GeoMix simulation study — MCMC outputs and predictions

Title: GeoMix simulation study — MCMC outputs and predictions Description: This deposit contains the pre-computed MCMC outputs and model predictions from the simulation study in: Wakefield, B. (2025). GeoMix: A Bayesian Hierarchical Model for Joint Inversion of Geotechnical and Geophysical Data. The files allow the simulation study figures and tables to be reproduced without re-running the computationally intensive MCMC chains. All files are in R's native .rds format and are read directly by the analysis scripts in the companion code repository (https://github.com/bradleywakefield/geomix-paper). Folder structure results/simulation/ ├── GeoMix_1/ … GeoMix_4/ ← 4 parallel GeoMix MCMC chains │ ├── batch_1.rds … batch_32.rds │ └── geomix_setup.rds ├── LGFM_1/ … LGFM_4/ ← 4 parallel LGFM MCMC chains │ ├── batch_1.rds … batch_32.rds │ └── geomix_setup.rds ├── predictions/ │ ├── GeoMix_predictions.rds │ ├── GeoMix_params.rds │ ├── LGFM_predictions.rds │ ├── LGFM_params.rds │ ├── GP_predictions.rds │ ├── GPwC_predictions.rds │ ├── GW_predictions.rds │ ├── GWwC_predictions.rds │ ├── LM_predictions.rds │ ├── data.rds │ └── pred_df.rds └── tables/ ├── GeoMix_class.csv … GeoMix_worst_rhat.csv └── LGFM_class.csv … LGFM_worst_rhat.csv File descriptions MCMC chain folders (GeoMix_1–4/, LGFM_1–4/) Each folder holds one MCMC chain. Sampling was run with 4 parallel chains, 4,000 iterations per chain, thinned and saved in 32 batches of 125 iterations. batch_1.rds – batch_32.rds — Raw MCMC draws. Each file is a numeric matrix (125 iterations × 8,110 parameters). Columns index the full joint parameter vector of the model (stratigraphy field Y1 at every lattice node, plus covariance and mixing parameters). geomix_setup.rds — The setupGeoMixModel() configuration object used to initialise that chain. Contains model constants, data lists, initial values, lattice structure, Gibbs/HMC control settings, and the hexagonal grouping used for parallel depth-slice sampling. predictions/ Post-processed outputs used directly by scripts/simulation/03_results.R to produce all figures and tables. GeoMix_predictions.rds — GeoMix predictive distribution at test locations. List with elements $mean (posterior mean), $sd (posterior SD), and $samples (posterior predictive draws) for the latent geotechnical property Y2. GeoMix_params.rds — Merged posterior samples across 4 chains. List with elements $params (named list of posterior draws for each parameter: a0, a1, sigma2, tau2, sigma2_L, sigma2_D, lD, lL, gamma, h, Y1prob, Y1) and $samples (the full draws object). LGFM_predictions.rds — As above for the Latent Gaussian Field Model (LGFM) baseline. LGFM_params.rds — As above for LGFM. GP_predictions.rds, GPwC_predictions.rds — Gaussian Process predictions (without/with Z1 covariate). List with $mean and $samples. GW_predictions.rds, GWwC_predictions.rds — GeoWarp predictions (without/with Z1 covariate), same structure. LM_predictions.rds — Linear model predictions, same structure. data.rds — Training split of the synthetic dataset (3,200 observations). Columns: d, x, y, ID, locID, dID, group, Y1, Y2, mu, sig, Z1, Z2, SU, SU1–SU8. pred_df.rds — Test split (4,800 observations, Z2 withheld during fitting). Same columns as data.rds except Z2 is absent. tables/ MCMC convergence diagnostics for GeoMix and LGFM, computed from the merged 4-chain posterior. {model}_params.csv — Per-parameter summary: param, ess_bulk, ess_tail, rhat, sd, mcse_mean, mcse_sd, q025, q50, q975, mcse_over_sd, class. {model}_class.csv — Diagnostics aggregated by parameter class: median/max R-hat, median/min bulk and tail ESS, median/max MCSE/SD. {model}_overall.csv — Single-row overall chain health summary: counts of R-hat > 1.01 and > 1.05, max/median R-hat, min/median bulk and tail ESS. {model}_worst_ess.csv — The 10 parameters with the lowest bulk ESS. {model}_worst_rhat.csv — The 10 parameters with the highest R-hat. How to use Download and unzip the deposit into the repository root so that results/simulation/ is present. Run source("run_simulation.R") in R. At the ask_run_mode() prompt, select load to skip MCMC and proceed directly to figure and table generation using these files. Software All files were produced using R 4.4.1 with the geomix R package (https://github.com/bradleywakefield/geomix) and nimble 1.4.1. Package dependencies are fully specified in renv.lock in the companion repository.