Renewable Energy

Enart Energy Selects Wind Turbine Blade Materials with Data Using EDMAP

Enart Energy was carrying out wind turbine blade material selection based on laboratory test cycles. The EDMAP model developed by Corius was trained on the dynamic mechanical analysis data of 8 epoxy materials; it predicts the temperature-dependent E′ value with R²=0.9997 and the tanδ value with R²=0.9980 accuracy. Material comparison can now be done in minutes instead of a test cycle lasting weeks.

Problem

Blade material selection under varying temperature conditions depended on laboratory test cycles; the process was slow and the comparison criterion remained subjective

Solution

EDMAP (EpoxyDMA-Predictor) — an MLP-based machine learning model; trained on the DMA data of 8 epoxy materials, it predicts mechanical properties from temperature and material identity with R² ≥ 0.998 accuracy

Result

Material comparison dropped from a laboratory test cycle lasting weeks to minutes; the model was moved into the production flow in ONNX format

R² = 0.9997 Above industry standards Model accuracy (E′)
R² = 0.9980 Ready for production deployment Model accuracy (tanδ)
8 materials Epoxy adhesive formulations Training dataset
Weeks to minutes Material comparison time Model output instead of a laboratory test cycle
ONNX Production integration Runs in the cloud and on edge devices
E′ · E″ · tanδ Predicted properties Produced together in the same inference

Which Operational Problems Was Enart Energy Facing?

Enart Energy was carrying out the selection of the epoxy composite materials used in wind turbine blades based on laboratory tests. Turbine blades are manufactured from glass fibre-reinforced epoxy composites operating under high stress; temperature (the range from −10 °C to +45 °C) markedly affects the material's mechanical behaviour. With no systematic material selection process for this variable, the team either chose unnecessarily expensive premium materials or accepted the risk of temperature-induced fatigue. Before installation, material selection in the sector commonly depends on long laboratory test cycles, which slows down the decision process.

  • The temperature–mechanical property relationship is determined by laboratory tests; the test cycle for each formulation took weeks
  • Material selection relied on intuitive experience; the risk of temperature-induced fatigue and delamination damage depended on that decision
  • Storage modulus (E′) and damping (tanδ) values were measured by manual DMA tests and the results tracked in spreadsheets
  • Tests were performed only at the prototype stage; there was no way to retroactively predict the temperature scenarios expected in the field
  • There was no standard numerical criterion for comparing different epoxy formulations

How Did Corius Solve These Problems?

Corius developed the machine learning model EDMAP (EpoxyDMA-Predictor) for Enart Energy. Using temperature and material identity as input, the model predicts Storage Modulus (E′), Loss Modulus (E″) and Tan Delta (tanδ) values in real time. Trained on an MLP architecture, the model was exported to ONNX format so it can be integrated into different operating environments.

01
Data Collection and Exploration
Months 1–2

The DMA (Dynamic Mechanical Analysis) test data used in the project was compiled. The data was cleaned and normalised according to epoxy formulations and temperature profiles, and the model training set was created.

  • Digitisation and consolidation of the laboratory DMA test records
  • Extraction of temperature–property curves for 8 epoxy materials
  • Filling of missing values by interpolation and outlier detection
  • Training / validation / test split (70/15/15)
02
Model Development and Validation
Months 3–4

A multilayer perceptron (MLP) architecture was designed and hyperparameter optimisation was performed. An accuracy of R²=0.9997 for E′ and R²=0.9980 for tanδ was reached. The model was exported to ONNX format.

  • MLP architecture: 3 hidden layers, ReLU activation, dropout regularisation
  • Hyperparameter optimisation
  • Separate output heads for E′, E″ and tanδ
  • ONNX Runtime export for production deployment
  • Cross-validation with engineer-approved scenarios
03
Integration and Pilot
Month 5

EDMAP was integrated into the material selection workflow. By entering the expected operational temperature range, engineers can now compare candidate epoxy formulations within minutes.

  • EDMAP service layer with a REST API wrapper
  • Integration with the internal materials database
  • Lightweight ONNX Runtime setup for platform-independent operation
  • Engineer training and preparation of a user guide
Technologies Used
Python (scikit-learn) PyTorch ONNX Runtime FastAPI Pandas / NumPy

In Which Scenarios Is EDMAP Used?

Region-Based Material Selection

Engineers enter the annual min/max temperature range of the region where the plant will be installed. For each epoxy candidate the model presents E′, E″ and tanδ curves for comparison and recommends the best balance in terms of mechanical performance.

Material comparison time Weeks to minutes
Formulations that can be compared 8 epoxy materials

Temperature Scenario Simulation

The temperature ranges to which the blade is exposed (summer heat waves or frost events) can be simulated on the model to assess in advance whether the material exceeds its critical threshold.

Input The location's operational temperature range
Output E′, E″ and tanδ curves

Formulation Benchmarking

New epoxy formulations are pre-screened with EDMAP before being subjected to the standard DMA test; only the leading candidates are sent for full testing.

Pre-screening Ranking by model output
Full test Only the leading candidates

What Kind of Results Did the EDMAP Project Produce?

R² = 0.9997 Above industry standards Model accuracy (E′)
R² = 0.9980 Ready for production deployment Model accuracy (tanδ)
8 materials Epoxy adhesive formulations Training dataset
Weeks to minutes Material comparison time Model output instead of a laboratory test cycle
ONNX Production integration Runs in the cloud and on edge devices
E′ · E″ · tanδ Predicted properties Produced together in the same inference

We now base blade material selection on temperature data. Without waiting for a laboratory test cycle, we can compare the formulations best suited to the temperature range in which the blade will operate, using the model. This both accelerated material decisions and took them out of the realm of experience.

Let's build a data-driven decision-making infrastructure with a machine learning model tailored to your sector.

Shall we add AI to your engineering processes? Let's build a data-driven decision-making infrastructure with a machine learning model tailored to your sector.