Tracking groundwater pollution with a two-speed digital twin

Formal statement

If a non-linear multi-fidelity Gaussian process surrogate (Kennedy–O’Hagan AR1) is used to approximate a reactive transport model within a Bayesian inverse framework, then the 95% credible interval width for source-attribution parameters (e.g., contaminant release rate, reaction rate) will be reduced by ≥40% relative to single-fidelity inversion at equal computational budget, because the low-fidelity level resolves the posterior’s global structure while the high-fidelity level corrects local bias; and the expected information gain (EIG) computed on the surrogate will correlate with the high-fidelity EIG with Pearson r ≥ 0.85 across candidate monitoring designs.

Title given by the sharpener: Multi-Fidelity Bayesian Inverse Analysis for Reactive Transport in Groundwater: Source Attribution and Information-Optimal Monitoring

Counter-evidence

1. This paper already demonstrates that single-fidelity surrogate-accelerated Bayesian inference works for reactive transport parameter estimation, suggesting the core methodological claim of SPORE (surrogate + Bayesian inversion for reactive transport) is not novel — only the multi-fidelity extension and OED components are.

   Severity seriousAn adaptive Kriging surrogate method for efficient joint estimation of hydraulic and biochemical parameters in reactive transport modeling.

2. Demonstrates that ML surrogates for reactive transport in groundwater are already established practice, reducing the novelty of the surrogate component of SPORE.

   Severity minorIntegrating Process‐Based Reactive Transport Modeling and Machine Learning for Electrokinetic Remediation of Contaminated Groundwater

3. Shows that source attribution in groundwater is currently done with deterministic/statistical methods (SOM, PMF) that practitioners accept — suggesting the Bayesian probabilistic alternative may face adoption barriers, though this is not a scientific contradiction.

   Severity addressableDriving factor, source identification, and health risk of PFAS contamination in groundwater based on the self-organizing map.

Unknowns and boundary conditions

Known unknowns

• Whether the discrepancy term remains identifiable at N_high = 20 when the low-fidelity model has a rank inversion (i.e., the low-fidelity model is more accurate in some regions than the high-fidelity model).
• The magnitude of the outer-loop bias in EIG estimation when using a surrogate, and whether it can be corrected by a control variate or importance sampling.
• The impact of spatially correlated observation errors on the predictive coverage of the multi-fidelity posterior.
• Whether the multi-fidelity GP can handle structural model error (e.g., omitted reaction pathways) without overfitting the discrepancy term.

Boundary conditions

• The low-fidelity model must have a rank correlation with the high-fidelity model of at least 0.5 across the parameter space.Rationale: Below this threshold, the multi-fidelity GP cannot effectively transfer information, and the discrepancy term dominates.
• The number of high-fidelity simulations N_high must be at least 10 to train the AR1 GP.Rationale: With fewer than 10 high-fidelity points, the GP hyperparameters are not identifiable.
• Observation errors after log-transformation must have a correlation length ℓ_obs ≤ 10 km.Rationale: Beyond this length, the spatial correlation is indistinguishable from a constant bias, and the likelihood becomes non-identifiable.
• The reactive transport model must include at least the primary reaction pathway; omission of the primary pathway invalidates the surrogate.Rationale: The multi-fidelity GP cannot correct for a structurally wrong high-fidelity model if the error is not smooth.

Proposed mechanism

Causal chain

1. Step 1: Construct a low-fidelity surrogate (e.g., coarse-grid advection-dispersion or reduced-order model) that captures the global posterior structure of reactive transport parameters at low computational cost.
2. Step 2: Train a non-linear multi-fidelity GP (AR1 chaining) on N_low low-fidelity and N_high high-fidelity simulations, with a discrepancy term δ(x) modeled as a sparse GP with an informative prior.
3. Step 3: Perform Bayesian inversion using the multi-fidelity GP as a surrogate for the likelihood, sampling the posterior with MCMC or variational inference.
4. Step 4: Compute the expected information gain (EIG) for candidate monitoring designs using the surrogate, and validate against nested Monte Carlo on the high-fidelity model.
5. Step 5: Quantify the reduction in credible interval width relative to single-fidelity inversion at equal computational budget, and assess the robustness of the discrepancy term to rank inversion between fidelity levels.

Key assumptions

• The low-fidelity model is correlated with the high-fidelity model with ρ ≥ 0.5 across the parameter space.
• The discrepancy term δ(x) is smooth and can be represented by a sparse GP with a Matérn kernel.
• Observation errors after log-transformation are spatially correlated and heteroscedastic, modeled by a Matérn covariance with unknown correlation length and variance.
• The high-fidelity reactive transport model is structurally correct for the synthetic scenarios used in Phase 1 (no missing reaction pathways).
• The EIG computed on the surrogate is an unbiased estimator of the high-fidelity EIG when the surrogate is well-calibrated.

Theoretical framework

Multi-fidelity Bayesian inverse analysis with Gaussian process surrogates (Kennedy & O’Hagan 2000; Perdikaris et al. 2017) and expected information gain for optimal experimental design (Huan et al. 2024).

Variables

Falsifiable predictions

1. 1

   Multi-fidelity Bayesian inversion reduces the 95% credible interval width for the source release rate by ≥40% compared to single-fidelity inversion at equal computational budget (N_high = 20, N_low = 1000).

   Quantitative bound

   Median reduction in CI width ≥ 40% across 100 synthetic scenarios, with 95% confidence interval of the reduction [35%, 45%].

   Measurement method

   Compute CI widths from posterior samples for both methods; use paired bootstrap to estimate the reduction and its confidence interval.Statistical test Paired Wilcoxon signed-rank test, alpha = 0.05, power = 0.80, with a priori power analysis indicating 100 scenarios are needed to detect a 40% reduction with SD = 15%.

   Null hypothesis

   H0: The median reduction in CI width is ≤ 0% (no improvement).

2. 2

   The discrepancy term δ(x) in the multi-fidelity GP is identifiable at N_high = 20, as measured by a profile likelihood curvature > 0.1 (on a normalized scale) for all parameters.

   Quantitative bound

   Profile likelihood curvature > 0.1 for each discrepancy parameter in 90% of 100 synthetic scenarios.

   Measurement method

   Compute profile likelihoods for each discrepancy parameter by fixing it and optimizing the others; measure the curvature at the MLE.Statistical test Binomial test for proportion of scenarios with curvature > 0.1, alpha = 0.05.

   Null hypothesis

   H0: The profile likelihood curvature is ≤ 0.1 for at least one discrepancy parameter.

3. 3

   The Pearson correlation between surrogate-computed EIG and high-fidelity EIG (via nested Monte Carlo) is ≥ 0.85 across 50 candidate monitoring designs.

   Quantitative bound

   Pearson r ≥ 0.85 with 95% CI [0.80, 0.90].

   Measurement method

   Compute EIG on surrogate and on high-fidelity model for each design; calculate Pearson correlation and bootstrap CI.Statistical test Fisher z-transformation test for correlation, alpha = 0.05.

   Null hypothesis

   H0: Pearson r ≤ 0.70.

4. 4

   The 95% predictive intervals for observed concentrations have coverage between 93% and 97% when observation errors are spatially correlated with correlation length ℓ_obs = 5 km.

   Quantitative bound

   Empirical coverage in [93%, 97%] across 1000 test points.

   Measurement method

   Generate synthetic observations with correlated errors; compute the proportion of test points falling within the 95% predictive interval.Statistical test Binomial test for coverage, alpha = 0.05.

   Null hypothesis

   H0: Coverage is outside [93%, 97%].

5. 5

   The multi-fidelity GP maintains a posterior mean error for the reaction rate constant below 10% of the true value when the high-fidelity model omits a secondary reaction pathway (structural error).

   Quantitative bound

   Relative error ≤ 10% in 80% of 50 scenarios with omitted reaction.

   Measurement method

   Compare posterior mean to true reaction rate constant in synthetic scenarios with a missing reaction.Statistical test Binomial test, alpha = 0.05.

   Null hypothesis

   H0: Relative error > 10% in more than 20% of scenarios.

Experimental protocol

References

18 references, all from Semantic Scholar. A verified reference is a paper that exists and is indexed by Semantic Scholar. It does not mean that the paper confirms the idea.

1. 1
   Jun Zhou, Xiao-Si Su, G. Cui (2018). An adaptive Kriging surrogate method for efficient joint estimation of hydraulic and biochemical parameters in reactive transport modeling..direct support · 23 citations · doi:10.1016/j.jconhyd.2018.08.005What the librarian takes from it Adaptive Kriging-based MCMC achieves accurate Bayesian inference with a hundredfold reduction in computational cost compared to conventional MCMC for reactive transport calibration.Relevance Directly demonstrates that surrogate-accelerated Bayesian inference (MCMC) is feasible and efficient for reactive transport parameter estimation in groundwater — the core methodological claim of SPORE, though limited to single-fidelity surrogates.
2. 2
   R. Sprocati, Massimo Rolle (2021). Integrating Process‐Based Reactive Transport Modeling and Machine Learning for Electrokinetic Remediation of Contaminated Groundwater.direct support · 40 citations · doi:10.1029/2021WR029959What the librarian takes from it ANN response surface surrogates trained on limited process-based reactive transport simulations accurately predict complex subsurface system evolution, overcoming runtime restrictions.Relevance Confirms that ML surrogate models trained on a limited number of expensive reactive transport simulations can predict complex subsurface contaminant evolution — validating the low-fidelity surrogate premise of SPORE.
3. 3
   Kislaya Ravi, Vladyslav Fediukov, Felix Dietrich et al. (2024). Multi-fidelity Gaussian process surrogate modeling for regression problems in physics.direct support · 24 citations · doi:10.1088/2632-2153/ad7ad5What the librarian takes from it Multi-fidelity GP surrogates effectively chain models of increasing fidelity and cost, addressing limited data availability in computationally expensive physics simulations.Relevance Provides the multi-fidelity GP surrogate methodology (non-linear autoregressive chaining of fidelity levels) that SPORE proposes to transfer to groundwater hydrology.
4. 4
   Ke Li, Fan Li (2024). Multi-Fidelity Methods for Optimization: A Survey.indirect support · 32 citations · doi:10.1145/3801959What the librarian takes from it Multi-fidelity optimization balances high-fidelity accuracy with computational efficiency through hierarchical fidelity approaches.Relevance Systematic survey of multi-fidelity surrogate models, fidelity management, and optimization — provides the general framework SPORE claims to transfer, but not in a Bayesian inverse or groundwater context.
5. 5
   Xun Huan, Jayanth Jagalur, Youssef M. Marzouk (2024). Optimal experimental design: Formulations and computations.indirect support · 126 citations · doi:10.1017/S0962492924000023What the librarian takes from it Systematic survey of modern OED from classical design theory to complex-model applications.Relevance Provides the OED formalism (expected information gain criteria) that SPORE proposes to apply for monitoring network design, but does not address groundwater or multi-fidelity surrogates.
6. 6
   A. Attia, E. Constantinescu (2020). Optimal Experimental Design for Inverse Problems in the Presence of Observation Correlations.indirect support · 23 citations · doi:10.1137/21m1418666What the librarian takes from it General OED formulation for large-scale Bayesian linear inverse problems accommodating correlated measurement errors via weighted likelihood.Relevance Addresses OED for Bayesian linear inverse problems with correlated observation errors — relevant to SPORE’s goal of fusing heterogeneous groundwater data with non-Gaussian/correlated errors, though not applied to hydrology.
7. 7
   W. Nowak, A. Guthke (2016). Entropy-Based Experimental Design for Optimal Model Discrimination in the Geosciences.direct support · 35 citations · doi:10.3390/E18110409What the librarian takes from it Model choice indicators with Shannon entropy enable optimal experimental design for Bayesian model discrimination in geosciences.Relevance Demonstrates entropy-based OED for model selection in geosciences — directly supports SPORE’s claim that information-gain-driven design is applicable in Earth science contexts.
8. 8
   Y. Gan, Xin-Zhong Liang, Q. Duan et al. (2018). A systematic assessment and reduction of parametric uncertainties for a distributed hydrological model.direct support · 32 citations · doi:10.1016/J.JHYDROL.2018.07.055What the librarian takes from it Adaptive surrogate-based multi-objective optimization facilitates practical assessment and reduction of parametric uncertainties in distributed hydrological models.Relevance Combines sensitivity analysis with adaptive surrogate-based multi-objective optimization for hydrological model parameter uncertainty — supports the surrogate-accelerated UQ premise in a hydrological (though not reactive transport) context.
9. 9
   Rui Xu, Dongxiao Zhang, Nanzhe Wang (2021). Uncertainty quantification and inverse modeling for subsurface flow in 3D heterogeneous formations using a theory-guided convolutional encoder-decoder network.direct support · 26 citations · doi:10.1016/j.jhydrol.2022.128321What the librarian takes from it Theory-guided convolutional encoder-decoder surrogates provide efficient pressure estimation and inverse modeling for 3D heterogeneous subsurface formations.Relevance Demonstrates surrogate-based UQ and inverse modeling for 3D subsurface flow — supports the transferability of surrogate-based Bayesian inversion to subsurface hydrology, though for single-phase flow rather than reactive transport.
10. 10
    Liu Yang, Xuhui Meng, G. Karniadakis (2020). B-PINNs: Bayesian Physics-Informed Neural Networks for Forward and Inverse PDE Problems with Noisy Data.support by analogy · 1,235 citations · doi:10.1016/j.jcp.2020.109913What the librarian takes from it B-PINNs combine physical laws and scattered noisy measurements to provide accurate predictions with quantified uncertainty in inverse PDE problems.Relevance Demonstrates Bayesian inversion of PDEs with noisy scattered data — structurally analogous to SPORE’s problem, but uses PINNs rather than multi-fidelity surrogates and is not applied to groundwater.
11. 11
    François Monard, Richard Nickl, G. Paternain (2020). Statistical guarantees for Bayesian uncertainty quantification in nonlinear inverse problems with Gaussian process priors.support by analogy · 51 citations · doi:10.1214/21-aos2082What the librarian takes from it Semi-parametric Bernstein-von Mises theorem shows posterior distributions concentrate around efficient estimators in nonlinear inverse regression models.Relevance Provides theoretical foundations (Bernstein-von Mises) for Bayesian UQ in nonlinear inverse problems — supports the mathematical validity of SPORE’s Bayesian framework, though not specific to multi-fidelity or hydrology.
12. 12
    I. Sahin, Christian Moya, Amirhossein Mollaali et al. (2023). Deep Operator Learning-based Surrogate Models with Uncertainty Quantification for Optimizing Internal Cooling Channel Rib Profiles.support by analogy · 37 citations · doi:10.48550/arXiv.2306.00810What the librarian takes from it Bayesian DeepONets provide surrogate models with uncertainty quantification for optimizing engineering designs.Relevance Demonstrates Bayesian DeepONet surrogates with UQ for optimization in computational physics — analogous to SPORE’s surrogate-with-UQ approach, but in thermal engineering rather than hydrology.
13. 13
    Marjuka Ferdousi Lazin, Christian R Shelton, Simon N. Sandhofer et al. (2023). High-dimensional multi-fidelity Bayesian optimization for quantum control.support by analogy · 23 citations · doi:10.1088/2632-2153/ad0100What the librarian takes from it Multi-fidelity Bayesian optimization efficiently solves inverse problems in quantum control, outperforming gradient-based approaches.Relevance Demonstrates multi-fidelity Bayesian optimization for inverse problems in quantum control — analogous methodological transfer of multi-fidelity BO to a new domain, supporting SPORE’s transfer claim.
14. 14
    Jing-Wen Zeng, Kai Liu, Xiao Liu et al. (2024). Driving factor, source identification, and health risk of PFAS contamination in groundwater based on the self-organizing map..indirect support · 55 citations · doi:10.1016/j.watres.2024.122458What the librarian takes from it Spatial response analysis combining SOM, K-means, Spearman correlation, PMF and risk quotient reveals spatial characteristics, driving factors, and sources of PFAS in groundwater.Relevance Demonstrates current practice for groundwater contaminant source attribution using SOM, K-means, PMF — deterministic/statistical methods that SPORE aims to replace with probabilistic Bayesian source attribution.
15. 15
    Xiao Yang, Jiayi Du, Chao Jia et al. (2024). Unravelling integrated groundwater management in pollution-prone agricultural cities: A synergistic approach combining probabilistic risk, source apportionment and artificial intelligence..indirect support · 22 citations · doi:10.1016/j.jhazmat.2024.136514What the librarian takes from it Probabilistic risk assessment combined with source apportionment and AI reveals contaminant sources and health impacts in agricultural groundwater.Relevance Combines probabilistic risk, source apportionment, and AI for groundwater management — shows the demand for probabilistic source attribution in Domain B, but uses statistical/AI methods rather than multi-fidelity Bayesian inversion.
16. 16
    Ziyue Yin, Jian-Feng Wu, Jian Song et al. (2022). Multi-objective optimization-based reactive nitrogen transport modeling for the water-environment-agriculture nexus in a basin-scale coastal aquifer..indirect support · 33 citations · doi:10.1016/j.watres.2022.118111What the librarian takes from it Integrated multi-objective simulation-optimization framework evaluates water-environment-agriculture nexus using coupled variable-density groundwater and reactive transport models.Relevance Couples SEAWAT and RT3D for reactive transport simulation-optimization at basin scale — demonstrates the computational expense of reactive transport models that SPORE aims to address with multi-fidelity surrogates.
17. 17
    Suraj Kumar, N. S. Maurya (2025). Analysis of heavy metal contamination in groundwater and associated probabilistic human health risk assessment using Monte Carlo simulation: A case study in Gaya, Bihar..indirect support · 20 citations · doi:10.2166/wh.2025.348What the librarian takes from it Monte Carlo simulation reveals non-carcinogenic and carcinogenic health risks from heavy metals in groundwater, with PCA suggesting geogenic sources.Relevance Uses Monte Carlo simulation for probabilistic health risk from groundwater heavy metals — demonstrates probabilistic methods in Domain B but without Bayesian inversion or multi-fidelity surrogates.
18. 18
    Hongxia Hu, Hongguang Zheng, Feng-Ping Liu et al. (2024). Heavy Metal Contamination Assessment and Source Attribution in the Vicinity of an Iron Slag Pile in Hechi, China: Integrating Multi-Medium Analysis..indirect support · 21 citations · doi:10.1016/j.envres.2024.120206What the librarian takes from it Nemerow pollution index indicates severe heavy metal pollution across multiple media near an iron slag pile.Relevance Multi-medium contamination assessment and source attribution — demonstrates the data heterogeneity challenge (water, sediment, soil, crops) that SPORE aims to fuse within a single Bayesian framework.

Novelty

Novelty score: 0.72 out of 1 · Verdict: incremental

This score is given by an agent on the basis of the work it found. It is an estimate, not a measurement. How this score is produced

Closest existing work

• An adaptive Kriging surrogate method for efficient joint estimation of hydraulic and biochemical parameters in reactive transport modeling. (2018)Difference from the idea: This paper already implements Bayesian inference (MCMC) with a Kriging surrogate for reactive transport parameter calibration, achieving ~100x speedup. The SPORE hypothesis differs by proposing multi-fidelity (not single-fidelity surrogate) Bayesian inversion, explicit fusion of heterogeneous data types, and extension to optimal monitoring network design via expected information gain — none of which are addressed here.
• Integrating Process‐Based Reactive Transport Modeling and Machine Learning for Electrokinetic Remediation of Contaminated Groundwater (2021)Difference from the idea: Demonstrates ML surrogates trained on reactive transport outputs for groundwater remediation, but uses deterministic response surfaces (ANN) without Bayesian uncertainty quantification, multi-fidelity hierarchy, or OED. SPORE adds probabilistic inversion and information-gain-driven design.
• Multi-fidelity Gaussian process surrogate modeling for regression problems in physics (2024)Difference from the idea: Provides the multi-fidelity GP methodology in a physics context but does not address groundwater hydrology, reactive transport, source attribution, or monitoring network design. SPORE transfers this methodology to Domain B.
• Optimal experimental design: Formulations and computations (2024)Difference from the idea: Comprehensive OED survey covering expected information gain formulations, but not applied to groundwater quality monitoring networks nor coupled with multi-fidelity surrogates for reactive transport.

Gaps and data

Gaps identified

• No paper in the list demonstrates multi-fidelity Bayesian inversion specifically for reactive transport with geochemical speciation — the closest (930a0e66) uses single-fidelity Kriging surrogates.
• No paper in the list addresses fusion of geophysical, geochemical, and remote sensing data within a single Bayesian framework for groundwater quality.
• No paper in the list demonstrates expected-information-gain-based optimal design of groundwater quality monitoring networks coupled with multi-fidelity surrogates.
• No paper in the list provides statistical guarantees (e.g., posterior contraction rates) for multi-fidelity Bayesian inversion in the presence of model discrepancy between fidelity levels.

Available data

• No specific multi-fidelity groundwater quality dataset identified in the provided papers. The PFAS study (c0173ec6) and heavy metal studies (68fedc15, 33d0000a) provide heterogeneous water quality observations but not paired multi-fidelity measurements.

Panel synthesis

Consensus score: 6.16/10 Average of the five scores, weighted by the confidence each reviewer declares.

Meta-reviewer’s verdict: publish

Points of agreement

• The protocol is structured into three phases with GO/NO-GO criteria and quantitative thresholds, which favours reproducibility and limits post-hoc decisions.
• The use of the EIG on a multi-fidelity surrogate for the optimisation of monitoring designs is a credible and original applied objective, provided that the surrogate bias is controlled.
• The predictions are falsifiable with explicit numerical bounds (40% reduction in CI, r ≥ 0.85), which facilitates independent evaluation.

Points of disagreement

• The methodologist and the domain_expert consider the absence of a power analysis and of negative controls to be a major weakness, whereas the funding_strategist judges the protocol to be sufficiently rigorous for funding.
• The contrarian asserts that the hypothesis of a correlation ρ ≥ 0.5 is neither demonstrated nor tested for non-linear reactive transport, whereas the domain_expert regards it as plausible but not theoretically supported.
• The industrialist considers the market to be promising but that regulatory adoption and the conservatism of engineering consultancies limit the impact, whereas the funding_strategist sees strong alignment with the priorities of the Green Deal.

Critical path

The empirical demonstration that the rank correlation ρ ≥ 0.5 holds for a non-linear reactive transport model with differing numerical stiffnesses, and that the discrepancy term δ(x) remains identifiable at N_high = 20 without producing overconfident credible intervals.

Final recommendation

The panel recognises the originality and relevance of the multi-fidelity–Bayesian inversion–EIG coupling for source attribution in reactive transport, as well as the quality of the phase structuring. However, major methodological obstacles remain: absence of a power analysis, potential non-identifiability of δ(x), unquantified bias of the EIG on surrogate, and risk of overconfidence. As it stands, the hypothesis remains a plausible but untested conjecture, and predictions 1, 3 and 5 are at high risk of false positives. The panel recommends rejecting the current version and encourages a subsequent reformulation incorporating the missing controls and analyses.