IAP-26-055

Reconstructing Groundwater Pollution Release Histories under Geological Uncertainty Using Machine Learning and Adaptive Monitoring

Groundwater contamination can persist long after the original pollution event, which threatens freshwater quality and aquatic ecosystems. A central challenge for environmental assessment is reconstructing the release histories of groundwater contaminants. In particular, distinguishing active contaminant sources from historical pollution that continues to migrate through an aquifer is often difficult when subsurface geology and transport pathways are poorly constrained. Yet, reconstructing release histories is important to design more effective monitoring and remediation schemes.
In heterogeneous alluvial aquifers, interconnected high-permeability lithofacies, low-permeability zones and multi-scale flow pathways can produce early contaminant breakthrough and long-time concentration tailing. Studies of the Macrodispersion Experiment (MADE) aquifer demonstrate the importance of sedimentary architecture and geological connectivity for such anomalous transport behaviour (Bianchi & Zheng, 2016; Bianchi & Pedretti, 2017). However, reproducing these transport signatures does not establish whether an unknown contaminant release history can be uniquely inferred from limited environmental observations. In particular, an oversimplified geological conceptual model may explain a late-time concentration tail by incorrectly inferring that pollutant input is still continuing.
Numerical source inversion, machine-learning surrogates and sequential experimental design have been used for contaminant source identification (Chang et al., 2024; Shen & Huan, 2023). However, important scientific questions remain regarding how geological connectivity and transport-model uncertainty affect the identifiability of contaminant release histories, the risk of falsely confident source attribution, and the value of additional observations.
This project will address these questions through three linked objectives:
1) Hydrogeological controls on source-history identifiability: Investigate how lithofacies connectivity, preferential flow and associated non-Fickian transport signatures govern the ability to distinguish historical, ceased releases from ongoing contaminant input.
2) ML-assisted source reconstruction under geological uncertainty: Develop and evaluate machine-learning models integrated with probabilistic inversion methods, to reconstruct contaminant release histories and assess how uncertain geological structures, limited source-location uncertainty and surrogate errors influence posterior reliability and false source attribution.
3) Adaptive monitoring for resolving source-history ambiguity: Investigate how informative monitoring locations and sampling times can be selected using Bayesian experimental design and reinforcement learning (RL), to discriminate competing source-history and geological hypotheses under monitoring constraints.
The study will use MADE-informed geological and transport models as a physically grounded test bed. Publicly available MADE geological training images and published tracer studies will inform model construction; accessible tracer observations may support validation. Hypothetical historical and ongoing pollution scenarios will be generated through controlled numerical experiments.

The research will integrate physics-based numerical modelling, probabilistic inference, scientific machine learning and adaptive monitoring design.
1. Physics-based groundwater modelling: The MADE aquifer will provide the primary geological reference, based on published lithofacies interpretations and the publicly available MADE training-image collection (Brunetti et al., 2019a, 2019b). An established groundwater flow and solute-transport simulator, MODFLOW, will be employed to investigate how geological heterogeneity, lithofacies connectivity and complex solute transport affect contaminant migration under different release histories.
2. Probabilistic source reconstruction: Bayesian and ensemble-based data assimilation methods will be employed to reconstruct unknown contaminant release histories. The effects of geological uncertainty and conceptual model mismatch on source identifiability and reconstruction reliability will be investigated.
3. Machine learning-assisted modelling and inference: Machine learning surrogate models, such as convolutional neural networks or neural operators, will be developed to accelerate repeated contaminant transport simulations and probabilistic source reconstruction.
4. Adaptive monitoring design: Bayesian Value of Information analysis will be used to identify informative monitoring locations and sampling times for reducing uncertainty in source reconstruction. Reinforcement learning (RL), potentially within a Partially Observable Markov Decision Process (POMDP) framework, will be explored to develop sequential monitoring strategies that adapt to newly acquired observations.
The proposed methods will be evaluated using controlled numerical experiments based on MADE-informed geological models, with available tracer observations and independent benchmarks used to assess transport-model reliability.

Year 1

• Review groundwater contaminant transport, MADE studies, geological heterogeneity and connectivity, and gain proficiency in MODFLOW.
• Analyse available MADE geological and tracer data, and develop a numerical benchmark.
• Generate heterogeneous geological models and simulate different contaminant release histories.
• Investigate source-history identifiability under geological uncertainty and establish baseline probabilistic inversion methods.

Year 2

• Quantify the influence of connectivity, transport tailing, geological entropy and observation density on release-history distinguishability.
• Extend Bayesian inference to source reconstruction under geological and source-location uncertainty.
• Develop ML surrogates to accelerate source reconstruction and evaluate their error propagation into source attribution and posterior calibration.
• Prepare a journal manuscript on hydrogeological identifiability and present at a national research conference.

Year 3

• Develop Bayesian monitoring strategies to reduce uncertainty in contaminant source reconstruction.
• Explore RL and POMDP frameworks for adaptive monitoring design.
• Test monitoring strategies under geological and source-location uncertainty.
• Prepare a journal manuscript on monitoring information value; present at an international conference.

Year 3.5

• Integrate the findings into practical identifiability and monitoring guidance.
• Finalise manuscripts, release suitable reproducible benchmark cases and computational workflows, and complete the PhD thesis and viva preparation.

The student will receive interdisciplinary training in hydrogeology, numerical contaminant transport, environmental data interpretation, data assimilation, inverse problems, machine learning and sequential decision-making.
IAPETUS cohort training, research seminars and opportunities for engagement with relevant environmental and geological organisations will develop broader perspectives on groundwater protection.
The student will be encouraged to present at national and international conferences, publish in peer-reviewed journals and develop transferable skills in collaborative research, project management, scientific writing and open science.