Method of learning electrode model

Abstract

To provide a method of learning an electrode model allowing for suppressing the occurrence of excessive learning.SOLUTION: A method of learning an electrode model of a 3D structure from a 2D sectional image of an electrode includes the steps of: lowering the resolution of training data created from the 2D sectional image of the electrode acquired; random number-generating a 3D structure same in resolution as the training data lowered in resolution by a generator; and discriminating a truth / falseness of the generated 3D-structure by a discriminator; and learning a discrimination result by the generator and the discriminator.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a method for training an electrode model. [Background technology]

[0002] Various techniques have been proposed for electrodes such as those disclosed in Patent Document 1. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2023-098419 Summary of the Invention [Problem to be solved by the invention]

[0004] Conventional techniques for generating 3D (three-dimensional) electrode models from 2D (two-dimensional) cross-sectional images are only capable of generating simple 3D structures. It is difficult to generate complex, high-resolution 3D structures from 2D cross-sectional images that include nanoscale to microscale electrode structural features. In conventional technology, when a generative adversarial network (GAN) is used to generate and train a 3D structure from a 2D cross-sectional image of an electrode, the high resolution of the electrode image can lead to overfitting. The higher the resolution of the training image and the more complex its features, the easier it becomes for the classifier to distinguish between the generated image and the training image, and the learning speed of the classifier exceeds that of the generator, resulting in overfitting.

[0005] The present disclosure has been made in consideration of the above-described circumstances, and has as its main object to provide a method for learning an electrode model that can suppress the occurrence of overlearning. [Means for solving the problem]

[0006] That is, the present disclosure includes the following aspects. <1> 1. A method for learning a 3D structural electrode model from 2D cross-sectional images of an electrode, comprising: reducing the resolution of training data created from the acquired 2D cross-sectional images of the electrodes; A step of randomly generating a 3D structure having the same resolution as the training data reduced in resolution by a generator; A method for learning an electrode model, comprising the steps of: a classifier discriminating the authenticity of the generated 3D structure; and the generator and the classifier learning the discrimination result. [Effects of the Invention]

[0007] The electrode model learning method of the present disclosure can prevent overlearning from occurring. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a flowchart showing an example of a method for learning an electrode model according to the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments of the present disclosure will be described. Note that matters other than those specifically mentioned in this specification that are necessary for implementing the present disclosure (for example, the general process of the electrode model learning method that does not characterize the present disclosure) can be understood as design matters for those skilled in the art based on prior art in the relevant field. The present disclosure can be implemented based on the contents disclosed in this specification and common general technical knowledge in the relevant field.

[0010] The present disclosure provides a method for learning a 3D structural electrode model from 2D cross-sectional images of an electrode, the method comprising: reducing the resolution of training data created from the acquired 2D cross-sectional images of the electrodes; A step of randomly generating a 3D structure having the same resolution as the training data reduced in resolution by a generator; The present invention provides a method for learning an electrode model, the method including a step of using a classifier to determine whether the generated 3D structure is true or false, and the generator and the classifier learning the determination result.

[0011] In developing solid-state batteries, it is necessary to obtain 2D cross-sectional images of electrodes using a scanning electron microscope (SEM) and observe the electrode state. To compare and analyze batteries, the composition components of each image are colored and then analyzed for characteristics directly related to battery performance, such as the degree of bending. However, in the example of evaluating the degree of bending, evaluation on a 2D plane and evaluation on a 3D structure lead to different results. Furthermore, highly dispersed electrodes often cannot be evaluated on a 2D plane because they are not completely conductive. Therefore, it is necessary to obtain the 3D structure of the electrode. Other simulations, such as 3D charge / discharge calculations and stress analysis, also require the 3D structure of the electrode.

[0012] Conventional generation of 3D structures from 2D cross-sectional images does not involve step-by-step learning, and generation using step-by-step learning is only used for learning and generating 2D cross-sectional images. This disclosure incorporates step-by-step learning to generate 3D structures from 2D cross-sectional images of electrodes in an adversarial learning generative network. In stepwise learning, the resolution of the training images is reduced to allow the model to learn coarse features first, and then the resolution of the training data is gradually increased to allow the model to learn finer features. According to the present disclosure, by performing stepwise learning, the problem of overfitting of the classifier can be solved, and the characteristics of nanoscale and microscale electrodes can be accurately learned and generated.

[0013] The electrode model learning method of the present disclosure is a method for learning an electrode model with a 3D structure from a 2D cross-sectional image of the electrode. The electrode model learning method of the present disclosure includes a resolution reduction step, a random number generation step, and a learning step.

[0014] [Resolution reduction step] The resolution reduction step is a step of reducing the resolution of training data created from the acquired 2D cross-sectional images of the electrodes. The 2D cross-sectional image may be acquired by SEM, and the resolution of the 2D cross-sectional image may be 1200x1000 pixels or more. The 2D cross-sectional image may be color-coded to indicate the compositional components. The training data (training images) may be created by dividing a 2D cross-sectional image with a resolution of 1200x1000 pixels into, for example, 64x64 pixels. The resolution of the reduced-resolution training data may be 4x4 pixels or more, or 64x64 pixels or more. In the training image, the particle size distribution (long side) may be at least 4 pixels or more, or may be at least 5 pixels, and may be at most 60 pixels or less. In the training images, materials (voids) with an area of ​​5 pixels^2 or less may be scattered among materials with an area of ​​100 pixels^2 or more. In the training images, 80% or more of the material particles with an area of ​​100^2 pixels or more may have a circularity (in any shape range) of 0.7 or less. The training images may have three or more types of substances present, four or more types of substances present, or five or more types of substances present.

[0015] [Random number generation step] The random number generation step is a step in which a generator randomly generates a 3D structure with the same resolution as the training data whose resolution has been reduced. For example, the generator may randomly generate a 3D structure with a resolution of 4x4x4 pixels from training data with a reduced resolution of 4x4 pixels.

[0016] [Learning Steps] The learning step is a step in which a classifier determines whether the generated 3D structure is true or false, and the generator and the classifier learn the determination result. The randomly generated 3D structure may be divided to obtain cross sections in the x, y, and z directions. For example, a generated 3D structure with a resolution of 4x4x4 pixels may be divided to obtain cross sections in the x, y, and z directions with a resolution of 4x4 pixels. A classifier may determine the authenticity of each cross section of the generated 3D structure. The discrimination results may be fed back to the generator and the classifier, allowing the generator and the classifier to learn.

[0017] [Decision step] After the learning step, it may be determined whether the number of learning times has reached a specified number. If the number of learning times has not reached the specified number, the random number generation step may be performed again, and if the number of learning times has reached the specified number, the resolution of the training data may be increased. It may be determined whether the resolution of the enhanced training data is higher than the original resolution. If the resolution of the enhanced training data is lower than the original resolution, the random number generation step may be performed again, and if the resolution of the enhanced training data is higher than the original resolution, the 3D structure learning generation may be completed.

[0018] FIG. 1 is a flowchart showing an example of a method for learning an electrode model according to the present disclosure. A 2D cross-sectional image of the electrode is acquired using an SEM, and the composition components are color-coded. The resolution of the 2D cross-sectional image is, for example, 1200 x 1000 pixels. Divide the 2D cross-sectional image to create training data. For example, create training data with a resolution of 64x64 pixels from a 2D cross-sectional image with a resolution of 1200x1000 pixels. The training data is reduced in resolution. For example, the resolution is reduced from 64x64 pixels to 4x4 pixels (resolution reduction step). The generator randomly generates a 3D structure with the same resolution as the reduced-resolution training data (random number generation step). For example, a 3D structure with a resolution of 4x4x4 pixels is randomly generated from the reduced-resolution training data with a resolution of 4x4 pixels. Divide the generated 3D structure and obtain cross sections in the x, y, and z directions. For example, divide the generated 3D structure with a resolution of 4x4x4 pixels and obtain cross sections in the x, y, and z directions with a resolution of 4x4 pixels. A classifier determines whether each cross section of the generated 3D structure is true or false. The discrimination results are fed back to the generator and the classifier, and the generator and the classifier are trained (learning step). It is determined whether the number of learning attempts has reached the specified number. If the number of learning attempts has not reached the specified number, the random number generation step is performed again. If the number of learning attempts has reached the specified number, the resolution of the training data is increased. For example, the resolution may be increased from 4x4 pixels to 8x8 pixels or higher, 32x32 pixels or higher, or 64x64 pixels or higher. Determine whether the resolution of the increased-resolution training data is higher than the original resolution. If the increased-resolution training data is lower than the original resolution, perform the random number generation step again. If the increased-resolution training data is higher than the original resolution, complete the 3D structure learning generation. For example, if the resolution of the original image is 64x64 pixels and the resolution of the training data with increased resolution is 128x128 pixels, the training data with increased resolution is determined to have a higher resolution than the original.

[0019] An example of a method for manufacturing an actual electrode is shown below. The solvent, conductive additive, and active material, Si, are placed in a container and subjected to alternating 30-second cycles using an ultrasonic homogenizer and shaker. The binder is added and shaken for 10 minutes, after which the solid electrolyte is added. After addition, the mixture is subjected to alternating 30-second cycles using an ultrasonic homogenizer and shaker. The mixed slurry is applied to a current collector foil, dried at 100°C for 30 minutes, and densified using a 5 ton / cm hot roll press to form an electrode on the current collector foil.

[0020] The electrode of the present disclosure is typically used as at least one of the positive and negative electrodes of a battery. The battery comprises a positive electrode, an electrolyte layer, and a negative electrode. The type of battery is not particularly limited, but examples include lithium ion batteries. The battery may be a primary battery or a secondary battery. The battery may be a liquid battery using an electrolytic solution as an electrolyte, or may be a solid battery. In the present disclosure, a solid-state battery refers to a battery containing a solid electrolyte. The solid-state battery may be a semi-solid-state battery that contains a solid electrolyte and a liquid-based material, or an all-solid-state battery that does not contain a liquid-based material. Examples of uses of the battery include power sources for vehicles such as hybrid electric vehicles (HEVs), plug-in hybrid electric vehicles (PHEVs), electric vehicles (BEVs), gasoline-powered vehicles, and diesel-powered vehicles. In particular, the battery may be used as a driving power source for hybrid electric vehicles (HEVs), plug-in hybrid electric vehicles (PHEVs), or electric vehicles (BEVs). The battery may also be used as a power source for mobile objects other than vehicles (for example, trains, ships, and aircraft), and as a power source for electrical appliances such as information processing devices.

Claims

[Claim 1] 1. A method for learning a 3D structural electrode model from 2D cross-sectional images of an electrode, comprising: reducing resolution of training data created from the acquired 2D cross-sectional images of the electrodes; A generator randomly generates a 3D structure having the same resolution as the reduced-resolution training data; A method for learning an electrode model, comprising the steps of: a classifier discriminating the authenticity of the generated 3D structure; and the generator and the classifier learning the discrimination result.

Images