Scholarlink

<p>Machine learning (ML) is increasingly transforming geosciences by reshaping how geological data are generated, interpreted, and integrated across fields from seismology to global Earth system modeling. This review synthesizes the methodological and epistemic changes, arguing that ML constitutes a methodological revolution rather than a paradigm shift, as it supplements rather than supplants the mechanistic foundations of geoscientific explanation. We identify three structural limitations of purely data-driven models. These are: (1) systemic extrapolation failure in non-stationary regimes; (2) limited physics-based explainability, which restricts mechanistic inference; and (3) concerns regarding reproducibility, bias, and sustainability. To address this, we formalize the Data–Model Coevolution Paradigm, a framework advocating for an iterative evolution of ML architectures and physical theory through a feedback loop between representation learning and process-based understanding. By integrating ML with physical reasoning via hybrid modeling, differentiable simulation, and operator learning, this coevolutionary approach offers a roadmap toward trustworthy, physically consistent, and scientifically generative AI in geoscience.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning advances and data model coevolution in geoscience

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月18日

查看全文
  • Abdelrhim Eltijnai,
  • Musaab A. A. Mohammed

摘要

Machine learning (ML) is increasingly transforming geosciences by reshaping how geological data are generated, interpreted, and integrated across fields from seismology to global Earth system modeling. This review synthesizes the methodological and epistemic changes, arguing that ML constitutes a methodological revolution rather than a paradigm shift, as it supplements rather than supplants the mechanistic foundations of geoscientific explanation. We identify three structural limitations of purely data-driven models. These are: (1) systemic extrapolation failure in non-stationary regimes; (2) limited physics-based explainability, which restricts mechanistic inference; and (3) concerns regarding reproducibility, bias, and sustainability. To address this, we formalize the Data–Model Coevolution Paradigm, a framework advocating for an iterative evolution of ML architectures and physical theory through a feedback loop between representation learning and process-based understanding. By integrating ML with physical reasoning via hybrid modeling, differentiable simulation, and operator learning, this coevolutionary approach offers a roadmap toward trustworthy, physically consistent, and scientifically generative AI in geoscience.

Graphical Abstract