材料领域知识嵌入的机器学习

Machine Learning Embedded with Materials Domain Knowledge

数据驱动的机器学习因其能够快速拟合历史数据中的潜在模式并实现材料性能的精准预测,已被广泛应用于材料性能优化和新材料设计。然而,由于缺乏描述符间关联关系、材料性能驱动机制等材料领域知识的指导,数据驱动的机器学习在实际应用中常常出现与材料基础理论认知或原理不一致的结果。本工作通过分析材料数据的特点和数据驱动的机器学习建模原理,厘清了数据驱动的机器学习应用于材料领域面临的三大矛盾:高维度与小样本数据的矛盾、模型准确性与易用性的矛盾、模型学习结果与领域专家知识的矛盾。藉此提出材料领域知识嵌入的机器学习作为上述矛盾的调和策略。进一步,面向“目标定义–数据准备–数据预处理–特征工程–模型构建–模型应用”的机器学习全流程,通过剖析相关的基础性和探索性工作,探讨了在机器学习各阶段实现材料领域知识嵌入的关键技术。最后,展望了材料领域知识嵌入机器学习的发展机遇和挑战。

Data-driven Machine Learning(ML) has been widely used in materials performance optimization and novel materials design due to its ability to quickly fit potential data patterns and achieve accurate prediction. However, the results of data-driven ML are often inconsistent with the materials basic theory or principle, which results mainly from the lack of the guidance of materials domain knowledge, e.g., the correlation among descriptors and the driving mechanism associated with the properties. Herein, by analyzing the characteristics of materials data and the modeling principle of data-driven ML methods, we clarify the three main contradictions occurring to the application of ML in materials science, i.e., the contradictions between high dimension and small sample, accuracy and usability of models, learning results and domain knowledge. Following this, we propose the ML method embedded with materials domain knowledge to reconcile these three contradictions. Further, surrounding the whole ML process including target definition, data collection and preprocessing, feature engineering, model construction and application, we explore some key techniques to realize domain knowledge embedding by summarizing the related basic and exploratory efforts. Finally, opportunities and challenges facing the ML method embedded with domain knowledge are also discussed.