Machine learning based dual flat-spherical indentation approach for rough metallic surfaces

Abstract

Instrumented indentation test (IIT) is regarded as one of the non-destructive test methods for mechanical characterization and safety assessment of engineering components. However, indentation test is very sensitive to surface topography i.e surface roughness, requiring specimens with ideally flat and smooth surface for producing reliable and reproducible test results. Recently, a dual flat-spherical indentation (DFSI) technique was proposed to extract material properties directly from rough metallic surface via indentation tests. Integration of DFSI technique with machine learning (ML) approach can facilitate reliable and fast characterization of rough metallic surfaces. Here, ML models including artificial neural network (ANN) and physics-informed artificial neural network (PI-ANN) are established, and then trained by using a database of indentation parameters generated via DFSI simulations. Spherical indentation techniques available in the literature are used to calculate the physics-informed loss in the PI-ANN model. The performances of trained ANN and PI-ANN models are compared, and an optimal ML model with hyperparameters is suggested based on validation study; PI-ANN model with sigmoid activation function shows a better performance than the ANN model and baseline model. The model performances are evaluated by performing DFSI experiments with SM45C and SS304. The applicability of ML based-DFSI method is demonstrated for extracting the mechanical properties from the rough surface of general metals.