Machine Learning Helps Evaluate Mechanical Performance of Composites

Materials engineers use CAE software advances to expedite composite development.

Hexagon has submitted this post. 

Written by: Rani Harb, Solutions Lead Materials & Data, Design & Engineering, Hexagon.

In the past few decades, machine learning research has yielded a wealth of technology advances that are just beginning to be leveraged by the engineering community. Materials engineers, material suppliers and original equipment manufacturers (OEMs) can now take advantage of machine learning to develop composite materials with greater performance and in shorter development cycles.

Figure: Illustration of the Design Space and Mechanical Response Under Various Conditions- Empty cell represent a condition or formulation not tested.

Illustration of the Design Space and Mechanical Response Under Various Conditions- Empty cell represent a condition or formulation not tested. (Image courtesy of Hexagon.)

Consider the following scenario. A group of materials engineers are tasked with developing new advanced composite materials to meet specific mechanical performance. The design space is immense and typically includes the resin type, reinforcement type and amount, processing agents, treatments, etc. In addition to the intrinsic variables, the composite material will exhibit different thermo-mechanical behaviors depending on the operating conditions—such as temperature, humidity, strain-rate, UV exposure and more.

Physical experimental testing or the use of simulation software to describe the governing physics have been the traditional approaches to evaluate the performance of a composite under various conditions. Due to time, financial and resource constraints, materials engineers must compromise and test a fraction of possible material formulations considered to be the best candidates. This compromise means optimal materials may be overlooked, and material suppliers may spend years evaluating and incrementally improving formulations to meet a set of conflicting requirements.

Machine Learning Can Expedite Materials Development

Machine learning research has given rise to the use of meta-models, sometimes called surrogate models or emulators. Engineers are familiar with the linear model, which is constructed via linear regression or least squares regression. A linear model imposes a linear relationship between input variables and output responses, and this model type can be limited and unsuitable for materials engineers. Meta-models are more sophisticated and capable of capturing highly nonlinear responses. An original database compiled from dozens or hundreds of composite tests can be used to train a machine learning algorithm, such as neural networks, offering remarkable flexibility.

In addition, data augmentation is performed using meta-modeling in the temperature and strain-rate dimensions. The surrogate models are then used to predict the output of composite configurations yet to be manufactured and physically tested. The most significant benefit of a meta-model is that it takes seconds to provide an output—such as a stress or strain relationship. This process presents a significant contrast to the days or weeks required to physically manufacture and test composites. The more test data that is available to train a meta-model, the more accurate meta-models will be.

Figure: Data Augmentation Along 3 Dimensions: Orientation, Fiber Content and Temperature

Data Augmentation Along 3 Dimensions: Orientation, Fiber Content and Temperature. (Image courtesy of Hexagon.)

Simulation Technologies Further Enrich the Process

In addition to leveraging the power of physics-informed meta-models, simulation technologies such as mean-field homogenization and finite element analysis (FEA) are used to further enrich the data along orientation and fiber content dimensions. To accurately simulate composite materials, the material models used must be able to reflect the varying anisotropic behavior of composites. Mean-field homogenization has been documented to produce accurate material models of composites, including short or long fiber reinforced composites. When combining FEA and mean-field homogenization, accurate mechanical simulations yield insight to the stress, strain, fatigue or creep behavior of fiber-reinforced plastics.

As for how materials engineers might leverage machine learning, the first stage involves evaluating a fixed number of composite candidates. Let’s say 100 physical tests are conducted. Reverse engineering is performed to determine the composite’s constituent mechanical properties. The reverse engineered mechanical properties and material models are then used to simulate 500 different composite candidates. The 100 physical tests and 500 simulated tests are used to train a machine learning algorithm such as a neural network. The neural nets are then deployed to evaluate hundreds or thousands of composite configurations rapidly and accurately—in seconds or minutes. With such algorithms, materials engineers are able to evaluate more material candidates and identify optimal composite formulations in a fraction of the time compared to traditional means. Materials engineers and suppliers stand to greatly benefit from machine learning.

Figure: Illustration of Reverse Engineering of Constituent Properties Using Micromechanics (Digimat)

Illustration of Reverse Engineering of Constituent Properties Using Micromechanics (Digimat). (Image courtesy of Hexagon.)

What Machine Learning Tools are Available to Materials Engineers?

Hexagon’s Manufacturing Intelligence division develops a number of software tools for materials engineers to leverage machine learning, including ODYSSEE, Digimat, MSC Nastran and Marc. ODYSSEE provides the meta-models discussed earlier, and features a number of additional machine learning capabilities. Digimat is used to determine the material models, such as constitutive relations of composite materials and reverse engineer phase properties by leveraging mean-field homogenization. The material models, based on mean-field homogenization, can then be incorporated in mechanical simulations in either MSC Nastran or Marc. MSC Nastran and Marc are FEA software tools commonly used for stress and deformation analysis.

There are substantially more machine learning capabilities available. To learn more, visit MSC Software’s website.