May 15 - 10:25 AM to 10:55 AM
Presented by Mr. Brice Shireman, Principal R&D Engineer, Boston Scientific
Mr. Darrin Beekman, R&D Engineer II, Boston Scientific
Mr. Ismail Guler, R&D Fellow, Boston Scientific
Mr. Bryan Plunger, R&D Engineer II, Boston Scientific
Mr. Zack Graves, Sr. Application Engineer, SmartUQ

Slotted metallic tubes play an important role in the medical device industry as the mechanical backbone for various devices. Conventional response surface models built from physical bench tests often lack the accuracy required for setting tolerances on the dimensions of these slotted tubes. The miniature size of the slots coupled with the precision capability of manufacturing equipment make it difficult to produce multi-factor samples at the exact input levels required by the sampling plan, which can reduce model accuracy. Therefore, a new approach is required for slotted-tube tolerance setting. This study explores the ability of Gaussian Process (GP) emulators, a class of surrogate model, to predict the functional performance of tubes for dimensional tolerancing applications. A series of GP emulators were constructed using training data from FEA based computer experiments that explore variation of model and geometric parameters of interest. Experimental data was used to calibrate the model parameters in the GP emulators to improve the model accuracy. Residual error between the calibrated GP emulators and the experimental data, the discrepancy, was mapped using a secondary GP emulator. A combination of SolidWorks®, ANSYS®, and Abaqus™ software packages were used to run computer experiments with 140 unique FEA models that simulate bending of a tube segment. Each model run had a unique set of five dimensions used to define the tube. A modified Latin hypercube approach determined the sampling of the five-factor design space. The output of the FEA simulations, specifically the maximum equivalent plastic strain (PEEQ) and the bending stiffness, were used to create continuous response emulators in SmartUQ® software. FEA-predicted PEEQ values were translated to fatigue performance (cycles to failure) using the Coffin-Manson low-cycle fatigue relationship. To aid in calibration, sample dimensions were selected to encompass the practical edges of the design space that could be used for typical tube dimensions. To efficiently test this practical design space, a 24 factorial sampling design with two semi-center points was utilized. After the parts were manufactured, they were characterized. Two bench tests were used to measure the performance of the parts: a rotary bending fatigue test and a cantilever stiffness test. Using the performance and dimensional characterization data from 18 unique batches of parts, the emulators were calibrated via a Frequentist technique within SmartUQ®. The calibration was subsequently used to tune the Coffin-Manson coefficients and the Young’s modulus of the tube material to most closely emulate the results observed from physical testing. This calibration technique within SmartUQ® not only tuned the tube parameters, but also produced a discrepancy model, which predicts the error between the emulator prediction and physical result across the multi-factor design space. Adding the predictions of the physical emulator and discrepancy emulator together significantly reduced the total error in the fatigue predictions. However, the discrepancy emulator did not improve the prediction of stiffness, which suggests small model-form error for this emulator. To assess the accuracy of the calibrated emulators, a validation was performed using a Leave-One-Batch-Out cross validation approach. Seventeen of the 18 calibration batches were used to calibrate the FEA and generate a discrepancy model. The resulting emulator was used to predict the performance of the remaining batch, which was not included in the calibration set. This was done 18 times using 18 unique calibration sets and leaving out each of the 18 batches in turn. The final accuracy of the emulators was reported in the form of Root-Mean-Square-Error (RMSE) from the Leave-One-Batch-Out cross validation results. The validation step quantified the prediction uncertainty between the emulator and physical testing in order to appropriately guard-band any predictions used for tolerance setting. The FEA based, calibrated emulators coupled with the discrepancy emulators exemplified high prediction accuracy of complex miniature geometrical interactions. The emulators can be used as a tool to simulate the conditions the end-product will see in a clinical environment. This allows the optimization of the design to ultimately improve the product performance and enhance the patient safety.

May 17 - 8:30 AM to 9:00 AM
Presented by Zack Graves, Sr. Application Engineer

The growing use of simulations in the engineering design process promises to reduce the need for extensive physical testing, decreasing both development time and cost. However, as statistician George E. P. Box said, “Essentially, all models are wrong, but some are useful.” For this reason, model validation such as outlined in ASME V&V 10 and V&V 40 is a crucial step in assessing the credibility of a simulation model before it can be used in lieu of physical data. Experimental data collected during validation can be compared to simulation outputs for model assessment in a number of ways. This talk will focus on two such methods, Bayesian calibration and an area metric method. Bayesian calibration is a method capable of both optimizing the tuning of model parameters to improve simulation accuracy and estimating any remaining discrepancy between the experimental and simulation data sets. The discrepancy model when analyzed can indicate areas where the model needs the most improvement. Additionally, as model discrepancy is assumed to exist in this framework, robust calibration is enabled even for inaccurate models. The area metric quantifies simulation discrepancy by making a comparison between the cumulative distribution functions (CDFs) of the simulation and the available experimental data. The simulation model is sampled to create a CDF for the output of interest which is plotted along with a CDF for the experimental data. The area metric is taken to be the sum of the area enclosed by the pair of CDFs. This presentation will discuss the benefits of and differences between Bayesian calibration and the area metric method as applied to model validation. A balloon expandable coronary stent model will be used as an example for illustrative purposes.