Material attribute prediction method and device based on modular deep learning

Abstract

The invention provides a material attribute prediction method and device based on modular deep learning, and the method and device are applied to a cloud material knowledge sharing platform, the cloud material knowledge sharing platform is loaded with a material knowledge base, the material knowledge base comprises multiple groups of parameter modules, and one group of parameter modules corresponds to one heterogeneous material task. The heterogeneous material task is used for realizing attribute prediction of the material; the method comprises the following steps: receiving a target attribute prediction problem of a target material uploaded by a user; querying in a material knowledge base based on the target attribute prediction problem, determining a corresponding target parameter module, and returning the target parameter module to the user; wherein the target parameter module is used for enabling a user to train a model based on the target parameter module, and performing attribute prediction of a target material based on the trained model. Through the method provided by the invention, the module most relevant to the given downstream task is selected from the material knowledge base, and the downstream material attribute prediction task is effectively adapted.

Description

Technical Field

[0001] The present invention relates to the field of material big data technology, and in particular to a material property prediction method and device based on modular deep learning. Background Art

[0002] Existing deep learning-based material knowledge sharing solutions can be divided into two categories: (1) pre-training a neural network on a specific material task and fine-tuning the network on downstream tasks; (2) pre-training a neural network skeleton on all source tasks and directly integrating all networks for downstream task fine-tuning. The disadvantages of existing technologies are: (1) Lack of flexibility. Existing solutions share knowledge from the same source task for each downstream task, and are unable to decouple and filter upstream heterogeneous material knowledge, which easily leads to negative transfer of material knowledge; (2) Parameter inefficiency. Existing solutions all require copying the complete pre-trained network to achieve knowledge sharing. Given the over-parameterization characteristics of neural networks, existing solutions use increased space costs to store material knowledge, which is not conducive to the transmission and sharing of knowledge.

[0003] How to effectively achieve knowledge decoupling and sharing between different material properties is a technical problem that needs to be solved at present. Summary of the invention

[0004] The present invention provides a material property prediction method and device based on modular deep learning, so as to solve the defects existing in the prior art.

[0005] The present invention provides a material property prediction method based on modular deep learning, which is applied to a cloud-based material knowledge sharing platform, wherein the cloud-based material knowledge sharing platform is loaded with a material knowledge base, wherein the material knowledge base includes: multiple groups of parameter modules, wherein one group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize material property prediction; The steps include: Receive target attribute prediction problem of target materials uploaded by users; Based on the target property prediction problem, the material knowledge base is queried, the corresponding target parameter module is determined, and the target parameter module is returned to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0006] According to a material property prediction method based on modular deep learning provided by the present invention, before receiving a target property prediction problem of a target material uploaded by a user, the method further includes: Obtaining a pre-trained material representation skeleton; wherein the pre-trained material representation skeleton is used to mount a parameter module; For a plurality of preset heterogeneous material tasks, a corresponding parameter module is designed for each heterogeneous material task and mounted on the pre-trained material representation skeleton; wherein the parameter module is used to store heterogeneous knowledge corresponding to the heterogeneous material task; The parameter module corresponding to each heterogeneous material task is stored to obtain the material knowledge base.

[0007] According to a material property prediction method based on modular deep learning provided by the present invention, for a plurality of preset heterogeneous material tasks, a corresponding parameter module is designed for each heterogeneous material task and mounted on the pre-trained material representation skeleton, comprising: For each heterogeneous material task, generate a corresponding initial parameter module, and mount the initial parameter module onto the pre-trained material representation skeleton; The heterogeneous knowledge of the heterogeneous material task is routed to the corresponding initial parameter module through a routing network to obtain the parameter module corresponding to the heterogeneous material task.

[0008] According to a material property prediction method based on modular deep learning provided by the present invention, the parameter module corresponding to each heterogeneous material task in the plurality of heterogeneous material tasks is stored to obtain the material knowledge base, including: Jointly train parameter modules; When the training meets the convergence condition, the trained parameter module is stored to obtain the material knowledge base.

[0009] According to a material property prediction method based on modular deep learning provided by the present invention, the method of querying in the material knowledge base based on the target property prediction problem to determine the corresponding target parameter module includes: Based on the target attribute prediction problem, a plurality of target lightweight parameter submodules are obtained by querying in the material knowledge base through an adaptive modular assembly AMA algorithm; wherein the combination of the plurality of target lightweight parameter submodules is: a module combination that minimizes the generalization error; The multiple target lightweight parameter sub-modules are assembled and spliced ​​to obtain a target parameter module corresponding to the target attribute prediction problem.

[0010] According to a material property prediction method based on modular deep learning provided by the present invention, based on the target property prediction problem, an adaptive modular assembly AMA algorithm is used to query in the material knowledge base to obtain multiple target lightweight parameter sub-modules, including: The prediction algorithm based on label propagation is used to calculate and obtain the prediction performance of each parameter module on the target heterogeneous material task corresponding to the target attribute prediction problem; Based on the prediction performance of each parameter module, the multiple target lightweight parameter sub-modules are obtained through modular weight optimization based on a target algorithm; wherein the target algorithm is an optimization algorithm that does not require training.

[0011] According to a material property prediction method based on modular deep learning provided by the present invention, the multiple target lightweight parameter submodules are assembled and spliced ​​to obtain a target parameter module corresponding to the target property prediction problem, including: Assembling the multiple target lightweight parameter submodules, and splicing the assembled multiple target lightweight parameter submodules to the pre-trained material representation skeleton to generate a deep learning model of the target heterogeneous material task corresponding to the target attribute prediction problem; The deep learning model is jointly fine-tuned to perform a target attribute prediction parameter module based on the jointly fine-tuned deep learning model.

[0012] The present invention also provides a material property prediction device based on modular deep learning, which is applied to a cloud-based material knowledge sharing platform, wherein the cloud-based material knowledge sharing platform is loaded with a material knowledge base, wherein the material knowledge base includes: a plurality of groups of parameter modules, wherein a group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize material property prediction; Includes the following modules: A receiving module, used for receiving target attribute prediction problems of target materials uploaded by users; A determination module is used to query the material knowledge base based on the target property prediction problem, determine the corresponding target parameter module, and return the target parameter module to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements any of the above-described methods for predicting material properties based on modular deep learning.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for predicting material properties based on modular deep learning.

[0015] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned methods for predicting material properties based on modular deep learning.

[0016] The present invention provides a material property prediction method and device based on modular deep learning, which is applied to a cloud-based material knowledge sharing platform. The cloud-based material knowledge sharing platform is loaded with a material knowledge base, and the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize the material property prediction; by receiving the target property prediction problem of the target material uploaded by the user; based on the target property prediction problem, querying in the material knowledge base, determining the corresponding target parameter module, and returning the target parameter module to the user; wherein the target parameter module is used to: enable the user to train the model based on the target parameter module, and predict the properties of the target material based on the trained model. It can be seen that the present invention trains a series of parameter modules containing decoupled material knowledge for data-rich tasks, establishes a material knowledge base, selects the parameter module most relevant to a given downstream task from the material knowledge base, and has high efficiency; trains the model based on the selected parameter module and performs fine-tuning to effectively adapt to the downstream material property prediction task. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a flow chart of the material property prediction method based on modular deep learning provided by the present invention.

[0019] Figure 2 It is a complete flow chart of the material property prediction method based on modular deep learning provided by the present invention.

[0020] Figure 3 It is a structural schematic diagram of a material property prediction device based on modular deep learning provided by the present invention.

[0021] Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] Before describing the method of the present invention in detail, the terms involved in the present invention are first schematically explained.

[0024] Modular Deep Learning: A deep learning approach that decomposes complex models into independent submodules.

[0025] Pre-training: Training a machine learning model on a large dataset to learn general representations that are independent of the specific task.

[0026] Fine-tuning: Based on the pre-trained model, improve the model performance by continuing training on a specific task dataset.

[0027] Label propagation: A semi-supervised learning algorithm that performs classification by propagating labels from labeled data to unlabeled data.

[0028] Genetic algorithm: An optimization algorithm that simulates the natural evolution process and iterates through selection, crossover and mutation operations.

[0029] Combine the following Figure 1-Figure 4 The present invention describes a material property prediction method and device based on modular deep learning.

[0030] Figure 1 is a flow chart of the material property prediction method based on modular deep learning provided by the present invention, such as Figure 1 As shown, it is applied to a cloud-based material knowledge sharing platform, the cloud-based material knowledge sharing platform is loaded with a material knowledge base, the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, the heterogeneous material task is used to realize the property prediction of the material; The method includes the following: Step 100: Receive a target attribute prediction problem of a target material uploaded by a user.

[0031] It should be noted that the embodiment of the present invention is implemented by a cloud-based material knowledge sharing platform, which is used to store, manage and index uploaded parameter modules, which can be lightweight parameter modules and provide search, screening and download functions so that users can efficiently find the required parameter modules.

[0032] Figure 2 This is a complete flow chart of the material property prediction method based on modular deep learning provided by the present invention. Figure 2As shown, the method provided by the embodiment of the present invention is divided into two parts: modular knowledge decoupling (Modular Knowledge Decoupling for Source Tasks) and adaptive module assembly (Adaptive Module Assembly for Target Tasks).

[0033] For upstream heterogeneous material tasks, this embodiment uses modular deep learning technology to achieve knowledge decoupling and storage. Specifically, a general pre-trained material representation skeleton (Backbone) is selected, and a certain number of parameter modules (knowledge module) are mounted for each upstream task. The heterogeneous knowledge of the upstream tasks is adaptively routed to different parameter modules through the routing network. It should be noted that a group of parameter modules corresponds to a heterogeneous material task, and each task corresponds to more than one parameter module. After the training converges, these modules containing task knowledge are stored in the material knowledge base.

[0034] Before step 100, the process of constructing the material knowledge base provided in this embodiment is first described, which specifically includes: Step S100: obtaining a pre-trained material representation skeleton; wherein the pre-trained material representation skeleton is used to mount a parameter module.

[0035] It should be noted that heterogeneous materials refer to materials with new structures and properties prepared by a specific method from two or more materials with different crystal structures, compositions or morphological characteristics. The pre-trained material representation skeleton is the basic part of the deep learning model and is designed to extract and express the basic characteristics of the material. In this embodiment, the pre-trained material representation skeleton is used to mount the corresponding parameter modules of multiple upstream tasks.

[0036] Step S200: for a plurality of preset heterogeneous material tasks, design a corresponding parameter module for each heterogeneous material task and mount it onto the pre-trained material representation skeleton; wherein the parameter module is used to store heterogeneous knowledge corresponding to the heterogeneous material task.

[0037] Step S200 specifically includes: Step S210: For each heterogeneous material task, generate a corresponding initial parameter module, and mount the initial parameter module on the pre-trained material representation skeleton.

[0038] Specifically, an initial parameter module is generated for each heterogeneous material task and mounted on the pre-trained material representation skeleton. These parameter modules are used to learn and store relevant knowledge of the respective tasks.

[0039] Step S220: Routing the heterogeneous knowledge of the heterogeneous material task to the corresponding initial parameter module through a routing network to obtain a parameter module corresponding to the heterogeneous material task.

[0040] Specifically, according to the specific requirements of the task, the heterogeneous knowledge of the task is adaptively routed to the corresponding initial parameter module for parameter learning and knowledge storage, thereby obtaining a customized parameter module for each task.

[0041] Step S300: store the parameter module corresponding to each heterogeneous material task to obtain the material knowledge base.

[0042] Step S300 specifically includes: Step S310: jointly train the parameter modules.

[0043] It should be noted that during training, the pre-trained material representation skeleton needs to be frozen, and only the parameter module is trained to optimize the performance of the entire model on multiple tasks.

[0044] Step S320: When the training meets the convergence condition, the trained parameter module is stored to obtain the material knowledge base.

[0045] Specifically, after the training reaches the convergence condition, the trained parameter module corresponding to each task and its related information are extracted from the model and stored to build and update the material knowledge base.

[0046] Step 200: query the material knowledge base based on the target property prediction problem, determine the corresponding target parameter module, and return the target parameter module to the user; wherein the target parameter module is used to enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0047] Specifically, see Figure 2 For the downstream module assembly, this embodiment designs an AMA (Adaptive Modular Assembly) algorithm that does not require training to select the most relevant parameter modules for each downstream task. The AMA algorithm first uses a prediction algorithm based on label propagation to estimate the prediction performance of each module on the downstream task. After that, the AMA algorithm uses modular weight optimization based on a genetic algorithm and adaptively selects the module combination that minimizes the generalization error. After the module assembly, these modules will be spliced ​​onto the pre-trained material representation skeleton and jointly fine-tuned. The following is a detailed explanation.

[0048] Step 200 searches the material knowledge base based on the target attribute prediction problem to determine the corresponding target parameter module, including: Step 210: Based on the target attribute prediction problem, an adaptive modular assembly (AMA) algorithm is used to query the material knowledge base to obtain a plurality of target lightweight parameter submodules; wherein the combination of the plurality of target lightweight parameter submodules is a module combination that minimizes the generalization error.

[0049] Step 210 specifically includes: Step 211: Calculate based on the prediction algorithm of label propagation to obtain the prediction performance of each parameter module on the target heterogeneous material task corresponding to the target attribute prediction problem.

[0050] Specifically, the label propagation algorithm is applied to simulate the performance of the modules on the target task according to the label information of the target attribute prediction problem, and the prediction performance score of each module is calculated.

[0051] Step 212: Based on the prediction performance of each parameter module, the plurality of target lightweight parameter submodules are obtained through modular weight optimization based on a target algorithm; wherein the target algorithm is an optimization algorithm that does not require training.

[0052] Specifically, the target algorithm is an optimization algorithm that does not require training. Taking the genetic algorithm as an example, a modular weight optimization model is constructed based on the genetic algorithm framework. The predicted performance score is used as the optimization target. Through iterative search and optimization, the combination of multiple target lightweight parameter sub-modules that minimizes the overall generalization error is determined.

[0053] Step 220: Assemble and splice the multiple target lightweight parameter sub-modules to obtain a target parameter module corresponding to the target attribute prediction problem.

[0054] Step 220 specifically includes: Step 221, assembling the multiple target lightweight parameter sub-modules, and splicing the assembled multiple target lightweight parameter sub-modules to the pre-trained material representation skeleton to generate a deep learning model of the target heterogeneous material task corresponding to the target attribute prediction problem.

[0055] Specifically, the assembled modules are combined and spliced ​​onto the pre-trained material representation skeleton to construct a deep learning model for the target attribute prediction problem. This model integrates the general feature extraction capabilities of the pre-trained material representation skeleton and the specific task knowledge of the target parameter module.

[0056] Step 222: jointly fine-tune the deep learning model to perform a target attribute prediction parameter module based on the jointly fine-tuned deep learning model.

[0057] Specifically, the constructed deep learning model is jointly fine-tuned, and the model parameters are further optimized by training on the target data set to improve the accuracy and generalization ability of the model in the target property prediction problem. Finally, the jointly fine-tuned deep learning model parameter module is obtained for subsequent material property prediction tasks.

[0058] The embodiment of the present invention uploads the parameter module to the cloud material knowledge sharing platform. When faced with a material property prediction problem to be solved, the cloud material knowledge sharing platform determines the most relevant knowledge module for users to download, thereby achieving effective knowledge transfer and accurate property prediction.

[0059] The above is a description of the steps of the material property prediction method based on modular deep learning provided by the present invention. From the description of the above steps, it can be seen that the material property prediction method based on modular deep learning provided by the present invention is applied to a cloud-based material knowledge sharing platform, and the cloud-based material knowledge sharing platform is loaded with a material knowledge base, and the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize the property prediction of the material; by receiving the target property prediction problem of the target material uploaded by the user; based on the target property prediction problem, query in the material knowledge base, determine the corresponding target parameter module, and return the target parameter module to the user; wherein, the target parameter module is used to: enable the user to train the model based on the target parameter module, and predict the properties of the target material based on the trained model. It can be seen that the present invention trains a series of parameter modules containing decoupled material knowledge for data-rich tasks, establishes a material knowledge base, selects the parameter module most relevant to a given downstream task from the material knowledge base, and has high efficiency; trains the model based on the selected parameter module and performs fine-tuning to effectively adapt to the downstream material property prediction task.

[0060] The material property prediction device based on modular deep learning provided by the present invention is described below. The material property prediction device based on modular deep learning described below and the material property prediction method based on modular deep learning described above can be referenced to each other.

[0061] Figure 3 is a schematic diagram of the structure of the material property prediction device based on modular deep learning provided by the present invention, such as Figure 3 As shown, the material property prediction device based on modular deep learning provided by the present invention is applied to a cloud-based material knowledge sharing platform, wherein the cloud-based material knowledge sharing platform is loaded with a material knowledge base, wherein the material knowledge base includes: multiple groups of parameter modules, wherein one group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize material property prediction; The device comprises: A receiving module 301 is used to receive a target attribute prediction problem of a target material uploaded by a user; The determination module 302 is used to query the material knowledge base based on the target property prediction problem, determine the corresponding target parameter module, and return the target parameter module to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0062] The material property prediction device based on modular deep learning provided by the present invention is applied to a cloud-based material knowledge sharing platform, and the cloud-based material knowledge sharing platform is loaded with a material knowledge base, and the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize the material property prediction; by receiving the target property prediction problem of the target material uploaded by the user; based on the target property prediction problem, querying in the material knowledge base, determining the corresponding target parameter module, and returning the target parameter module to the user; wherein the target parameter module is used to: enable the user to train the model based on the target parameter module, and predict the properties of the target material based on the trained model. It can be seen that the present invention trains a series of parameter modules containing decoupled material knowledge for data-rich tasks, establishes a material knowledge base, selects the parameter module most relevant to a given downstream task from the material knowledge base, and has high efficiency; trains the model based on the selected parameter module and performs fine-tuning to effectively adapt to the downstream material property prediction task.

[0063] Based on the above embodiment, in this embodiment, the device further includes a library generation module, which is specifically used to: Before receiving the target attribute prediction problem of the target material uploaded by the user, a pre-trained material representation skeleton is obtained; wherein the pre-trained material representation skeleton is used to mount the parameter module; For a plurality of preset heterogeneous material tasks, a corresponding parameter module is designed for each heterogeneous material task and mounted on the pre-trained material representation skeleton; wherein the parameter module is used to store heterogeneous knowledge corresponding to the heterogeneous material task; The parameter module corresponding to each heterogeneous material task is stored to obtain the material knowledge base.

[0064] Based on the above embodiment, in this embodiment, the library generation module is specifically used for: For each heterogeneous material task, generate a corresponding initial parameter module, and mount the initial parameter module onto the pre-trained material representation skeleton; The heterogeneous knowledge of the heterogeneous material task is routed to the corresponding initial parameter module through a routing network to obtain the parameter module corresponding to the heterogeneous material task.

[0065] Based on the above embodiment, in this embodiment, the library generation module is specifically used for: Jointly train parameter modules; When the training meets the convergence condition, the trained parameter module is stored to obtain the material knowledge base.

[0066] Based on the above embodiment, in this embodiment, the determining module 302 is specifically used for: Based on the target attribute prediction problem, a plurality of target lightweight parameter submodules are obtained by querying in the material knowledge base through an adaptive modular assembly AMA algorithm; wherein the combination of the plurality of target lightweight parameter submodules is: a module combination that minimizes the generalization error; The multiple target lightweight parameter sub-modules are assembled and spliced ​​to obtain a target parameter module corresponding to the target attribute prediction problem.

[0067] Based on the above embodiment, in this embodiment, the determining module 302 is specifically used for: The prediction algorithm based on label propagation is used to calculate and obtain the prediction performance of each parameter module on the target heterogeneous material task corresponding to the target attribute prediction problem; Based on the prediction performance of each parameter module, the multiple target lightweight parameter sub-modules are obtained through modular weight optimization based on a target algorithm; wherein the target algorithm is an optimization algorithm that does not require training.

[0068] Based on the above embodiment, in this embodiment, the determining module 302 is specifically used for: Assembling the multiple target lightweight parameter submodules, and splicing the assembled multiple target lightweight parameter submodules to the pre-trained material representation skeleton to generate a deep learning model of the target heterogeneous material task corresponding to the target attribute prediction problem; The deep learning model is jointly fine-tuned to perform a target attribute prediction parameter module based on the jointly fine-tuned deep learning model.

[0069] Figure 4 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 4As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430 and a communication bus 440, wherein the processor 410, the communication interface 420 and the memory 430 communicate with each other through the communication bus 440. The processor 410 may call the logic instructions in the memory 430 to execute the material property prediction method based on modular deep learning, which is applied to a cloud-based material knowledge sharing platform, wherein the cloud-based material knowledge sharing platform is loaded with a material knowledge base, wherein the material knowledge base includes: multiple groups of parameter modules, wherein a group of parameter modules corresponds to a heterogeneous material task, wherein the heterogeneous material task is used to realize the material property prediction; The method comprises: Receive target attribute prediction problem of target materials uploaded by users; Based on the target property prediction problem, the material knowledge base is queried, the corresponding target parameter module is determined, and the target parameter module is returned to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0070] In addition, the logic instructions in the above-mentioned memory 430 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0071] On the other hand, the present invention further provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the material property prediction method based on modular deep learning provided by the above methods, applied to a cloud-based material knowledge sharing platform, the cloud-based material knowledge sharing platform is loaded with a material knowledge base, the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, the heterogeneous material task is used to realize material property prediction; The method comprises: Receive target attribute prediction problem of target materials uploaded by users; Based on the target property prediction problem, the material knowledge base is queried, the corresponding target parameter module is determined, and the target parameter module is returned to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0072] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when executed by a processor to execute the material property prediction method based on modular deep learning provided by the above methods, and is applied to a cloud-based material knowledge sharing platform, wherein the cloud-based material knowledge sharing platform is loaded with a material knowledge base, and the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, and the heterogeneous material task is used to realize material property prediction; The method comprises: Receive target attribute prediction problem of target materials uploaded by users; Based on the target property prediction problem, the material knowledge base is queried, the corresponding target parameter module is determined, and the target parameter module is returned to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

[0073] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0074] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A material property prediction method based on modular deep learning, characterized in that: Applied to a cloud-based material knowledge sharing platform, the cloud-based material knowledge sharing platform is loaded with a material knowledge base, the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, the heterogeneous material task is used to realize the property prediction of the material; The method comprises: Receive target attribute prediction problem of target materials uploaded by users; Based on the target property prediction problem, the material knowledge base is queried, the corresponding target parameter module is determined, and the target parameter module is returned to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

2. The material property prediction method based on modular deep learning according to claim 1, characterized in that: Before receiving the target attribute prediction problem of the target material uploaded by the user, the method further includes: Obtaining a pre-trained material representation skeleton; wherein the pre-trained material representation skeleton is used to mount a parameter module; For a plurality of preset heterogeneous material tasks, a corresponding parameter module is designed for each heterogeneous material task and mounted on the pre-trained material representation skeleton; wherein the parameter module is used to store heterogeneous knowledge corresponding to the heterogeneous material task; The parameter module corresponding to each heterogeneous material task is stored to obtain the material knowledge base.

3. The material property prediction method based on modular deep learning according to claim 2, characterized in that: For a plurality of preset heterogeneous material tasks, a corresponding parameter module is designed for each heterogeneous material task and mounted on the pre-trained material representation skeleton, including: For each heterogeneous material task, generate a corresponding initial parameter module, and mount the initial parameter module onto the pre-trained material representation skeleton; The heterogeneous knowledge of the heterogeneous material task is routed to the corresponding initial parameter module through a routing network to obtain the parameter module corresponding to the heterogeneous material task.

4. The material property prediction method based on modular deep learning according to claim 2, characterized in that: The storing of the parameter module corresponding to each heterogeneous material task in the plurality of heterogeneous material tasks to obtain the material knowledge base includes: Jointly train parameter modules; When the training meets the convergence condition, the trained parameter module is stored to obtain the material knowledge base.

5. The material property prediction method based on modular deep learning according to claim 1, characterized in that: The step of querying the material knowledge base based on the target attribute prediction problem to determine a corresponding target parameter module includes: Based on the target attribute prediction problem, a plurality of target lightweight parameter submodules are obtained by querying in the material knowledge base through an adaptive modular assembly AMA algorithm; wherein the combination of the plurality of target lightweight parameter submodules is: a module combination that minimizes the generalization error; The multiple target lightweight parameter sub-modules are assembled and spliced ​​to obtain a target parameter module corresponding to the target attribute prediction problem.

6. The material property prediction method based on modular deep learning according to claim 5, characterized in that: Based on the target attribute prediction problem, the adaptive modular assembly AMA algorithm is used to query in the material knowledge base to obtain multiple target lightweight parameter sub-modules, including: The prediction algorithm based on label propagation is used to calculate and obtain the prediction performance of each parameter module on the target heterogeneous material task corresponding to the target attribute prediction problem; Based on the prediction performance of each parameter module, the multiple target lightweight parameter sub-modules are obtained through modular weight optimization based on a target algorithm; wherein the target algorithm is an optimization algorithm that does not require training.

7. The material property prediction method based on modular deep learning according to claim 5, characterized in that: The assembling and splicing of the multiple target lightweight parameter submodules to obtain a target parameter module corresponding to the target attribute prediction problem includes: Assembling the multiple target lightweight parameter submodules, and splicing the assembled multiple target lightweight parameter submodules to the pre-trained material representation skeleton to generate a deep learning model of the target heterogeneous material task corresponding to the target attribute prediction problem; The deep learning model is jointly fine-tuned to perform a target attribute prediction parameter module based on the jointly fine-tuned deep learning model.

8. A material property prediction device based on modular deep learning, characterized in that: Applied to a cloud-based material knowledge sharing platform, the cloud-based material knowledge sharing platform is loaded with a material knowledge base, the material knowledge base includes: multiple groups of parameter modules, one group of parameter modules corresponds to a heterogeneous material task, the heterogeneous material task is used to realize the property prediction of the material; The device comprises: A receiving module, used for receiving target attribute prediction problems of target materials uploaded by users; A determination module is used to query the material knowledge base based on the target property prediction problem, determine the corresponding target parameter module, and return the target parameter module to the user; wherein the target parameter module is used to: enable the user to train a model based on the target parameter module, and predict the properties of the target material based on the trained model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the material property prediction method based on modular deep learning as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the material property prediction method based on modular deep learning as described in any one of claims 1 to 7 is implemented.

Images