K. Brown
Boston University,
United States
Keywords: machine learning
Summary:
Machine learning is a powerful tool for aiding in the design of structures with optimized mechanical performance. Indeed, there is a strong precedent for using machine learning, and other forms of parametric or topological optimization, as part of the design process. In these processes, a design is sequentially altered and evaluated to identify structures with optimized performance. However, the vast majority of approaches for implementing optimization strategies do so by referring to a computational model of the system in question to provide ground truth, or measurements of the underlying property to be optimized. This is appropriate for optimizing properties such as elastic modulus or Poisson’s ratio that are readily and accurately attainable using simulation. For many properties, particularly those related to mechanical failure, experiments remain the best, and in some cases only, way of determining ground truth. In this talk, we discuss our recent progress developing methods for using machine learning tools to advance mechanical designs in situations where ground truth can only be determined experimentally. In particular, we will discuss two approaches to bring machine learning into these arenas of mechanical design. In the first, an artificial neural network is trained using finite element analysis (FEA) to optimize the design of composite structures. Importantly, experiments are used to train the FEA model and the consequences for discrepancies between the model and the experimental system are discussed. In the second example, we discuss strategies for performing mechanical optimization experimentally. Specifically, we will discuss recent progress in transformatively increasing both the rate at which experiments are conducted and the rationale for choosing experiments. Taken together, these examples show that mechanical design can be aided by machine learning even when experiments are an integral part of the design process.