Bayesian methods for the identification of model parameters for water transport in porous media

Paola Stolfi, E. Onofri, Gabriella BrettiarXiv 2026 · DOI · arXiv

TL;DR

This paper tackles the challenge of inferring hydraulic parameters for water transport in porous media, a problem made difficult by complex, nonlinear physics and sparse data. Using a novel application of Approximate Bayesian Computation with Sequential Monte Carlo (ABC-SMC), the authors not only accurately recover model parameters but also reveal the inherent "identifiability structure" of the problem, showing how certain parameter combinations are more or less discernible from the data.

What Problem Does This Paper Address?

Understanding how water moves through porous materials is crucial for fields ranging from civil engineering and environmental science to the preservation of historical artifacts. This movement is often governed by complex, nonlinear partial differential equations (PDEs) that describe how water saturates a material over time. The challenge arises when we try to determine the specific material properties (like hydraulic parameters) that control this water transport. This is an "inverse problem": instead of predicting water movement from known properties, we're trying to infer properties from observed water movement.

The core difficulty lies in the fact that the observed data (typically, the total amount of water absorbed over time) is a "sparse integral measurement." This means it's a simplified, averaged view of a complex, spatially varying process. The nonlinear nature of the governing equations, combined with this averaging, leads to an "ill-posed" inverse problem. In simpler terms, different combinations of material properties can produce very similar observed water absorption patterns, making it hard to pinpoint the true values. This phenomenon is known as anisotropic parameter sensitivity and structured correlations, which can trap traditional calibration methods.

Methods

The authors employ Approximate Bayesian Computation with Sequential Monte Carlo (ABC-SMC) as their primary inferential framework. This is a "likelihood-free" method, meaning it doesn't require an explicit mathematical formula for the probability of observing the data given the parameters (the likelihood function), which is often intractable for complex models like PDEs. Instead, ABC-SMC works by:

  1. Simulating: Generating synthetic data from the model using candidate parameter values drawn from a prior distribution.
  2. Comparing: Measuring the "distance" between the simulated data and the observed data using summary statistics.
  3. Iterating: Progressively refining the parameter estimates by adapting a tolerance threshold and using a population of "particles" (parameter sets) that evolve over generations.

What's novel here is not just the application of ABC-SMC to this specific problem, but its use as an analytical tool. The authors demonstrate that the geometry of the posterior distribution (how the probability is spread across parameter values) produced by ABC-SMC directly encodes the identifiability structure of the inverse problem. They use Principal Component Analysis (PCA) on the posterior distribution to quantify this structure, identifying "principal components" that represent the most and least informative directions in the parameter space.

Datasets

The paper uses two types of datasets:

  • Synthetic Data: Generated from known parameter values for both the Natalini-Nitsch (NN) and BkP diffusivity models. This allows for a controlled validation of the parameter recovery and identifiability analysis.
  • Real Laboratory Data:
    • Imbibition experiments on common brick specimens (from a previous study [6]).
  • Imbibition experiments on Ajarte limestone specimens (from a previous study [7]).

These real-world datasets are used to demonstrate the practical applicability of the ABC-SMC framework and the resulting identifiability insights on materials relevant to cultural heritage.

Key Insights and Ideas

  • Posterior Geometry as a Diagnostic: The shape, spread, and correlations within the posterior distribution obtained from ABC-SMC are not just a measure of uncertainty but a direct reflection of the underlying identifiability of the model parameters. This is a powerful insight that goes beyond simple parameter estimation.
  • PCA Reveals Hidden Structure: Applying PCA to the posterior distribution provides a principled and computationally efficient way to understand parameter sensitivity. Low-variance principal components highlight parameter combinations that are difficult to distinguish from the data, revealing "near-invariance directions."
  • Porosity's Dual Role: The porosity parameter ($n_0$) exhibits a unique identifiability behavior due to its dual role as a structural parameter and a scaling factor for the total absorbed water. This leads to an approximate "scaling symmetry" in the forward model, creating posterior ridges rather than sharp point estimates.
  • Material Heterogeneity Matters: The Ajarte limestone, being more heterogeneous than brick, leads to a more ill-posed inverse problem, reflected in wider marginal posteriors and stronger correlations between parameters.
  • ABC-SMC is Robust: The adaptive ABC-SMC algorithm proves to be numerically stable and effective even for complex, nonlinear PDE models and under conditions of higher model uncertainty.

Key Results

  • Accurate Parameter Recovery: Both the NN and BkP models' parameters were accurately recovered in synthetic experiments and within acceptable confidence intervals for real-world brick and Ajarte limestone data.
  • Identifiability Structure Revealed: The posterior distributions clearly showed anisotropic parameter sensitivity and structured correlations, particularly for parameters defining the boundaries of the absorption function (e.g., $a$ and $b$) and compensatory relationships (e.g., between $c$ and $k_{log}$ in the NN model).
  • PCA Quantifies Sensitivity: For common brick, the first two principal components captured over 50% of the posterior variance, indicating a relatively low-dimensional identifiable subspace. For Ajarte, PC1 accounted for ~40%, with subsequent components contributing comparably, suggesting a similar global dimensionality but potentially more complex material-dependent conditioning.
  • Porosity Uncertainty: The porosity parameter ($n_0$) consistently showed wider posterior distributions, especially for Ajarte, highlighting its partial identifiability.
  • Model Performance: Both the NN and BkP models, when calibrated using ABC-SMC, provided excellent fits to the imbibition curves for both brick and Ajarte, demonstrating the effectiveness of the chosen models and the calibration procedure.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the spatial integration of the observable (water absorption curve) inherently limits the information available to discriminate between certain parameter combinations.
  • For Ajarte limestone, increased parameter uncertainty was reported, directly linked to the material's higher structural heterogeneity and the more ill-posed inverse problem.
  • Future Work:
    • Spatially Resolved Data: The authors suggest integrating spatially resolved moisture profiles (e.g., from neutron radiography) to break the near-invariance directions caused by spatial integration and reduce posterior degeneracy.
  • Formal Model Comparison: Extending the ABC-SMC framework to approximate Bayes factors would allow for principled model selection between different parameterizations (like NN and BkP) based on their posterior evidence.
  • Experimental Design Guidance: The insights from posterior geometry and PCA can directly guide future experimental design by identifying which observational channels would be most effective in reducing uncertainty for specific parameter combinations.

Summary updated

Abstract

The structure of the nonlinear inverse problem arising from capillarity-driven imbibition in porous media is investigated, considering a degenerate parabolic PDE with compactly supported diffusivity and boundary-driven fluxes as the governing forward model. The inverse problem -- inferring hydraulic model parameters from sparse integral absorption measurements -- is inherently ill-posed: the nonlinear forward operator induces anisotropic parameter sensitivity and structured correlations that render the calibration landscape non-convex and partially unidentifiable. To characterise this structure rigorously, Approximate Bayesian Computation with Sequential Monte Carlo (ABC-SMC) is adopted as a likelihood-free inferential framework, bypassing the analytical intractability of the likelihood while providing full posterior distributions over the parameter space. Two physically motivated parameterisations of the diffusivity function are analysed -- the Natalini-Nitsch (NN) and the BkP formulations. It is shown that the posterior geometry obtained via ABC-SMC encodes, in directly readable form, the sensitivity structure of the nonlinear forward operator: the principal component decomposition of the posterior covariance provides a natural hierarchy of parameter sensitivity, with low-variance eigendirections identifying the parameter combinations to which the forward map is most responsive. This geometric decomposition constitutes a principled and computationally efficient alternative to classical sensitivity analysis, arising as a byproduct of the calibration procedure. These findings are established through both synthetic experiments, confirming accurate parameter recovery, and real laboratory imbibition data from materials of cultural heritage relevance.
BibTeX
@misc{stolfi2026bayesian,
  author = {Paola Stolfi and E. Onofri and Gabriella Bretti},
  title = {{Bayesian methods for the identification of model parameters for water transport in porous media}},
  year = {2026},
  eprint = {2603.29859},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.29859},
}
Figure