<p>The complex process–structure–property (PSP) relationships of additively manufactured continuous carbon fiber reinforced thermoplastic composites (CCFRTP) govern microstructural formation, and mechanical performance, yet remain difficult to characterize due to the interplay of numerous processing conditions. Existing works are limited by high costs and narrow processing windows, while most data-driven studies rely on small datasets, deterministic predictions, and lack generalizability. This motivates the development of machine learning frameworks that enhance predictive accuracy and provide reliable uncertainty estimates. In this study, a dataset was curated from 95 independent studies, comprising of 654 porosity and 324 tensile strength datapoints with process parameters, filament properties, and physics-informed features. To predict PSP mappings, bootstrapped ensembles of machine learning models—including random forest, extremely randomized trees, support vector regression, kernel ridge, and artificial neural networks—were trained and optimized to capture complex, nonlinear PSP interactions. For predicting the porosity, the support vector regression ensemble achieved the best performance with a test coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">${R}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mrow> <mi>R</mi> </mrow> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>) of 0.8888 and predictive coverage of 95.38%, whereas for tensile strength, the artificial neural network ensemble achieved the best performance with a test <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${R}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mrow> <mi>R</mi> </mrow> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> of 0.8411 and predictive coverage of 96.92%. Furthermore, Shapley additive explanations were utilized for model interpretability and to investigate feature interactions, revealing that thermal and deposition process parameters predominantly affect porosity, while fiber orientation efficiency and fiber volume fraction are the primary determinants of tensile strength. The findings demonstrate that ensemble learning can effectively capture complex PSP relationships, yielding accurate predictions, uncertainty estimates with reasonable empirical coverage, and physically meaningful insights to inform process optimization, defect mitigation, and design of 3D-printed CCFRTPs.</p>

Data–driven prediction and uncertainty quantification of process–structure–property relationships for additively manufactured continuous carbon fiber reinforced polymer composites

期刊论文
发表日期：

2026年3月26日

查看全文

Tongwha Kim,
Kamran Behdinan

摘要

The complex process–structure–property (PSP) relationships of additively manufactured continuous carbon fiber reinforced thermoplastic composites (CCFRTP) govern microstructural formation, and mechanical performance, yet remain difficult to characterize due to the interplay of numerous processing conditions. Existing works are limited by high costs and narrow processing windows, while most data-driven studies rely on small datasets, deterministic predictions, and lack generalizability. This motivates the development of machine learning frameworks that enhance predictive accuracy and provide reliable uncertainty estimates. In this study, a dataset was curated from 95 independent studies, comprising of 654 porosity and 324 tensile strength datapoints with process parameters, filament properties, and physics-informed features. To predict PSP mappings, bootstrapped ensembles of machine learning models—including random forest, extremely randomized trees, support vector regression, kernel ridge, and artificial neural networks—were trained and optimized to capture complex, nonlinear PSP interactions. For predicting the porosity, the support vector regression ensemble achieved the best performance with a test coefficient of determination ( ${R}^{2}$ $R^{2}$ ) of 0.8888 and predictive coverage of 95.38%, whereas for tensile strength, the artificial neural network ensemble achieved the best performance with a test ${R}^{2}$ $R^{2}$ of 0.8411 and predictive coverage of 96.92%. Furthermore, Shapley additive explanations were utilized for model interpretability and to investigate feature interactions, revealing that thermal and deposition process parameters predominantly affect porosity, while fiber orientation efficiency and fiber volume fraction are the primary determinants of tensile strength. The findings demonstrate that ensemble learning can effectively capture complex PSP relationships, yielding accurate predictions, uncertainty estimates with reasonable empirical coverage, and physically meaningful insights to inform process optimization, defect mitigation, and design of 3D-printed CCFRTPs.