Introduction: Fatigue behavior in additively manufactured (AM) 316L austenitic stainless steel is influenced by variations in material properties, processing routes, and testing procedures. Sensitivity to even slight variations in these conditions negatively impacts any direct and simple comparison between individual data sets and precludes the ability to identify larger trends in fatigue behavior and performance. The complicated solidification and thermal history AM alloys experience makes it more important to be able to leverage larger bodies of available data to understand how processing defects and unique microstructures are influencing fatigue behavior. The large number of factors contributing to the variability in the available fatigue data for AM processed 316L austenitic stainless steel coupled with the small sample sizes of each data set limits broader insight into how fatigue properties are impacted by the AM process. By aggregating available smaller data sets and developing protocols for data categorization and standardization, statistical techniques can be used to discern the nuanced and varied behavior of fatigue in AM 316L.
Experimental Procedure: Fatigue data from literature was compiled encompassing both wrought and laser-powder bed fusion (L-PBF) processed 316L austenitic stainless steel. Data were standardized to equivalent fully-reversed axial stress amplitudes by accounting for changes in load ratio and loading configuration, thus reducing the dimensionality of the obtained data set. Commonly reported variables pertaining to mechanical properties, processing conditions, and surface conditions were categorized.
The objectives of the statistical meta-analysis were to describe the general scatter of fatigue behavior, quantify contributions to varied fatigue life, and to discern between defect and microstructure driven behavior from the fatigue response, or S-N curve. All statistical analyses were performed using the R programming language. General scatter was determined for data subsetted by each categorized variable. Simple linear regression models provide coefficients of determination and mean fatigue lives. Contributions to fatigue life were then quantified using an analysis of covariance (ANCOVA) model with an additional blocking factor to improve model precision. An iterative process of model formulation, variable selection, and validity checking was performed to obtain a statistically and practically sufficient model. Each compiled S-N curve dataset was then regressed using Basquin’s equation to obtain the fatigue response. After performing a similar procedure to identify contributions to fatigue response in an ANCOVA model, fatigue responses were clustered using hierarchical and k-means clustering techniques to determine which groups of material exhibited similar behavior.
Results from the meta-analysis were extended to runout and strain-controlled fatigue data. Fatigue behavior groups in runout were compared using a Weibull analysis with a Kolmogorov-Smirnov test to test group membership. A set of specimens were produced and tested under strain-controlled fatigue for comparison to compiled data. Specimens were hot-isostatic pressed (HIP) to mitigate defect effects. Testing was conducted at strain amplitudes greater than 0.002 mm/mm.
Results and Discussion: Linear models and data handling practices were effective in capturing scatter and contributions to both fatigue life and response in AM processed 316L fatigue. Load ratio and loading method could be effectively standardized to remove their main effects from models for fatigue life. Models for general scatter and fatigue life were able to quantify and verify several of the commonly relied upon relationships that tend to be used qualitatively. Impacts of surface machining, application of stress relief, annealing, or HIP treatments, and build angle were obtained.
Analysis of fatigue responses, in addition to quantifying contributions to changes in S-N curve slope, was able to discern two statistically and practically distinct groups of fatigue behavior. By considering the categorized variables and parallels between grouped datasets, these two groups in AM 316L could broadly be attributed to defect driven and microstructure driven failure. Conditions where the presence of processing defects or their coupling with the applied stresses diminished favor the microstructure driven failure group. HIP treatment, machining, and choosing more favorable orientations with processing defects favor microstructure driven fatigue. While standardization removed the main effects of load ratio and loading method for fatigue life, their impact was still observed for fatigue response. More severe loading conditions favored defect driven fatigue.
Extending these results to strain-controlled and runout data demonstrated a general dichotomy between defect and microstructure driven fatigue. HIP strain-controlled fatigue data obtained for this study exhibited nearly wrought behavior albeit a decreased fatigue strength in contrast to the stark deviation of defect driven material in both the low-cycle and very-high cycle fatigue regimes. Weibull analysis of runout data confirmed that the two fatigue behavior comes from different statistical distributions with carry over to the very-high cycle fatigue regime.
Conclusion: Published fatigue data with varied testing conditions can be effectively compiled and analyzed using statistical and meta-analytic techniques in a way extensible to multiple regimes of fatigue failure and new datasets. Protocols for this type of meta-analysis were developed and able to provide quantitative insight to both commonly observed qualitative knowledge and reveal new insights regarding the fatigue behavior of AM 316L. In particular, analysis of the fatigue response broadly showed defect driven and microstructure driven fatigue behavior in AM 316L. This dichotomy may be related back to the assessed categorized variables and how they act to either mitigate the presence of processing defects or diminish their interaction with imposed stresses.