Scholarlink

<p>This study presents a novel multi-feature deep learning framework that integrates Transformer architecture with readily computable molecular descriptors to predict CO adsorption mechanisms at single metal oxide interfaces. The framework employs specialized encoders for structural, electronic, and kinetic descriptors, utilizing cross-feature attention mechanisms to capture the multifaceted nature of catalytic processes. Unlike traditional approaches requiring expensive DFT calculations, our method uses empirical descriptors and pre-computed parameters that can be rapidly obtained for practical catalyst screening applications. Comprehensive evaluation across seven distinct metal oxide systems demonstrates superior performance over traditional machine learning methods, achieving mean absolute errors below 0.12&#xa0;eV for adsorption energy prediction and correlation coefficients exceeding 0.92. Systematic ablation studies reveal the hierarchical importance of different data modalities, with structural information providing the most critical contribution. Case studies on CeO₂, TiO₂, and ZnO validate the model’s capability to distinguish material-specific mechanisms and provide mechanistic insights consistent with experimental observations. The multi-feature approach successfully predicts coverage-dependent effects, surface termination influences, and defect-mediated processes, establishing a foundation for data-driven catalyst design and mechanism elucidation in sustainable catalysis applications.</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

Multi-feature deep learning framework for predicting CO adsorption mechanisms at metal oxide interfaces: a transformer-based approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月31日

查看全文
  • Yu Yan

摘要

This study presents a novel multi-feature deep learning framework that integrates Transformer architecture with readily computable molecular descriptors to predict CO adsorption mechanisms at single metal oxide interfaces. The framework employs specialized encoders for structural, electronic, and kinetic descriptors, utilizing cross-feature attention mechanisms to capture the multifaceted nature of catalytic processes. Unlike traditional approaches requiring expensive DFT calculations, our method uses empirical descriptors and pre-computed parameters that can be rapidly obtained for practical catalyst screening applications. Comprehensive evaluation across seven distinct metal oxide systems demonstrates superior performance over traditional machine learning methods, achieving mean absolute errors below 0.12 eV for adsorption energy prediction and correlation coefficients exceeding 0.92. Systematic ablation studies reveal the hierarchical importance of different data modalities, with structural information providing the most critical contribution. Case studies on CeO₂, TiO₂, and ZnO validate the model’s capability to distinguish material-specific mechanisms and provide mechanistic insights consistent with experimental observations. The multi-feature approach successfully predicts coverage-dependent effects, surface termination influences, and defect-mediated processes, establishing a foundation for data-driven catalyst design and mechanism elucidation in sustainable catalysis applications.