文章信息/Info

Title:

Data and Model Driven Intelligent Computing Framework for Perovskite Materials

文章編號(hào):

1674-3962(2025)04-0309-09

作者:

程晉榮; 何鵬飛; 李藝欣; 雷詠梅

1.上海大學(xué)材料科學(xué)與工程學(xué)院 ，上海 200444 2.上海大學(xué)計(jì)算機(jī)工程與科學(xué)學(xué)院，上海 200444

Author(s):

CHENG Jinrong;  HE Pengfei;  LI Yixing;  LEI Yongmei

1.School of Materials Science and Engineering, Shanghai University, Shanghai 200444, China 2.School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China

關(guān)鍵詞:

SISSO算法; 智能計(jì)算; 主動(dòng)學(xué)習(xí); 鈣鈦礦材料

Keywords:

SISSO algorithm;  intelligent computing;  active learning;  perovskite materials

分類號(hào):

TQ174.1; TP181

DOI:

10.7502/j.issn.1674-3962.202412002

文獻(xiàn)標(biāo)志碼:

A

摘要:

鈣鈦礦材料因其復(fù)雜的化學(xué)成分、多樣的晶體結(jié)構(gòu)和豐富的物理特性，成為現(xiàn)代材料科學(xué)研究熱點(diǎn)之一。結(jié)合模型驅(qū)動(dòng)方法和數(shù)據(jù)驅(qū)動(dòng)方法，構(gòu)建特征工程融合主動(dòng)學(xué)習(xí)的材料智能計(jì)算框架，提高模型精度和系統(tǒng)性能。通過數(shù)據(jù)布局和動(dòng)態(tài)調(diào)度協(xié)同優(yōu)化，提出針對(duì)材料特征的確定獨(dú)立篩選和稀疏算子（SISSO）并行計(jì)算方法，緩解SISSO算法在建立特征工程模型時(shí)面臨的精度較低與計(jì)算成本較高的問題，降低數(shù)據(jù)質(zhì)量對(duì)模型的影響。構(gòu)建面向材料數(shù)據(jù)的主動(dòng)學(xué)習(xí)方法，以處理材料數(shù)據(jù)標(biāo)記的復(fù)雜性，剔除噪聲數(shù)據(jù)。

Abstract:

Perovskite materials have become one of the hotspots in modern materials science research due to their complex chemical compositions, diverse crystal structures and rich physical properties. In this paper, by combining the model�瞕riven approach and the data-driven approach, a materials intelligent computing framework integrating feature engineering and active learning is constructed to improve the model accuracy and system performance. Through the collaborative optimization of data layout and dynamic scheduling, a sure independence screening and sparsifying operator (SISSO) parallel computing method for material features is proposed to alleviate the problems of low accuracy and high computational cost faced by the SISSO algorithm when establishing the feature engineering model and reduce the impact of data quality on the model. An active learning method oriented to material data is constructed to deal with the complexity of material data labeling and eliminate noisy data.