Presentation Information
[P2-10]Effective fine-tuning of image generative AI on FePt microstructures using the phase-field method
*Toshiyuki Koyama1,2 (1. Nagoya University (Japan), 2. National Institute for Materials Science (Japan))
Keywords:
Data science,Phase-field simulation,Image generative AI,FePt,Machine learning
The phase-field (PF) method is a powerful framework for calculating the formation of internal microstructures in various materials based on a continuum model. It has developed over the past four decades to become one of the main simulation methods for phase transformations and microstructure changes. During this period, the establishment of computational thermodynamics (CALPHAD method) and the material property calculations using material microstructure information have also advanced in various fields, simultaneously. Consequently, this series of computational methods is evolving into a robust system supporting overall material design. On the other hand, the machine learning techniques in the rapidly developing field of data-driven science have been further extended in recent years with the emergence of generative AI (Artificial Intelligence). In this study, we propose the efficient fine-tuning method on the microstructure generative AI by using the PF method with the denoising diffusion probabilistic models.
Currently, the image generative AI technology is advancing rapidly, as well as the language generative AI, and the PF method is a good complement to image generative AI. This complementarity is due to the fact that applying image generative AI to the materials microstructures requires large training data sets and is difficult to cover all of them, experimentally. Since the PF method can provide various microstructure developments under specified conditions, for instance with a constant volume fraction, it is beneficial to pre-train image generative AI with the output microstructure data from the PF simulation, and then fine-tune the pre-trained AI with additional experimental microstructures. Figure 1 is an example of such an approach, and shows the results after training of FePt microstructures of high-density magnetic recording media. (A) is a base model of image generative AI, and here, Texture Diffusion released on CIVITAI [1] is adopted. (B) is a FePt microstructure produced by the PF simulation, where the white part is the FePt phase, and the black part is a segregant phase such as an amorphous carbon, and 100 different microstructure morphologies are prepared as a training data set. (C) is the output of image generative AI that was fine-tuned on (A) with the information from (B) using Dream Booth technique, showing six examples. Note that the morphology of microstructure simulated by the PF method is reproduced, accurately. (D) is a micrograph of an actual FePt microstructure [2], and this single image was used for further additional fine-tuning on the image generative AI of (C), resulting in the final output (E) with five examples displayed, where LoRA (Low-Rank Adaptation) technique was employed for the final tuning. It is clear that realistic microstructure morphology can be generated with extreme accuracy. This method of preliminary training with morphology information from the PF method and additional learning with few experimental microstructures seems to be an effective approach from the standpoint of making efficient use of small number of experimental data.
[1] CIVITAI (https://civitai.com/models)
[2] A. Bolyachkin, H. Sepehri-Amin, I. Suzuki, H. Tajiri, Y.K. Takahashi, K. Srinivasan, H. Ho, H. Yuan, T. Seki, A. Ajan, and K. Hono, Acta Mater. 227, 117744 (2022).
Acknowledgements: Part of this research was supported by the JST SIP ("Materials Integration" for Revolutionary Design System of Structural Materials), MEXT Data Creation and Utilization Materials Research and Development Project (Data Creation and Utilization Magnetic Materials Research Center JPMXP1122715503), and JST-CREST (JPMJC22C3).
Currently, the image generative AI technology is advancing rapidly, as well as the language generative AI, and the PF method is a good complement to image generative AI. This complementarity is due to the fact that applying image generative AI to the materials microstructures requires large training data sets and is difficult to cover all of them, experimentally. Since the PF method can provide various microstructure developments under specified conditions, for instance with a constant volume fraction, it is beneficial to pre-train image generative AI with the output microstructure data from the PF simulation, and then fine-tune the pre-trained AI with additional experimental microstructures. Figure 1 is an example of such an approach, and shows the results after training of FePt microstructures of high-density magnetic recording media. (A) is a base model of image generative AI, and here, Texture Diffusion released on CIVITAI [1] is adopted. (B) is a FePt microstructure produced by the PF simulation, where the white part is the FePt phase, and the black part is a segregant phase such as an amorphous carbon, and 100 different microstructure morphologies are prepared as a training data set. (C) is the output of image generative AI that was fine-tuned on (A) with the information from (B) using Dream Booth technique, showing six examples. Note that the morphology of microstructure simulated by the PF method is reproduced, accurately. (D) is a micrograph of an actual FePt microstructure [2], and this single image was used for further additional fine-tuning on the image generative AI of (C), resulting in the final output (E) with five examples displayed, where LoRA (Low-Rank Adaptation) technique was employed for the final tuning. It is clear that realistic microstructure morphology can be generated with extreme accuracy. This method of preliminary training with morphology information from the PF method and additional learning with few experimental microstructures seems to be an effective approach from the standpoint of making efficient use of small number of experimental data.
[1] CIVITAI (https://civitai.com/models)
[2] A. Bolyachkin, H. Sepehri-Amin, I. Suzuki, H. Tajiri, Y.K. Takahashi, K. Srinivasan, H. Ho, H. Yuan, T. Seki, A. Ajan, and K. Hono, Acta Mater. 227, 117744 (2022).
Acknowledgements: Part of this research was supported by the JST SIP ("Materials Integration" for Revolutionary Design System of Structural Materials), MEXT Data Creation and Utilization Materials Research and Development Project (Data Creation and Utilization Magnetic Materials Research Center JPMXP1122715503), and JST-CREST (JPMJC22C3).
