the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Nov 2025
Digital Twin – Driven Machine Learning Optimization Framework for Multizone Curing Control in Wind Turbine Blade Manufacturing
Abstract. The Vacuum-Assisted Resin Infusion Molding (VARIM) process is widely used in wind turbine blade manufacturing due to its cost-effectiveness and reliability. However, challenges such as prolonged curing cycles and defects caused by non-uniform cure remain persistent. To address these issues, multizone heating systems have been developed to enable independent temperature control across blade sections. Yet, optimizing the temperature profile in each zone is computationally intensive, requiring detailed modelling of curing kinetics and heat transfer mechanisms. To overcome these challenges, in this work, a machine learning (ML) based digital twin of the VARIM process was developed using a time-distributed long short-term memory (LSTM) network trained on data generated by a high-fidelity multiphysics solver. The model achieved a predictive accuracy of 96.7 % in replicating the resin curing behavior. Its time-distributed architecture effectively captures the spatial – temporal dependencies across multiple zones, allowing precise prediction of the degree-of-cure evolution. Paired with a gradient-free optimization algorithm, the digital twin reduced curing time by 12.5 % while improving cure uniformity. This AI-driven framework eliminates costly trial-and-error experimentation, and provides a scalable, adaptive solution for improving both quality and productivity in wind turbine blade manufacturing, with strong potential for extension to other composite manufacturing processes.
- Preprint
(2047 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on wes-2025-226', Anonymous Referee #1, 17 Nov 2025
-
RC2: 'Comment on wes-2025-226', Anonymous Referee #2, 19 Dec 2025
This manuscript proposes an approach to optimize the cure cycle of VARIM composite parts by leveraging a multi-zone heating process. Each zone is heated to a temperature optimized by an algorithm using a machine learning (ML) surrogate to evaluate candidate solutions. An experimental proof-of-concept is reported, aiming to demonstrate improved mechanical properties of the manufactured part. The manuscript is generally well-structured, clear, and addresses a relevant topic in wind turbine manufacturing.
While the work is very promising, the current manuscript requires major revisions to address several significant issues that limit its scientific rigor and reproducibility:
- The title and graphical abstract report the development of a digital twin with an active feedback loop for controlling the multi-zone heating system. In practice, the authors implement a digital model for forward prediction of resin cure without data feedback or control. The digital twin nature of the approach should be discussed.
- The manuscript lacks critical details about the ML implementation, including dataset size, model architecture, training procedure, and hyperparameter selection. Since the code is not provided, it makes it challenging to reproduce the results. Additionally, potential limitations, such as generalization to unseen geometry or arbitrary initial conditions, are not discussed.
- The Nelder–Mead algorithm is used for optimization, but its limitations (e.g., local search, sensitivity to initialization) are not discussed. No comparison is provided with alternative approaches, such as simulation-only optimization or modern derivative-free methods like Bayesian optimization.
- The manuscript lacks a quantitative assessment of the results compared to appropriate baselines. For example, it would be valuable to compare the time gains and mechanical properties achieved with ML-based optimization versus those achieved with the conventional optimization approach.
Major comments:
- Abstract, line 20: the 12.5 % curing time reduction and improved curing uniformity are not supported by experimental evidence in the manuscript, nor discussed. The authors should consider removing this claim or providing the supporting evidence.
- Line 108: "To accurately simulate these coupled processes...". What is the definition of an accurate result? The author could provide references to supporting literature.
- Table 1: Consider reporting the values used in this work.
- Line 154: "calibrated multiphysics model". The authors reported the calibration of the cure kinetics only.
- Line 157: The manuscript mentions that CNN-RNN models were chosen, but the authors do not report a CNN.
- Line 191: Consider providing a schematic to help the reader understand the data structure.
- Line 206: Critical details are missing: what library is used to implement the network? Is the code available to ensure reproducibility? What is the dataset size? Is a validation dataset used? What is the data split?
- Line 275: How many simulations were performed on the geometry? Do the simulations include the optimal configuration?
- Figures 6 and 7: Check the curing time unit. If the unit is correct, the manufacturing settings seem unsuitable for an industrial process. Please, discuss.
- Figure 6: The machine learning model matches well with the simulations, which is reasonable, especially if the optimized values are within the domain of the training data. The authors could strengthen the manuscript by characterizing the experimental degree of cure, for instance, by performing DSC on cutout samples and measuring the residual enthalpy.
- Figure 8: Please annotate the figure and describe the testing setup.
- Table 2, Figure 10, line 335:
- The manuscript requires additional data points to support the claims, particularly given the variability.
- Please indicate how the yield strength, compressive modulus, and ultimate strength were derived from the experimental results.
- The cure uniformity was not characterized.
- Figure 10 seems to be truncated.
Minor comments:
- Line 113: "and 𝛥𝑇(𝑡) The". Uppercase.
- Line 126, Figure 1: the SI symbol for Kelvin is K, and °C for degree Celsius.
- Line 312: Compliance calibration is a common practice and may not require detailed information.
Citation: https://doi.org/10.5194/wes-2025-226-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 258 | 106 | 18 | 382 | 12 | 12 |
- HTML: 258
- PDF: 106
- XML: 18
- Total: 382
- BibTeX: 12
- EndNote: 12
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The research paper addresses the optimisation of the curing process for wind turbine blades manufacturing using machine learning coupled with digital twin. For such a purpose a multi-zone heating mould for the VARIM process has been developed. The machine learning tool is supposed to reduce significantly the computatonal time needed for the optimisation.
The paper addresses an important topic in composite manufacturing and put forward a viable solution to optimise cure cycles and maximise part quality. However the paper does not report several relvant information for the understanding of the results and marginally discusses them with a lot left to figure out for the reader. Additionally, too little comparison effort is made to assess the benefits against standard manufacturing. Overall, although the main idea is relevant, the paper does not meet the requirments for publication and I therefore suggest major revisions.
Major remarks:
Minor remarks: