日本地球惑星科学連合2019年大会

講演情報

[J] ポスター発表

セッション記号 S (固体地球科学) » S-CG 固体地球科学複合領域・一般

[S-CG62] 固体地球科学における機械学習の可能性

2019年5月26日(日) 17:15 〜 18:30 ポスター会場 (幕張メッセ国際展示場 8ホール)

コンビーナ:内出 崇彦(産業技術総合研究所 地質調査総合センター 活断層・火山研究部門)、小田 啓邦(産業技術総合研究所地質情報研究部門)

[SCG62-P08] FORCsensei: 機械学習を用いたFirst Order Reversal Curve (FORC)データ処理の最適化

David Heslop1,2Andrew Roberts1,2、*小田 啓邦2Xiang Zhao1,2Richard Harrison3Adrian Muxworthy4Pengxiang Hu1,2佐藤 哲郎2 (1.オーストラリア国立大学、2.産業技術総合研究所地質情報研究部門、3.ケンブリッジ大学、4.インペリアルカレッジロンドン)

キーワード:機械学習、FORC、磁気ヒステリシス、2次微分

First-order reversal curve (FORC) distributions are a powerful diagnostic tool for characterizing and quantifying magnetization processes in fine magnetic particle systems. Estimation of FORC distributions requires computation of the second-order mixed derivative of noisy magnetic hysteresis data. This operation amplifies measurement noise and for weakly magnetic systems it can compromise estimation of a FORC distribution. Several processing schemes, typically based on local polynomial regression, have been developed to produce smoothed FORC distributions that suppress detrimental noise. As FORC analysis has become more quantitative, the smoothed distribution needs to be consistent with the measurement data from which it was estimated. This can be a challenging task even for expert users, who must subjectively adjust parameters that define the form and extent of smoothing until a satisfactory FORC distribution is obtained. For non-expert users, estimation of FORC distributions using inappropriate smoothing parameters can produce heavily distorted results corrupted by processing artifacts, which in turn can lead to spurious inferences concerning the magnetic system under investigation. We have developed a statistical machine learning framework based on probabilistic model comparison to guide estimation of FORC distributions. An intuitive approach is presented that reveals which regions of a FORC distribution may have been smoothed inappropriately. Our machine learning approach will objectively select an optimal FORC distribution probabilistically, which will automate the derivative estimation process for both expert and non-expert users.