Beyond the fit: assessing bias-variance trade-off in johnson-cook constitutive modeling for AA5182

Physics-based predictive models often contain known inadequacies due to their simplified representation of underlying physical systems, introducing systematic deviations from reality. Despite this, they tend to generalize better than their purely empirical counterparts and are often the only tools available to evaluate or predict the performance of complex engineering systems. In this paper, we present a principled framework for calibrating simulation models with known deficiencies against experimental data and for evaluating their predictive ability. Discussions explore the modeling of a cylindrical AA5182 aluminum alloy specimen tested in compression at temperatures ranging from 200 − 500 °C, and at strain rates of 0.001 and 1 s - 1 . A finite element model of the specimen is constructed in Abaqus/CAE 2025 using the Johnson-Cook constitutive model, a common choice for simulating alloy behavior due to its computational efficiency. The Johnson-Cook model has several alloy-specific parameters that must be calibrated against experimental data, introducing uncertainty into the finite element model predictions. Additionally, its simplified treatment of strain hardening leads to systematic discrepancies between simulated and observed material responses, which can be inferred through an independent, empirically trained error function. Herein, we explicitly account for these uncertainties and discrepancies and evaluate the bias-variance trade-off in the model predictions. We propose a metric, Calibration Flexibility Measure, to quantitatively evaluate the trade-off between goodness-of-fit and model complexity. Key decisions, such as selection of calibration parameters, integration of prior knowledge, and definition of model-form error are explored to assess and characterize their effects on posterior distributions and predictive performance.