Training Method for Generation Artificial Intelligence Model for Steel Heat Treatment Process

Abstract

The method for training a generative artificial intelligence model for a steel heat treatment process according to the present invention comprises the steps of: generating a first text sequence regarding a heat treatment process of a steel product using a large language model (LLM); collecting first microstructure numerical / image information and first physical property information of the steel product for each of the first text sequences; and, in training an artificial intelligence model by receiving as a training data set the first physical property information obtained through actual experiments, the first microstructure numerical / image information corresponding to the first physical property information, and the first text sequence regarding a heat treatment process corresponding to the first physical property information, wherein the large language model embedding model (LLM embedding model) receives the first text sequence and generates a first embedding vector for identifying semantic relationships; the artificial intelligence model receives the first physical property information and generates a second microstructure numerical / image information; the artificial intelligence model receives the second microstructure numerical / image information and generates a second embedding vector; and the large language model decoding model (LLM decoding A method for training a generative artificial intelligence model for a steel heat treatment process is provided, comprising the steps of: receiving a second embedding vector generated by the artificial intelligence model and generating a second text sequence; and training the artificial intelligence model based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, and the difference between the first microstructure numerical / image information and the second microstructure numerical / image information.

Description

Technology Field

[0001] The present invention relates to steel heat treatment process technology, and more specifically, to a method for training a generative artificial intelligence model for a steel heat treatment process. Background Technology

[0003] Since the physical properties of steel products vary depending on the heat treatment, the development of heat treatment technology is important to maintain product quality.

[0005] Figure 1 shows the change in specific physical properties obtained by the heat treatment process.

[0006] That is, referring to Fig. 1, the physical properties that change due to the heat treatment process include hardness, strength, ductility, and toughness.

[0007] Traditionally, finding the appropriate process to obtain the desired physical properties required numerous trials and errors.

[0008] Therefore, technology is required to automate material property information such as hardness, strength, ductility, and toughness resulting from these heat treatment processes. Prior art literature

[0010] KR10-2024-0065047 A The problem to be solved

[0011] The present invention is proposed to solve the technical problem described above and provides a structured training method for an artificial intelligence model that expresses a heat treatment process for obtaining desired physical properties of steel products as a word sequence and generates it, that is, a training method for a generative artificial intelligence model for a steel heat treatment process. means of solving the problem

[0013] According to one embodiment of the present invention for solving the above problem, the method comprises the steps of: generating a first text sequence regarding a heat treatment process of a steel product using a large language model (LLM); collecting first microstructure numerical / image information and first physical property information of the steel product for each of the first text sequences; and training an artificial intelligence model by receiving the first physical property information obtained through actual experiments, the first microstructure numerical / image information corresponding to the first physical property information, and the first text sequence regarding a heat treatment process corresponding to the first physical property information as a training data set, wherein the large language model embedding model (LLM embedding model) receives the first text sequence and generates a first embedding vector for identifying semantic relationships; the artificial intelligence model receives the first physical property information and generates a second microstructure numerical / image information; the artificial intelligence model receives the second microstructure numerical / image information and generates a second embedding vector; and the large language model decoding model (LLM decoding model) A method for training a generative artificial intelligence model for a steel heat treatment process is provided, comprising the steps of: receiving a second embedding vector generated by a model and generating a second text sequence; and training the artificial intelligence model based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, and the difference between the first microstructure numerical / image information and the second microstructure numerical / image information.

[0015] Additionally, according to another embodiment of the present invention, the method comprises the steps of: generating a first text sequence regarding a heat treatment process of a steel product using a large language model (LLM); collecting first microstructure numerical / image information and first physical property information of the steel product for each of the first text sequences; and training an artificial intelligence model by receiving the first physical property information obtained through actual experiments, the first microstructure numerical / image information corresponding to the first physical property information, and the first text sequence regarding a heat treatment process corresponding to the first physical property information as a training data set, wherein the large language model embedding model (LLM embedding model) receives the first text sequence and generates a first embedding vector for identifying semantic relationships; the artificial intelligence model receives the first embedding vector and generates second microstructure numerical / image information; the artificial intelligence model receives the second microstructure numerical / image information and generates second physical property information; and the artificial intelligence model receives the second physical property information and generates third microstructure numerical / image information. A method for training a generative artificial intelligence model for a steel heat treatment process is provided, comprising: a step in which the artificial intelligence model receives the third microstructure numerical / image information and generates a second embedding vector; a step in which the large language model decoding model (LLM decoding model) receives the second embedding vector generated by the artificial intelligence model and generates a second text sequence; and a step in which the artificial intelligence model trains based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, the difference between the first microstructure numerical / image information, the second microstructure numerical / image information, and the third microstructure numerical / image information, and the difference between the first physical property information and the second physical property information.

[0016] In addition, the physical property information in the present invention is characterized by including hardness, strength, ductility, and toughness. Effects of the invention

[0018] The training method for a generative artificial intelligence model for a steel heat treatment process according to the present invention provides a structured training method for an artificial intelligence model that expresses a heat treatment process for obtaining desired physical properties of a steel product as a sequence of words and generates it. Brief explanation of the drawing

[0020] Figure 1 is a drawing showing the change in specific physical properties obtained by a heat treatment process. Figure 2 is a diagram showing a method for analyzing the microstructure of a material. FIG. 3 is a diagram showing training data used in the training method for a prototype generation artificial intelligence model for a steel heat treatment process according to the present invention. Figure 4 is a drawing showing a heat treatment process prompt embedding. FIG. 5 is a diagram showing a text (token) sequence, i.e., a prompt, for the heat treatment process. FIG. 6 is a configuration diagram showing a training method according to a first embodiment of an artificial intelligence model, namely a unidirectional training method. FIG. 7 is a drawing showing an example of a heat treatment process generation process. FIG. 8 is a configuration diagram showing a training method according to a second embodiment of an artificial intelligence model, namely a prototype training method. Specific details for implementing the invention

[0021] Hereinafter, in order to explain in detail enough for a person skilled in the art to easily implement the technical concept of the present invention, embodiments of the present invention will be described with reference to the attached drawings.

[0023] Figure 2 is a diagram showing a method for analyzing the microstructure of a material.

[0024] Referring to Fig. 2, the heat treatment process can be represented as a sequence of text (tokens), and the microstructure analysis of the material can be performed by shooting an electron beam to extract numerical values / images related to the physical properties of the object.

[0025] In addition, material properties such as hardness, strength, ductility, and toughness can be expressed numerically.

[0027] Figure 3 is a diagram showing training data used in the training method for a generative artificial intelligence model for a steel heat treatment process of the present invention.

[0028] Referring to Fig. 3, as training data, A) a text (token) sequence for each heat treatment process, B) microstructure numerical values / images, i.e., numerical values ​​or images related to material properties, and C) numerical information regarding material properties are used.

[0030] The training method for a generative artificial intelligence model for a steel heat treatment process according to an embodiment of the present invention is as follows.

[0031] First, a step of generating a first text sequence for the heat treatment process of a steel product using a large language model (LLM) is processed.

[0032] In addition, a step of collecting first microstructure numerical / image information and first physical property information of a steel product for each of the first text sequences is processed.

[0034] In addition, the artificial intelligence model is trained by receiving first material property information obtained through actual experiments, first microstructure numerical / image information corresponding to the first material property information, and a first text sequence regarding the heat treatment process corresponding to the first material property information as a training data set.

[0036] The training method for the artificial intelligence model is as follows.

[0037] Figure 4 is a diagram showing a heat treatment process prompt embedding, and Figure 5 is a diagram showing a text (token) sequence, i.e., a prompt, for the heat treatment process.

[0038] Referring to FIGS. 4 and 5, a massive language model embedding model (LLM embedding model) receives a first text sequence regarding a heat treatment process as input and generates a first embedding vector to identify semantic relationships.

[0039] For reference, referring to FIG. 5, a first text sequence for a heat treatment process of a steel product can be generated using a large language model (LLM).

[0041] FIG. 6 is a configuration diagram showing a training method according to a first embodiment of an artificial intelligence model, namely a unidirectional training method.

[0042] Referring to FIG. 6, the artificial intelligence model training method according to the first embodiment is processed as follows.

[0043] First, a step is processed in which a massive language model embedding model (LLM embedding model) receives a first text sequence as input and generates a first embedding vector to identify semantic relationships.

[0044] In addition, a step is processed in which an artificial intelligence model receives first material property information and generates second microstructure numerical / image information.

[0045] In addition, a step is processed in which the artificial intelligence model receives second microstructure numerical / image information and generates a second embedding vector.

[0046] In addition, a step is processed in which a massive language model decoding model (LLM decoding model) receives the second embedding vector generated by the artificial intelligence model and generates a second text sequence.

[0047] That is, a step is processed in which an artificial intelligence model is trained based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, and the difference between the first microstructure numerical / image information and the second microstructure numerical / image information.

[0049] Figure 7 is a diagram showing an example of a heat treatment process generation process.

[0050] Referring to FIG. 7, when a desired physical property value (physical property information) is input into an artificial intelligence model trained through the prototype generation artificial intelligence model training method for the steel heat treatment process of the present invention, automation is achieved in which a text sequence capable of achieving the said physical property information is provided.

[0052] FIG. 8 is a configuration diagram showing a training method according to a second embodiment of an artificial intelligence model, namely a prototype training method.

[0053] Referring to Fig. 8, first, a step is processed in which a large language model embedding model (LLM embedding model) receives a first text sequence as input and generates a first embedding vector to identify semantic relationships.

[0054] In addition, a step is processed in which an artificial intelligence model receives a first embedding vector and generates second microstructure numerical / image information.

[0055] In addition, a step is processed in which an artificial intelligence model receives second microstructure numerical / image information and generates second material property information.

[0056] In addition, a step is processed in which an artificial intelligence model receives second material property information and generates third microstructure numerical / image information.

[0057] In addition, a step is processed in which the artificial intelligence model receives third microstructure numerical / image information and generates a second embedding vector.

[0058] In addition, a step is processed in which a massive language model decoding model (LLM decoding model) receives the second embedding vector generated by the artificial intelligence model and generates a second text sequence.

[0059] That is, a training step is processed in which the artificial intelligence model is trained based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, the difference between the first microstructure numerical / image information, the second microstructure numerical / image information, and the third microstructure numerical / image information, and the difference between the first material property information and the second material property information.

[0060] In other words, training can proceed in a circular, alternating manner from both sides. That is, training can be performed with X as input and Y as the target, or with Y as input and X as the target.

[0062] As described above, the training method for a generative artificial intelligence model for a steel heat treatment process according to the present invention provides a structured training method for an artificial intelligence model that expresses a heat treatment process for obtaining desired physical properties of a steel product as a sequence of words and generates it.

[0064] As such, those skilled in the art to which the present invention pertains will understand that the present invention may be implemented in other specific forms without altering its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of the present invention is defined by the claims set forth below rather than by the detailed description above, and all modifications or variations derived from the meaning and scope of the claims and their equivalents should be interpreted as being included within the scope of the present invention.

Claims

Claim 1 A step of generating a first text sequence regarding a heat treatment process of a steel product using a large language model (LLM); a step of collecting first microstructure numerical / image information and first physical property information of the steel product for each of the first text sequences; and a step of training an artificial intelligence model by receiving as a training data set the first physical property information obtained through actual experiments, the first microstructure numerical / image information corresponding to the first physical property information, and the first text sequence regarding a heat treatment process corresponding to the first physical property information, wherein a large language model embedding model (LLM embedding model) receives the first text sequence and generates a first embedding vector for identifying semantic relationships; a step in which the artificial intelligence model receives the first physical property information and generates second microstructure numerical / image information; a step in which the artificial intelligence model receives the second microstructure numerical / image information and generates a second embedding vector; and a large language model decoding model (LLM decoding model) receives the second embedding vector generated by the artificial intelligence model A method for training a generative artificial intelligence model for a steel heat treatment process, comprising: a step of generating a second text sequence; and a step of training the artificial intelligence model based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, and the difference between the first microstructure numerical / image information and the second microstructure numerical / image information. Claim 2 A method for training a generative artificial intelligence model for a steel heat treatment process, characterized in that, in claim 1, the physical property information includes hardness, strength, ductility, and toughness. Claim 3 A step of generating a first text sequence regarding a heat treatment process of a steel product using a large language model (LLM); a step of collecting first microstructure numerical / image information and first physical property information of the steel product for each of the first text sequences; and a step of training an artificial intelligence model by receiving as a training data set the first physical property information obtained through actual experiments, the first microstructure numerical / image information corresponding to the first physical property information, and the first text sequence regarding a heat treatment process corresponding to the first physical property information, wherein a large language model embedding model (LLM embedding model) receives the first text sequence and generates a first embedding vector to identify semantic relationships; a step in which the artificial intelligence model receives the first embedding vector and generates a second microstructure numerical / image information; a step in which the artificial intelligence model receives the second microstructure numerical / image information and generates a second physical property information; a step in which the artificial intelligence model receives the second physical property information and generates a third microstructure numerical / image information; and a step in which the artificial intelligence model [receives] the third A method for training a generative artificial intelligence model for a steel heat treatment process, comprising: a step of receiving microstructure numerical / image information and generating a second embedding vector; a step in which the big language model decoding model (LLM decoding model) receives the second embedding vector generated by the artificial intelligence model and generates a second text sequence; and a step in which the artificial intelligence model trains based on the difference between the first text sequence and the second text sequence, the difference between the first embedding vector and the second embedding vector, the difference between the first microstructure numerical / image information, the second microstructure numerical / image information, and the third microstructure numerical / image information, and the difference between the first physical property information and the second physical property information. Claim 4 A method for training a generative artificial intelligence model for a steel heat treatment process, characterized in that, in paragraph 3, the physical property information includes hardness, strength, ductility, and toughness.

Images