You are here

Machine Learning and Data-driven Modeling to Simulate Damage, Fracture and Failure

John Emery, Sandia National Laboratories

Geoffrey Bomarito, NASA

Jacob Hochhalter, University of Utah

Kyle Johnson, Sandia National Laboratories

James Warner, NASA

Fracture and failure of engineering materials are among the most challenging solid mechanics phenomena to predict. For these phenomena, development of computational modeling approaches for improved predictions has been the primary focus for several decades.  These approaches include a broad range of pertinent material length and time scales as well as the requisite numerical methods for their implementation, e.g. Peridynamics and Phase Field methods. Nevertheless, challenges remain, including the need for constitutive models to capture variability in complex stress states for general ductile, anisotropic materials with rate and temperature dependence. While success has often been demonstrated for lab-scale testing, inclusion of these advanced modeling methods within engineering workflows and propagating this variability for engineering-scale uncertainty quantification (UQ) is a significant remaining technology gap.  This gap is made especially important given the limited data sets that are common to full-scale engineering scenarios. 
Abstracts are solicited that address these limitations by applying emerging data-driven and/or machine learning (ML) algorithms to identify and capture failure mechanisms, develop new models or model forms, replace existing models, or aggregate community data in a tractable and efficient manner to enable UQ in complex engineering scenarios either. Examples include, but are not limited to, ML algorithms trained on simulated data to replace analytical constitutive models, ML for scientific discovery and data aggregation, ML algorithms to enable multiscale analysis, and methods of overcoming sparse datasets, such as incorporation of physical constraints, e.g., physics-informed ML. Further, work to fundamentally change the engineering failure-prediction workflow is of keen interest. If ML can be trained to identify combinations of materials’ microstructural initial conditions and the service loadings that lead to failure, failure prediction will move beyond model-form and mesh dependencies and avoid the particulars of how to upscale and deliver microstructurally-informed failure predictions with minimal computational requirement.