Open Access
Subscription or Fee Access
Prediction of Fiber-reinforced Composite Material Properties Using Collaborative Filtering Techniques
Abstract
The traditional structural design process follows a deductive, “bottom-up†approach in which the manufacturing process determines the material structure (e.g., crystalline structure), which in turn determines the material properties, which in turn govern the ultimate structural performance. This causal chain of analyses is at the core of the integrated computational materials engineering (ICME). There are two key challenges in the implementation of ICME. First, accounting for the processingstructure- properties-performance relationship often requires very detailed computational simulations. These simulations can be very expensive to perform, which consequently makes an exploration of a material design space expensive, if not outright intractable. Second, for a given material, different processing methods and environmental conditions can give rise to dramatically different macro-level performance characteristics at the structural scale. A machine learning framework has been developed to address these challenges. The proposed framework reduces the amount of testing required to obtain material properties and avoids the need for microscale modeling. In this research, collaborative filtering (CF) is used to find missing material properties from similarities among material families at both the constituent level (i.e., fiber and matrix) and the laminate level. The properties obtained with CF can be used in the traditional deductive approach to determine performance at the part level. Several CF techniques were explored and implemented in a flexible workflow that may be used to find missing properties from an existing material database. These techniques include several machine learning algorithms, as well as an invariant-based approach to represent elastic properties recently developed by Stephen Tsai and Daniel Melo (2014) to enhance the material data set. Some of the techniques allow for stochastic prediction and provide statistical distributions of the missing properties instead of a single deterministic prediction. This information is then used to quantify performance uncertainty.
DOI
10.12783/asc33/25957
10.12783/asc33/25957
Full Text:
PDFRefbacks
- There are currently no refbacks.