Scholarlink

<p>This work presents a reproducible workflow that uses an ensemble of Chemprop’s Directed Message-Passing Neural Networks (D-MPNNs) for polymer property prediction, specifically the glass transition temperature (<i>T</i><sub>g</sub>). The workflow estimates epistemic uncertainty and defines an applicability domain, enabling users to identify which predictions are most reliable, prioritize data for further study, and interpret the latent D-MPNN features driving model decisions. Using a dataset of 902 polymers with SMILES representations and experimental <i>T</i><sub>g</sub> values, this work demonstrates that the D-MPNN ensemble achieves strong predictive performance (test set <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(R^2 = 0.765\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.765</mn> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(RMSE = 0.0673\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>R</mi> <mi>M</mi> <mi>S</mi> <mi>E</mi> <mo>=</mo> <mn>0.0673</mn> </mrow> </math></EquationSource> </InlineEquation>), comparable to previous studies employing classical machine learning models with carefully selected molecular descriptors.</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

Polymer property prediction using ensemble of directed message-passing neural networks with uncertainty estimation and applicability domain analysis

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Vidit Agrawal

摘要

This work presents a reproducible workflow that uses an ensemble of Chemprop’s Directed Message-Passing Neural Networks (D-MPNNs) for polymer property prediction, specifically the glass transition temperature (Tg). The workflow estimates epistemic uncertainty and defines an applicability domain, enabling users to identify which predictions are most reliable, prioritize data for further study, and interpret the latent D-MPNN features driving model decisions. Using a dataset of 902 polymers with SMILES representations and experimental Tg values, this work demonstrates that the D-MPNN ensemble achieves strong predictive performance (test set \(R^2 = 0.765\) R 2 = 0.765 , \(RMSE = 0.0673\) R M S E = 0.0673 ), comparable to previous studies employing classical machine learning models with carefully selected molecular descriptors.

Graphical abstract