TY - JOUR
T1 - Tensile strength prediction in directed energy deposition through physics-informed machine learning and Shapley additive explanations
AU - Cooper, Clayton
AU - Zhang, Jianjing
AU - Huang, Joshua
AU - Bennett, Jennifer
AU - Cao, Jian
AU - Gao, Robert X.
N1 - Funding Information:
CC acknowledges support by the National Science Foundation Graduate Research Fellowship under Grant No. 1937968 . JB and JC acknowledge support by the Army Research Laboratory under grant W911NF-18-2-0275 . JC, RG, and JZ acknowledge support by the National Science Foundation under ERC (HAMMER) grant EEC-2133630 .
Publisher Copyright:
© 2023 Elsevier B.V.
PY - 2023/6
Y1 - 2023/6
N2 - In directed energy deposition (DED), local material microstructure and tensile strength are determined by the thermal history experienced at each spatial location on the part. While prior research has investigated the effect of thermal history on mechanical properties, a tensile strength prediction model that is physically interpretable and parsimonious with good predictive accuracy is still needed. This paper investigates a data-driven predictive model with Shapley additive explanation (SHAP)-based model interpretation to address this issue. First, physically meaningful thermal features translated from prior experimental works are used as inputs to a neural network for tensile property prediction. SHAP values are then computed for the individual input features to quantify their respective influences on tensile property predictions and reduce model complexity using the metric of cumulative relative variance (CRV). Prediction of experimentally acquired Inconel 718 (IN718) tensile strength demonstrates that feature influences quantified by the developed method can be verified by findings from prior works, confirming the physical interpretability of the neural network prediction logic. In addition, model complexity reduction based on CRV has shown that fewer than 10% of the original features are required by the parsimonious model to achieve the same predictive accuracy of tensile strength as reported in prior literature, thereby demonstrating the effectiveness of SHAP-based feature reduction method in improving DED process characterization.
AB - In directed energy deposition (DED), local material microstructure and tensile strength are determined by the thermal history experienced at each spatial location on the part. While prior research has investigated the effect of thermal history on mechanical properties, a tensile strength prediction model that is physically interpretable and parsimonious with good predictive accuracy is still needed. This paper investigates a data-driven predictive model with Shapley additive explanation (SHAP)-based model interpretation to address this issue. First, physically meaningful thermal features translated from prior experimental works are used as inputs to a neural network for tensile property prediction. SHAP values are then computed for the individual input features to quantify their respective influences on tensile property predictions and reduce model complexity using the metric of cumulative relative variance (CRV). Prediction of experimentally acquired Inconel 718 (IN718) tensile strength demonstrates that feature influences quantified by the developed method can be verified by findings from prior works, confirming the physical interpretability of the neural network prediction logic. In addition, model complexity reduction based on CRV has shown that fewer than 10% of the original features are required by the parsimonious model to achieve the same predictive accuracy of tensile strength as reported in prior literature, thereby demonstrating the effectiveness of SHAP-based feature reduction method in improving DED process characterization.
KW - Directed energy deposition
KW - Interpretable machine learning
KW - Model pruning
KW - Shapley additive explanation
KW - Tensile strength prediction
UR - http://www.scopus.com/inward/record.url?scp=85149058997&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85149058997&partnerID=8YFLogxK
U2 - 10.1016/j.jmatprotec.2023.117908
DO - 10.1016/j.jmatprotec.2023.117908
M3 - Article
AN - SCOPUS:85149058997
SN - 0924-0136
VL - 315
JO - Journal of Materials Processing Technology
JF - Journal of Materials Processing Technology
M1 - 117908
ER -