Data-efficient neural operator framework for curing deformation field prediction in composite manufacturing process design

Composite manufacturing process design and optimisation require massive iterative evaluations that map process parameters to the resulting deformation fields. High-fidelity numerical models of composite curing are accurate but prohibitively time-consuming, making them impractical for iterative optimisation. Data-driven prediction methods can accelerate deformation field prediction, but typically require large amounts of high-fidelity training data. To address these limitations, this study presents a data-efficient multi-fidelity neural operator framework for curing deformation field prediction. The framework first applies aggregation-value-based sampling to sample a small set of informative high-fidelity samples, thereby reducing labelling costs. A physics-guided low-dimensional neural operator (LNO) is then developed, embedding Laplacian eigenfunctions of the geometry and POD modes of field data to learn in a reduced spectral space with lower complexity. Finally, transfer learning is employed to fine-tune the LNO pretrained on low-fidelity data with the selected high-fidelity samples, further reducing data requirements. Application to an aerospace CFRP component shows that the proposed approach achieves high prediction accuracy with substantially fewer high-fidelity simulations than baseline operator models. These results demonstrate its potential as a practical and data-efficient tool to support intelligent design and optimisation of composite curing processes.