Machine learning-based stress-strain prediction and process parameter optimisation in material extrusion additive manufacturing

Material extrusion (MEX) additive manufacturing (AM) has emerged as a widely adopted method for fabricating complex polymer components, particularly in customised and lightweight engineering applications. However, the layer-by-layer deposition process inherent in this technique results in considerable variability in mechanical performance. Among the most critical challenges is the difficulty in accurately predicting and controlling the full stress-strain behaviour of the printed parts. A reliable understanding of the stress-strain response is essential not only for assessing the structural safety but also for guiding the design of printing parameters tailored to the application-specific mechanical requirements. Despite progress in empirical modelling and data-driven approaches, most existing methods focus on isolated mechanical indicators such as yield strength or modulus. These approaches often demand large datasets and fail to generalise across changing process parameters or testing conditions. Furthermore, an inverse prediction of the process settings from the desired mechanical performance remains an underexplored area, particularly in scenarios where experimental data are limited. This research addresses these limitations by proposing a data-efficient, machine learning-based framework for the stress-strain curve prediction and inverse parameter design in material extrusion additive manufacturing. The proposed approach encompasses three integrated studies, each targeting a specific modelling challenge: inverse learning, cross-domain generalisation, and strain-rate-sensitive response prediction. The overall aim is to establish a generalisable methodology capable of supporting performance-driven design decisions with minimal experimental input. In the first study, an inverse learning strategy is developed using a long short-term memory (LSTM) network. This model is trained on experimentally derived stress-strain curves from sixteen different printing parameter sets, comprising forty-eight tensile tests. By learning the temporal relationship between strain and stress evolution, the model accurately predicts the process parameters required to reproduce a given target curve. The inverse model achieves a coefficient of determination (R2) of 0.8648 and a mean squared error (MSE) of 0.1348 on unseen test data, confirming its potential for supporting reverse parameter identification based on user-defined mechanical goals. The second study introduces a transfer learning framework that enhances model generalisation across distinct process parameter domains. A hybrid architecture combining LSTM and temporal convolutional networks (TCNs) is first trained on datasets involving variations in the print angle. The pre-trained model is then fine-tuned using a small dataset from a new parameter domain involving an infill variation. The adapted model achieves a root mean squared error (RMSE) of 0.09. These results demonstrate that the framework successfully transfers predictive knowledge between domains with minimal data, offering a practical solution for the data-scarce additive manufacturing settings. The third study explores the influence of tensile speed on stress-strain behaviour. Four sets of identically printed specimens are tested at varying strain rates. A decision tree (DT) model is constructed to predict the deformation responses under different loading conditions. The model effectively captures the rate-dependent variations in yield strength and post-yield response. Moreover, it successfully predicts the mechanical behaviour of a test curve excluded from training, validating its robustness in generalising across untested strain rates. This part of the study underscores the feasibility of interpretable, low-data models in capturing loading condition effects in MEX-fabricated components. All proposed models are validated against physical experiments using polylactic acid specimens fabricated by fused deposition modelling (FDM). Together, these three investigations form a coherent framework that supports both forward prediction of stress-strain responses and inverse identification of optimal process settings. The findings illustrate that with appropriate model design and transfer strategies, accurate full-curve predictions can be achieved without the need for extensive data collection. This work contributes to the development of intelligent, data-efficient design tools for additive manufacturing and establishes a foundation for future integration of digital twin systems and closed-loop process control environments.<p></p>

Author dc:creator

• Linlin Wang (21052232)

• Additive manufacturing
• Machine learning
• 3D printing
• Mechanical properties

Statement dc:rights

• All rights reserved
• Open Access after 2027-10-27

Identifier

10779/exe.32076291.v1

OAI identifier oai:identifier

oai:figshare.com:article/32076291