 2The variance increases with the number of network layers, while the variance decreases. Furthermore, when the number of network layers in the four target networks reaches five, the test R... 2 The learning process tends to stabilize, meaning that the target network T reaches saturation when the number of network layers is 5. Therefore, the training system 100 can set the number of layers of the target network T to 5 and set the number of subnetworks in each layer to the number shown in Figure d), but this specification is not limited to this.

[0231] Figure 17 A schematic diagram is given showing the relationship between the prediction performance of the target network O on the initial overpotential and the number of layers.

[0232] like Figure 17 As shown, the number of subnetworks in each level of the target network O is indicated by the numbers in parentheses. Specifically, a) shows the relationship between the prediction performance of target network O1 for the initial overpotential and the number of levels; b) shows the relationship between the prediction performance of target network O2 for the initial overpotential and the number of levels; c) shows the relationship between the prediction performance of target network O3 for the initial overpotential and the number of levels; and d) shows the relationship between the prediction performance of target network O4 for the initial overpotential and the number of levels. Figure 17 We can see that testing R 2 The variance increases with the number of network layers, while the variance decreases. Furthermore, when the number of network layers in the four target networks reaches five, the test R... 2 The learning process tends to stabilize, meaning that the target network O reaches saturation when the number of network layers is 5. Therefore, the training system 100 can set the number of layers of the target network O to 5 and set the number of subnetworks in each layer to the number shown in Figure d), but this specification is not limited to this.

[0233] Figure 18 A comparison of the prediction performance of different networks for Tafel slope and initial overpotential is presented. Specifically, a) shows the comparison of the prediction performance of different networks for Tafel slope; b) shows the comparison of the prediction performance of different models for initial overpotential.

[0234] like Figure 18 As shown, 15, 145, and 909 represent the number of input descriptors corresponding to different networks. For example, '15' represents the 15 most frequent descriptors in the preferred descriptor combinations based on the GA algorithm, '145' represents the number of descriptors used in published literature, and '909' is the number of descriptors proposed in this specification. ANN, XGBoost, and Hierarchical Neural Network (HNN) represent different AI networks, such as target network T or target network O.

[0235] '15-ANN' means training an ANN using samples described by 15 descriptors, '145-HNN' means training an HNN using samples described by 145 descriptors, '909-ANN' means training an ANN using samples described by 909 descriptors, '909-XGBoost' means training an XGBoost using samples described by 909 descriptors, and '909-HNN' means training an HNN using samples described by 909 descriptors.

[0236] See Figure 18 It can be determined that the predictive performance of '909-ANN' is lower than that of '15-ANN' and '909-XGBoost'. This indicates that after increasing the number of descriptors describing the sample catalysts, the ANN failed to resolve the contradiction between the large number of descriptors and the small number of sample catalysts; in fact, the predictive performance of the ANN decreased with more descriptors. The predictive performance of '909-HNN', due to its superior performance compared to all other networks, specifically the HNN provided in this specification, resolves the long-standing contradiction between the large number of descriptors and the insufficient training samples in traditional ANNs for materials science problems. In conjunction with the above, the target network T3 or target network O3 did not exhibit overfitting when predicting the various catalysts in Table 3 (including elements not present in the training samples of target networks T3 or O3), indicating that the predictive accuracy of the HNN provided in this specification for catalytic performance can be used for catalyst research.

[0237] In summary, the training method and system for the target network provided in this specification expand the number of descriptors describing the sample catalysts, enrich the descriptive dimensions of the sample catalysts, and allow for parallel training of subnetworks by grouping K descriptors into multiple levels of the target network, avoiding the contradiction between an excessive number of descriptors and insufficient training samples. First, the significant increase in the number of descriptors combined with the target network can improve the prediction accuracy of catalytic performance; even a small number of training samples can significantly enhance the predictive power of the target network, reducing experimental costs in catalyst research. Second, when grouping descriptors, each group includes the same core descriptor and some non-core descriptors, considering both the importance of the core descriptor and exploring the potential impact of different non-core descriptor combinations on catalyst performance. Combined with the predicted performance data given above, the prediction accuracy of the target network is higher than that of other networks, which is beneficial for accelerating the process of innovative catalyst material research.

[0238] Figure 19A flowchart of a catalyst performance prediction method P1900 according to an embodiment of this specification is shown. The prediction system 200 can execute the catalyst performance prediction method P1900 to predict the target performance of a target catalyst, the target catalyst comprising at least one element. Figure 19 As shown, the catalyst performance prediction method P1900 includes the following steps.

[0239] S1910: Based on the proportion information of the at least one element in the target catalyst and the attribute values ​​of the X attributes corresponding to each of the at least one element, generate K descriptors to describe the target catalyst.

[0240] In some embodiments, the X attributes include at least two of the following dimensions: periodic table dimension; thermodynamic dimension; physical dimension; and crystallographic dimension.

[0241] Specifically, the X attributes can be found in Table 4, and will not be repeated here.

[0242] In some embodiments, the K descriptors include at least one of the following:

[0243] The first set of descriptors is used to describe the statistical characteristics of the target catalyst in the X properties;

[0244] The second set of descriptors is used to describe the valence electron occupation state of the target catalyst;

[0245] The third set of descriptors is used to describe the ionicity of the target catalyst; and

[0246] The fourth set of descriptors is used to describe the configuration entropy of the target catalyst.

[0247] In some embodiments, the first set of descriptors is generated as follows:

[0248] For each of the X attributes, the prediction system 200 performs statistical analysis on the characteristics of the target catalyst in the attribute from multiple statistical dimensions based on the proportion information of the at least one element in the target catalyst and the attribute values ​​of the at least one element in the attribute, and obtains P statistical values; and generates a first set of descriptors based on the P statistical values ​​of the target catalyst in each of the X attributes, wherein the number of the first set of descriptors is equal to X*P.

[0249] Specifically, the generation method of the first set of descriptors corresponding to the target catalyst is similar to the generation method of the first set of descriptors corresponding to sample catalyst #1. For details on the generation method, please refer to the above text, which will not be repeated here.

[0250] In some embodiments, the second set of descriptors includes: occupancy information of Q valence electrons in the target catalyst, wherein the occupancy information of the z-th valence electron is obtained in the following manner:

[0251] For each of the at least one element, the prediction system 200 determines the number of each valence electron and the total number of valence electrons in the element based on the attribute values ​​of at least some of the X attributes corresponding to the element; for the z-th valence electron, the prediction system 200 determines the occupancy information of the z-th valence electron based on the proportion information of the at least one element in the target catalyst, the number of the z-th valence electrons corresponding to each of the at least one element, and the total number of valence electrons corresponding to each of the at least one element.

[0252] Specifically, the generation method of the second set of descriptors corresponding to the target catalyst is similar to that of the second set of descriptors corresponding to sample catalyst #1. For details on the generation method, please refer to the above text, which will not be repeated here.

[0253] In some embodiments, the third set of descriptors is obtained based on the proportion of the at least one element in the target catalyst, the electronegativity of each of the at least one element, and any of the following:

[0254] The maximum value of the electronegativity corresponding to each of the at least one element;

[0255] The average value of the electronegativity of each of the at least one element.

[0256] Specifically, the generation method of the third set of descriptors corresponding to the target catalyst is similar to that of the third set of descriptors corresponding to sample catalyst #1. For details on the generation method, please refer to the above text, which will not be repeated here.

[0257] In some embodiments, the fourth set of descriptors is obtained based on the proportion of the at least one element in the target catalyst, the Boltzmann constant, and room temperature.

[0258] Specifically, the generation method of the fourth set of descriptors corresponding to the target catalyst is similar to that of the fourth set of descriptors corresponding to sample catalyst #1. For details on the generation method, please refer to the above text, which will not be repeated here.

[0259] S1920: Among the K descriptors, N1 core descriptors and K1 non-core descriptors are determined. The core descriptors have a greater impact on the target performance than the non-core descriptors.

[0260] In some embodiments, the target performance is the Tafel slope, and the N1 core descriptors include multiples of the following:

[0261] MDT1, representing the average deviation of the electronegativity of the different elements in the at least one element;

[0262] MDT2 represents the maximum value of the weighted attribute value of the at least one element over the BCC atomic energy difference;

[0263] MDT3 represents the minimum weighted attribute value of the at least one element in a row of the periodic table;

[0264] MDT4, characterizing the maximum deviation of the BCC atomic energy difference among the at least one element; and

[0265] MDT5 represents the average deviation of the different elements in the work function among the at least one element.

[0266] For example, after the prediction system 200 determines the N1 core descriptors in MDT1 to MDT5, it uses the remaining descriptors in the K descriptors as non-core descriptors.

[0267] In some embodiments, the target performance is the initial overpotential, and the N1 core descriptors include multiples of the following:

[0268] MDO1 represents the maximum value of the weighted attribute values ​​of the at least one element in terms of the number of unfilled valence electrons (d).

[0269] MDO2 represents the average value of the weighted attribute deviation of different elements in mass density among the at least one element;

[0270] MDO3 represents the minimum value of the weighted attribute value deviation of different elements in the work function among the at least one element;

[0271] MDO4, representing the average deviation of the number of valence electrons d-filled among the at least one element; and

[0272] MDO5 represents the maximum value of the weighted attribute value deviation of different elements in the at least one element on the covalent radius.

[0273] For example, after the prediction system 200 determines the N1 core descriptors from MDO1 to MDO5, it uses the remaining descriptors among the K descriptors as non-core descriptors.

[0274] S1930: Input at least a portion of the K descriptors into the target network for prediction to obtain the performance index corresponding to the target performance. The target network is a network trained using Y sample catalysts. The target network includes multiple sub-networks with the same structure and arranged hierarchically. Each sub-network includes N input nodes and 1 output node. The output node of the sub-network in the i-th level serves as the input node of the sub-network in the (i+1)-th level. Each sub-network in the first level is input with the N1 core descriptors and N2 non-core descriptors. The N2 non-core descriptors are a portion of the K1 non-core descriptors, and the non-core descriptors input to different sub-networks are not completely identical. The ratio of N to Y is less than 1, and the ratio of K to Y is greater than 1.

[0275] In some embodiments, the prediction system 200 generates m1 sets of descriptors based on the K descriptors, each set including the N1 core descriptors and N-N1 non-core descriptors; the m1 sets of descriptors are input into the m1 sub-networks of the first layer of the target network to obtain m1 intermediate performance indicators of the target performance; the m1 sub-networks of the (i-1)th layer of the target network are then processed according to the order of i values ​​from 2 to F. 1-1 The target performance m obtained from the subnetwork i-1 Each intermediate performance metric is divided into m i Grouping inputs the m of the i-th layer of the target network i A subnetwork is used to obtain the target performance m. i There are several intermediate performance metrics; and the intermediate performance metrics of the target performance output by the subnetwork of layer F are used as the final performance metrics of the target performance.

[0276] For example, when the target performance is the Tafel slope, the prediction system 200 can invoke the target network T. Specifically, the prediction system 200 can input the m1 sets of descriptors respectively as follows: Figure 9 The target network T shown has m1 subnetworks in its first layer, from which m1 intermediate performance metrics of the Tafel slope are obtained; the m1 subnetworks in the (i-1)th layer of the target network T are then processed according to the order of i from 2 to F. 1-1 The Tafel slope m obtained from each subnetwork i-1 Each intermediate performance metric is divided into m i Group input to the target network T, layer m of the i-th layer i A subnetwork is used to obtain the m-th slope of Tafel. i There are several intermediate performance metrics; then the intermediate performance metrics of the Tafel slope output by the subnetwork of layer F are used as the final performance metrics of the Tafel slope.

[0277] For example, when the target performance is at the initial overpotential, the prediction system 200 can invoke the target network O. Specifically, the prediction system 200 can input the m1 sets of descriptors respectively as follows: Figure 9 The target network O shown has m1 subnetworks in its first layer, from which m1 intermediate performance indices for the initial overpotential are obtained; the m1 subnetworks in the (i-1)th layer of the target network O are then processed according to the order of i values ​​from 2 to F. 1-1 The initial overpotential m obtained from the subnetwork i-1 Each intermediate performance metric is divided into m i Group input to target network O, layer m of the i-th layer i A sub-network is used to obtain the m-th potential at the initial overpotential. i The intermediate performance index is then used as the final performance index of the Tafel slope, which is the initial overpotential output of the sub-network of the F layer.

[0278] It is understood that some technical details involved in the catalyst performance prediction method P1900 can be referred to the description in the target network training method P500, and will not be repeated here. The beneficial effects of the various embodiments in the above method P1900 can also be referred to the description of the beneficial effects of the embodiments in method P500, and will not be repeated here.

[0279] In summary, the catalyst performance prediction method and training system provided in this specification expand the number of descriptors describing the target catalyst, enrich the descriptive dimensions of the target catalyst, and allow for parallel prediction by grouping K descriptors into sub-networks through a multi-level target network. First, the significant increase in the number of descriptors, combined with the target network, improves the accuracy of predicting the catalytic performance of the target catalyst, reducing experimental costs in catalyst research. Second, when grouping descriptors, each group includes the same core descriptor and some non-core descriptors, considering both the importance of the core descriptor and exploring the potential impact of different non-core descriptor combinations on the catalytic performance of the target catalyst. Combined with the prediction data presented above, the target network exhibits higher prediction accuracy than other networks, which is beneficial for accelerating the process of innovative catalyst material research.

[0280] This specification, in another aspect, provides a computer-readable non-transitory storage medium storing at least one set of executable instructions for predicting catalyst performance or training a target network. When the executable instructions are executed by a processor, they instruct the processor to implement the steps of the catalyst performance prediction method P1900 or the target network training method P500 described herein. In some possible embodiments, various aspects of this specification can also be implemented as a program product comprising program code. When the program product is run on system 400, the program code causes system 400 to perform the steps of the catalyst performance prediction method P1900 or the target network training method P500 described herein. The program product for implementing the above methods may employ a portable compact disk read-only memory (CD-ROM) containing program code and may run on system 400. However, the program product of this specification is not limited thereto. In this specification, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system. The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. The computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing the operations described herein may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages.The program code can be executed entirely on system 400, partially on system 400, as a standalone software package, partially on system 400 and partially on a remote computing device, or entirely on a remote computing device.

[0281] The term "and / or" in the embodiments of this specification describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0282] The terms “first”, “second”, etc., used in this specification are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence.

[0283] Unless otherwise stated, the term "multiple" in this specification shall be understood as two or more.

[0284] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0285] In summary, after reading this detailed disclosure, those skilled in the art will understand that the foregoing detailed disclosure is presented by way of example only and is not restrictive. Although not explicitly stated herein, those skilled in the art will understand that this specification requires various reasonable changes, improvements, and modifications to the embodiments. These changes, improvements, and modifications are intended to be made by this specification and are within the spirit and scope of the exemplary embodiments described herein.

[0286] Furthermore, certain terms in this specification have been used to describe embodiments of this specification. For example, "an embodiment," "an embodiment," and / or "some embodiments" mean that a particular feature, structure, or characteristic described in connection with that embodiment may be included in at least one embodiment of this specification. Therefore, it is to be emphasized and understood that two or more references to "an embodiment" or "an embodiment" or "alternative embodiment" in various parts of this specification do not necessarily refer to the same embodiment. Moreover, specific features, structures, or characteristics may be suitably combined in one or more embodiments of this specification.

[0287] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and aiding in the understanding of a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art may readily identify some of the devices as separate embodiments when reading this specification. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. It is also valid when each secondary embodiment contains fewer than all the features of a single foregoing disclosed embodiment.

[0288] Every patent, patent application, publication of a patent application, and other material, such as articles, books, specifications, publications, documents, and literature (excluding any related historical examination documents), cited in this disclosure is incorporated herein for all purposes, including, for example, in the specification and claims of this disclosure. However, in the event of any inconsistency or conflict between the descriptions, definitions, and / or terms used in the foregoing and those used in this disclosure, the descriptions, definitions, and / or terms used in this disclosure shall prevail.

[0289] Finally, it should be understood that the embodiments disclosed herein are illustrative of the principles of the embodiments described in this specification. Other modified embodiments are also within the scope of this specification. Therefore, the embodiments disclosed in this specification are merely examples and not limitations. Those skilled in the art can implement the applications described in this specification using alternative configurations based on the embodiments in this specification. Therefore, the embodiments in this specification are not limited to the embodiments precisely described in the applications.

Claims

1. A method for predicting catalyst performance, configured to predict the target performance of a target catalyst, said target catalyst comprising at least one element, the method comprising: Based on the proportion information of the at least one element in the target catalyst and the attribute values ​​of the X attributes corresponding to each of the at least one element, K descriptors are generated to describe the target catalyst. N1 core descriptors and K-N1 non-core descriptors are determined from the K descriptors. The core descriptors have a greater impact on the target performance than the non-core descriptors. as well as At least a portion of the K descriptors are input into the target network for prediction to obtain the performance metric corresponding to the target performance, wherein... The target network is a network trained using Y sample catalysts. The target network comprises multiple sub-networks with identical structures arranged hierarchically. Each sub-network includes N input nodes and 1 output node. The output node of the sub-network in the i-th level serves as the input node of the sub-network in the (i+1)-th level. In the first layer, each sub-network is input with the N1 core descriptors and N-N1 non-core descriptors, where the N-N1 non-core descriptors are a subset of the K-N1 non-core descriptors. Furthermore, the non-core descriptors input to different subnetworks are not entirely the same. The ratio of N to Y is less than 1, and the ratio of K to Y is greater than 1; Wherein, X, K, N1, Y, and N are all integers greater than 1, and i is an integer greater than or equal to 1.

2. The method according to claim 1, wherein, The X attributes include at least two of the following dimensions: Dimensions of the periodic table; Thermodynamic dimension; Physical dimension; or Crystallographic dimension.

3. The method according to claim 1, wherein, The K descriptors include at least one of the following: The first set of descriptors is used to describe the statistical characteristics of the target catalyst in the X properties; The second set of descriptors is used to describe the valence electron occupation state of the target catalyst; The third set of descriptors is used to describe the ionicity of the target catalyst; as well as The fourth set of descriptors is used to describe the configuration entropy of the target catalyst.

4. The method according to claim 3, wherein, The first set of descriptors was generated in the following way: For each of the X attributes, based on the proportion information of the at least one element in the target catalyst and the attribute value of the at least one element in the attribute, P statistical values ​​are obtained from multiple statistical dimensions to statistically analyze the characteristics of the target catalyst in the attribute. as well as Based on the P statistical values ​​corresponding to each of the X properties of the target catalyst, a first set of descriptors is generated, and the number of descriptors in the first set is equal to X*P.

5. The method according to claim 3, wherein, The second set of descriptors includes: the occupancy information of Q valence electrons in the target catalyst, wherein the occupancy information of the z-th valence electron is obtained in the following manner: For each of the at least one element, based on the attribute values ​​of at least some of the X attributes corresponding to the element, determine the quantity of each type of valence electron and the total number of valence electrons in the element; and For the z-th valence electron, the occupancy information of the z-th valence electron is determined based on the proportion information of the at least one element in the target catalyst, the number of the z-th valence electrons corresponding to each of the at least one element, and the total number of valence electrons corresponding to each of the at least one element.

6. The method according to claim 3, wherein, The third set of descriptors is obtained based on the proportion information of the at least one element in the target catalyst, the electronegativity of each of the at least one element, and any one of the following: The maximum value of the electronegativity corresponding to each of the at least one element; or The average value of the electronegativity of each of the at least one element.

7. The method according to claim 3, wherein, The fourth set of descriptors is obtained based on the proportion of the at least one element in the target catalyst, the Boltzmann constant, and room temperature.

8. The method according to claim 1, wherein, The target performance is the Tafel slope, and the N1 core descriptors include multiples of the following: A descriptor characterizing the average deviation of the electronegativity of the different elements in the at least one element; A descriptor representing the maximum value of the weighted attribute values ​​of the at least one element over the BCC atomic energy difference; A descriptor representing the minimum value of the weighted attribute values ​​of the at least one element in a row of the periodic table; A descriptor characterizing the maximum deviation of the different elements in the BCC atomic energy difference among the at least one element; as well as A descriptor characterizing the average deviation of the different elements in the work function of the at least one element.

9. The method according to claim 1, wherein, The target performance is the initial overpotential, and the N1 core descriptors include multiples of the following: A descriptor representing the maximum value among the weighted attribute values ​​of the at least one element in terms of the number of unfilled valence electrons; A descriptor representing the average of the weighted attribute values ​​of different elements in mass density among the at least one element; A descriptor representing the minimum value of the weighted attribute value deviation of the different elements in the work function among the at least one element; A descriptor characterizing the average deviation of the number of valence electrons d-filled among the at least one element; as well as A descriptor representing the maximum value of the weighted attribute value deviation of different elements in the at least one element at the covalent radius.

10. The method according to claim 1, wherein, The target network comprises F layers. The step of inputting at least a portion of the K descriptors into the target network for prediction to obtain the performance metrics corresponding to the target performance includes: Based on the K descriptors, generate m1 groups of descriptors, each group of descriptors including the N1 core descriptors and N-N1 non-core descriptors; The m1 sets of descriptors are respectively input into the m1 sub-networks of the first layer of the target network to obtain the m1 intermediate performance indicators of the target performance; Following the order of i values ​​from 2 to F, the m values ​​of the (i-1)th layer of the target network are... 1-1 The target performance m obtained from the subnetwork i-1 Each intermediate performance metric is divided into m i The m-th layer of the target network is then input after the group. i A sub-network is used to obtain m of the target performance. i One intermediate performance metric; and The intermediate performance index of the target performance output by the sub-network of layer F is used as the final performance index of the target performance. Wherein, m1, F, and m i-1 All are integers greater than 1, and m i It is an integer greater than or equal to 1.

11. A method for training a target network, said target network being used to predict the target performance of a catalyst, the method comprising: Obtain experimental values ​​for the target performance of Y sample catalysts, each sample catalyst comprising at least one element; For each sample catalyst, based on the proportion information of the at least one element in the sample catalyst and the attribute values ​​of X attributes corresponding to each of the at least one element, K descriptors are generated to describe the sample catalyst. Among the K descriptors, N1 core descriptors and K-N1 non-core descriptors are determined. The core descriptors have a greater impact on the target performance than the non-core descriptors. A subset of the K descriptors corresponding to each sample catalyst is input into the target network for prediction, yielding the predicted value corresponding to the target performance, wherein: The target network comprises multiple sub-networks with identical structures arranged hierarchically. Each sub-network includes N input nodes and 1 output node. The output node of the sub-network in the i-th level serves as the input node of the sub-network in the (i+1)-th level. In the first layer, each sub-network is input with the N1 core descriptors and N-N1 non-core descriptors, where the N-N1 non-core descriptors are a subset of the K-N1 non-core descriptors. Furthermore, the non-core descriptors input to different subnetworks are not entirely the same. The ratio of N to Y is less than 1, and the ratio of K to Y is greater than 1; The parameters of the target network are updated with the training objective of minimizing the difference between the predicted value and the experimental value. Wherein, X, K, N1, Y, and N are all integers greater than 1, and i is an integer greater than or equal to 1.

12. The method according to claim 11, wherein, The X attributes include at least two of the following dimensions: Dimensions of the periodic table; Thermodynamic dimension; Physical dimension; or Crystallographic dimension.

13. The method according to claim 11, wherein, The K descriptors include at least one of the following: The first set of descriptors is used to describe the statistical characteristics of the sample catalyst on the X properties; The second set of descriptors is used to describe the valence electron occupation state of the sample catalyst; The third set of descriptors is used to describe the ionicity of the sample catalyst; as well as The fourth set of descriptors is used to describe the configuration entropy of the sample catalyst.

14. The method according to claim 13, wherein, The first set of descriptors was generated in the following way: For each of the X attributes, based on the proportion information of the at least one element in the sample catalyst and the attribute value of the at least one element in the attribute, P statistical values ​​are obtained from multiple statistical dimensions to statistically analyze the characteristics of the sample catalyst in the attribute. as well as Based on the P statistical values ​​corresponding to each of the X attributes of the sample catalyst, a first set of descriptors is generated, and the number of descriptors in the first set is equal to X*P.

15. The method according to claim 13, wherein, The second set of descriptors includes: the occupancy information of Q valence electrons in the sample catalyst, wherein the occupancy information of the z-th valence electron is obtained in the following manner: For each of the at least one element, based on the attribute values ​​of at least some of the X attributes corresponding to the element, determine the quantity of each type of valence electron and the total number of valence electrons in the element; and For the z-th valence electron, the occupancy information of the z-th valence electron is determined based on the proportion information of the at least one element in the sample catalyst, the number of the z-th valence electrons corresponding to each of the at least one element, and the total number of valence electrons corresponding to each of the at least one element.

16. The method according to claim 13, wherein, The third set of descriptors is obtained based on the proportion of the at least one element in the sample catalyst, the electronegativity of each of the at least one element, and any of the following: The maximum value of the electronegativity corresponding to at least one element; or The average value of the electronegativity corresponding to at least one element.

17. The method according to claim 13, wherein, The fourth set of descriptors is obtained based on the proportion of the at least one element in the sample catalyst, the Boltzmann constant, and room temperature.

18. The method according to claim 11, wherein, The target performance is the Tafel slope, and the N1 core descriptors include multiples of the following: A descriptor characterizing the average deviation of the electronegativity of the different elements in the at least one element; A descriptor representing the maximum value of the weighted attribute values ​​of the at least one element over the BCC atomic energy difference; A descriptor representing the minimum value of the weighted attribute values ​​of the at least one element in a row of the periodic table; A descriptor characterizing the maximum deviation of the different elements in the BCC atomic energy difference among the at least one element; as well as A descriptor characterizing the average deviation of the different elements in the work function of the at least one element.

19. The method according to claim 11, wherein, The target performance is the initial overpotential, and the N1 core descriptors include multiples of the following: A descriptor representing the maximum value among the weighted attribute values ​​of the at least one element in terms of the number of unfilled valence electrons; A descriptor representing the average of the weighted attribute values ​​of different elements in mass density among the at least one element; A descriptor representing the minimum value of the weighted attribute value deviation of the different elements in the work function among the at least one element; A descriptor characterizing the average deviation of the number of valence electrons d-filled among the at least one element; as well as A descriptor representing the maximum value of the weighted attribute value deviation of different elements in the at least one element at the covalent radius.

20. The method according to claim 11, wherein, The target network comprises F levels. The step of inputting a portion of the K descriptors corresponding to each sample catalyst into the target network for prediction to obtain the predicted value corresponding to the target performance includes: Based on the K descriptors, generate m1 groups of descriptors, each group of descriptors including the N1 core descriptors and N-N1 non-core descriptors; The m1 sets of descriptors are respectively input into the m1 sub-networks of the first layer of the target network to obtain the m1 intermediate prediction values ​​of the target performance; Following the order of i from 2 to F, the m values ​​of the (i-1)th layer of the target network are... 1-1 The target performance m obtained from the subnetwork i-1 The intermediate predicted values ​​are divided into m i After grouping, input the m of the i-th layer of the target network. i A sub-network is used to obtain m of the target performance. i One intermediate predicted value; and The intermediate predicted value of the target performance output by the sub-network of layer F is used as the final predicted value of the target performance; Wherein, m1, F, and m i-1 All are integers greater than 1, and m i It is an integer greater than or equal to 1.

21. A prediction system, characterized in that, The target performance of the catalyst is configured to be predicted, including: At least one storage medium storing at least one instruction set; and At least one processor is communicatively connected to the at least one storage medium. When the prediction system is running, the at least one processor reads the at least one instruction set and executes the method of any one of claims 1-10 according to the instructions of the at least one instruction set.

22. A training system, characterized in that, The target network configured to be trained for predicting the target performance of the catalyst includes: At least one storage medium storing at least one instruction set; and At least one processor is communicatively connected to the at least one storage medium. When the training system is running, the at least one processor reads the at least one instruction set and executes the method according to any one of claims 11-20 according to the instructions of the at least one instruction set.

Images