<p>This study introduces the Trustworthy Microstructural Recognition Framework (TMRF), aimed at enhancing microstructural recognition with a focus on explainability, human-computer interaction, and accountability. Evaluating microstructural images across seven classes, the framework employed transfer learning models in two experiments: Group 1 (three distinct classes) and Group 2 (all classes). DenseNet outperformed other models, achieving 97<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="43578_2025_1548_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> accuracy and 96<InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="43578_2025_1548_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> F1 scores. Explainable AI (XAI) improved interpretability, with Occlusion and SmoothGrad providing insertion fidelity scores of 60<InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="43578_2025_1548_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> and 50<InlineEquation ID="IEq4"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="43578_2025_1548_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">$\%$</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> for Group 1 and Group 2, respectively. Visualization highlighted the model’s detailed understanding of Group 1’s microstructures. Practical demonstration over the Time-Temperature-Transformation (TTT) diagram showcased Group 1’s impeccable micrograph identification. The TMRF Fact Sheet, emphasizing explainability, underscores the framework’s role in fostering trust in microstructural recognition systems.</p> Graphical abstract <p></p>

Tmrf: Trustworthy microstructure recognition framework with deep learning and explainable AI

期刊论文
发表日期：

2025年3月12日

查看全文

Ayush Pratap,
Pao-Ann Hsiung,
Neha Sardana

摘要

This study introduces the Trustworthy Microstructural Recognition Framework (TMRF), aimed at enhancing microstructural recognition with a focus on explainability, human-computer interaction, and accountability. Evaluating microstructural images across seven classes, the framework employed transfer learning models in two experiments: Group 1 (three distinct classes) and Group 2 (all classes). DenseNet outperformed other models, achieving 97 $\%$ $%$ accuracy and 96 $\%$ $%$ F1 scores. Explainable AI (XAI) improved interpretability, with Occlusion and SmoothGrad providing insertion fidelity scores of 60 $\%$ $%$ and 50 $\%$ $%$ for Group 1 and Group 2, respectively. Visualization highlighted the model’s detailed understanding of Group 1’s microstructures. Practical demonstration over the Time-Temperature-Transformation (TTT) diagram showcased Group 1’s impeccable micrograph identification. The TMRF Fact Sheet, emphasizing explainability, underscores the framework’s role in fostering trust in microstructural recognition systems.

Graphical abstract