(707e) Multi-Fidelity Deep Neural Network for the Prediction of Mechanical Properties of Polymer Matrix Composites for AI-Assisted Materials Design

Automotive material design is a traditional inverse problem of optimizing mechanical properties, and it requires a lot of predictions of mechanical properties and behavior depending on constructed materials. One of the efforts to minimize trial and errors from the design of engineering plastic materials to commercialization is the establishment of a Computer-Aided Engineering (CAE) system for automated material design. Even with an automated workflow, the mechanical properties and behavior of polymeric materials must be predicted to operate the CAE system. Engineering plastics used in automobile materials require excellent long-term heat resistance, unlike general-purpose plastics, and mechanical properties with minimum tensile strength of 500 kg/cm² and impact strength of 6 kgf·cm/cm² or more over a wide temperature range. In order to consider the aforementioned conditions, the information on mechanical properties and behaviors required when designing polymer matrix composites includes tensile strength, elastic modulus, maximum load, maximum stress, and breaking point stress. Many of them can be easily derived from stress-strain curves (S-S curve) and creep curves obtained from tensile tests, impact tests, and creep tests, or can be obtained from models with sufficient accuracy. Therefore, the prediction of the S-S curve and creep curve is the key to the mechanical property behavior, and if this part is automated, the entire CAE can be automated.

In this study, the S-S curve and creep curve prediction model considering microstructure for the combination of 2~5 components of matrix types such as amorphous polymer and semi-crystalline polymer and fillers such as ceramic powder, glass fiber (GF), carbon fiber(CF), carbon nano tube, etc., which are often used as materials for automobiles. This time, version 1 developed a model using point data for measuring the mechanical properties of pure materials, and version 2 was developed to predict the mechanical properties of pure materials without information on the mechanical properties of the pure material. In this presentation, as the first achievement of the study, a data-driven model was created based on the mechanical property behavior of the combination that the research team has already accumulated.

In this paper, a system for predicting mechanical property behavior of binary polymer matrix composites is developed using deep neural network methodology by learning tensile test data for two different compositions in a binary combination of matrices and fillers. Numerical analysis methods such as FEM, which are conventionally used to predict the behavior of S-S curve, take a long time to calculate, and for complex models, calculation performance is poor. Therefore, studies are being attempted to overcome this through (large-scale) data-based prediction. The polymer matrix composites for which mechanical properties are predicted are a combination of matrices such as PA6, PP, PC and PA6,6 and fillers such as Al2O3, Al2O5Si, Si3N4, and Bn. This material combination is currently actively used as a main material for transportation equipment such as electric vehicles, and there are many tensile tests and physical property data, which is an advantage for prediction using big data-based machine learning. Tensile test data is removed because there are many unnecessary string information, and all strings are converted to integers and used for prediction. Test data are stress (MPa), stroke elongation (%), stroke (mm), and load (N) data over time, measured at 0.001 sec intervals. Since it is data at a time interval of 0.001sec, about 1.5 million data had to be learned in the tensile test, so the time interval is adjusted to improve performance.

This prediction model predicts the stress value according to the strain at a given combination and composition by using the following tensile test data of the polymer composite resin. The strain at the given combination and composition and the elastic modulus, density, molecular weight, and structure (as SMILES) information of each matrix/filler pure material were entered. Increased performance. Tensile test data was randomly divided into train, validation, and test data set to proceed with learning. The deep neural network structure used 5 layers, and the hyperparameter values were obtained through bayesian optimizer. Finally, when the verification was conducted through cross-validation, compared with the existing experimental data, the performance of R^2=~0.83 value was shown, and the mechanical properties under conditions that were not used for training were predictable. As a result, this is after verification and performance verification of functional matrix test data mainly used in automobiles and electronic products, and the proposed hybrid prediction model showed improved performance compared to the existing first principle model or data-based model. The proposed model is predicted by reflecting not only data but also physical properties, and it is expected that it will be possible to predict only mechanical and physical properties without relying on experimental data in the future.

Although accuracy results are obtained with single fidelity previously, there is not always sufficient accuracy data for the matrix and filler, so studies based on multi-fidelity that use less expensive experimental data sets are also in progress. By adding one more single-fidelity model to the single-fidelity model above, we complete the Multi-fidelity Deep Neural Network in which two Single-fidelity NNs using different fidelity data are layered. Multi-fidelity DNN uses tensile test data (high-fidelity, HF) and CAE simulation data (low-fidelity, LF) universally, so it has an average of 5% performance with fewer data sets compared to models using only HF data. It is expected to show improvement. With the addition of the LF data-driven single-fidelity model, human errors and experimental errors in the test can be compensated for, thereby improving the predictive performance in situations where the mechanical property data set is insufficient.

The prediction of mechanical properties of polymer matrix composites has been less studied than inorganic or metals, and the information and properties of pure materials and data-driven machine learning without microstructures or microscopic images are the first attempts by our researchers. Ultimately, for materials mainly used in automobiles and electronic products, we want to support the function of automatically designing engineering plastics that meet the target properties with only pure material information without tensile test data in matrix/filler combinations in 2 to n components.