Lessons Learned from Integrating a Surrogate Model into a Finite Element Software

A common approach for performing multiscale modeling of materials is to define a micromechanics model that captures the finite element (FE) integration point behavior. Both open-source and commercial software often allow users to make this connection through user-defined subroutines. In a typical workflow, a predefined set of inputs are provided by the FE software, the external software executes using those inputs, and predefined outputs are returned to the FE software. In previous work, NASA researchers developed a surrogate model using the Tensorflow machine learning software that is intended to replace a computationally expensive physics-based micromechanics model. The NASA Multiscale Analysis Tool (NASMAT) was used to generate nonlinear material responses across a range of inputs for use in generating training data for the surrogate model. Both the physics-based and surrogate modeling approaches will be briefly summarized. Tensorflow was then coupled to the Abaqus finite element solver to perform a multiscale analysis via a custom-developed interface. Results are shown documenting the performance of the surrogate model compared to a traditional multiscale modeling approach where the physics-based model is called at each integration point in the FE model. Benefits and limitations of the implemented approach will be addressed. A key contribution of the presentation will be discussing the lessons learned from this exercise. This discussion will address the implication of software architecture choices, the role of the model execution environment, and key profiling metrics. Suggestions for improving the adoption of surrogate models within FE tools will also be addressed. While specific software tools were utilized in this effort, the lessons learned are broadly applicable to other available tools.