Calibrating and Quantifying Uncertainty for Complex Flow and Transport: Chromium and RDX Regional Models at LANL - 23485

Two unique contaminant plumes below Los Alamos National Laboratory (LANL) property are characterized using similar methods but different models. A plume of hexavalent chromium (Cr⁶⁺ or Cr, in this document) has been detected in excess of the 50 ppb New Mexico groundwater standard in the central portion of the regional aquifer; the highest measured concentration is approximately 1,200 ppb. Additionally, a plume of the organic compound hexahydro-1,3,5-trinitro-1,3,5-triazine (Royal Demolition Explosive [RDX]) with concentrations greater than the New Mexico tap water drinking standard of 9.66 ppb is present in the western portion of the regional aquifer near the mountain block; the highest measured concentration in the regional aquifer is approximately 26 ppb. Numerical modeling is used to improve understanding of the plumes and inform decision-making around remediation strategies. The modeling approach integrates all available information to inform decision-makers and stakeholders with reliable assessment, prediction, and uncertainty quantification. The steps developed for the modeling leverage all lines of evidence available, including site data, CSM and expert knowledge, as well as other information (such as focused data-based studies or pre-existing models). This site information is combined into the numerical model, which includes boundary and initial conditions, discretization in space and time, and simplifying assumptions. These lines of evidence also inform distribution development to constrain input parameters to plausible ranges and quantify the likelihood of certain values. Models are then calibrated against site data ('targets') and validated with data and conceptual information that were not used as targets. Modeling is inherently an iterative process and analytical tools are used at every step to improve process representation, input parameterization, and assumptions. Calibration is performed using a Bayesian approach and vetted through hands-on expert interaction with the model. Multiple top calibrations are identified using a combination of forward draws across parameter prior distributions, hand-tuning, and leveraging a Levenberg-Marquardt (LM) algorithm. The challenges of the calibration include high temporal and spatial resolution, high-dimensional parameter space (including spatially heterogeneous material properties), high-dimensional target data sets, and complex flow fields. A discrete set of top-performing calibrations are used to estimate model uncertainty. Each of these then initializes a Markov Chain Monte Carlo (MCMC) sampler to quantify parameter uncertainty and produce posterior parameter distributions. Both the RDX regional model (RRM) and Chromium regional model (CRM) show very close matches between simulated and observed targets at the site; including time-series of concentrations, hydraulic gradients (both direction and magnitude) local to the plumes, static water levels, and drawdown responses at monitoring wells from pump tests. Combined, these targets ensure confidence that the site physics and hydraulics are well represented in the model. The method presented here for calibrating complex numerical problems and estimating uncertainty is effective for both contaminant plumes at LANL, though each contains a unique decision context, dataset, CSM, and challenges. (authors)