How Much Can Random-Field Dimensionality Be Reduced for Surrogate-Based Bayesian Contaminant Source Identification in Heterogeneous Aquifers?

Contaminant source identification in heterogeneous aquifers is classically hindered by high parameter dimensions. The unknown hydraulic conductivity field must be represented by a spatially correlated random field, which introduces many unknowns alongside the source parameters. Jointly inferring them requires repeated forward-model evaluations, which quickly become computationally prohibitive. Therefore, the forward model is typically replaced by a surrogate. Since surrogate accuracy degrades as input dimension grows, the random field is compressed, e.g. using a Karhunen-Loève (KL) expansion, traditionally truncated to capture 90\% to 98\% of the spatial field variance, which may remain high dimensional. We hypothesize that, for inverse problems, fewer explicit KL inputs suffice for stable source-parameter inference because dispersion smooths fine-scale variations and sparse observations provide limited information about them. We propose an incremental inference framework that combines probabilistic Gaussian Process emulators with uncertainty-aware Bayesian inference. The emulators are trained on full-resolution model outputs, so the variability induced by omitted KL modes is represented approximately by an additional predictive variance. Adding this variance to the measurement-error in the likelihood regularizes the Bayesian update against surrogate truncation errors and helps prevent posterior overconfidence. By tracking a normalized Kullback-Leibler divergence across truncation levels, we provide a reference-free diagnostic for the stabilization of the source-parameter posteriors. In the synthetic cases considered, the posteriors stabilize with substantially fewer explicit KL inputs than traditional input-variance heuristics demand. The required number depends on what the observations can resolve and which parameters are inferred, and the framework provides a problem-specific diagnostic for identifying this truncation level.

Authors

Institutions

Publication Details

Published
2026-11-13
DOI
https://doi.org/10.22541/essoar.15012152/v1
Primary Topic
Groundwater flow and contamination studies
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

How Much Can Random-Field Dimensionality Be Reduced for Surrogate-Based Bayesian Contaminant Source Identification in Heterogeneous Aquifers?

Anneli Guthke, Maria Fernanda Morales Oreamuno, Sergey Oladyshkin, Stefania Scheurer et al.
2 links
Groundwater flow and contamination studies
preprint

How Much Can Random-Field Dimensionality Be Reduced for Surrogate-Based Bayesian Contaminant Source Identification in Heterogeneous Aquifers?

Anneli Guthke, Maria Fernanda Morales Oreamuno, Sergey Oladyshkin, Stefania Scheurer, Wolfgang Nowak
preprint en

Abstract

Contaminant source identification in heterogeneous aquifers is classically hindered by high parameter dimensions. The unknown hydraulic conductivity field must be represented by a spatially correlated random field, which introduces many unknowns alongside the source parameters. Jointly inferring them requires repeated forward-model evaluations, which quickly become computationally prohibitive. Therefore, the forward model is typically replaced by a surrogate. Since surrogate accuracy degrades as input dimension grows, the random field is compressed, e.g. using a Karhunen-Loève (KL) expansion, traditionally truncated to capture 90\% to 98\% of the spatial field variance, which may remain high dimensional. We hypothesize that, for inverse problems, fewer explicit KL inputs suffice for stable source-parameter inference because dispersion smooths fine-scale variations and sparse observations provide limited information about them. We propose an incremental inference framework that combines probabilistic Gaussian Process emulators with uncertainty-aware Bayesian inference. The emulators are trained on full-resolution model outputs, so the variability induced by omitted KL modes is represented approximately by an additional predictive variance. Adding this variance to the measurement-error in the likelihood regularizes the Bayesian update against surrogate truncation errors and helps prevent posterior overconfidence. By tracking a normalized Kullback-Leibler divergence across truncation levels, we provide a reference-free diagnostic for the stabilization of the source-parameter posteriors. In the synthetic cases considered, the posteriors stabilize with substantially fewer explicit KL inputs than traditional input-variance heuristics demand. The required number depends on what the observations can resolve and which parameters are inferred, and the framework provides a problem-specific diagnostic for identifying this truncation level.

University of Stuttgart (DE)
Deutsche Forschungsgemeinschaft, Singapore Institute of Manufacturing Technology
Groundwater flow and contamination studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon