This portfolio project demonstrates a simplified uncertainty-aware hydrogeophysical workflow using synthetic Time-Domain Electromagnetic (TEM)-type response data. The project explores how indirect geophysical signals can be translated into possible subsurface conditions, hydrological interpretation, and contaminated site decision-support information.
The workflow is intentionally developed as a synthetic and screening-level study. It is not a full electromagnetic inversion, calibrated groundwater flow model, or formal contaminated-site risk assessment.
Research problem: This project explores how simplified time-domain electromagnetic (TEM)-type responses can be interpreted under uncertainty for groundwater and contaminated-site screening.
Data: Synthetic TEM-type response curves generated from a simplified layered subsurface model. No field TEM dataset is used.
What I implemented: A Python workflow for synthetic response generation, Monte Carlo probabilistic inversion, prior-posterior uncertainty comparison, scenario testing, AI-assisted surrogate screening, and simplified groundwater-contaminant pathway interpretation.
Main technical lesson: Several subsurface parameter combinations can explain similar geophysical responses. Therefore, non-uniqueness and uncertainty must be considered before translating geophysical signals into hydrological interpretation.
Main limitation: This is a portfolio-level synthetic study, not a full electromagnetic forward model, calibrated groundwater model, or formal contaminated-site risk assessment.
Relevance: The project demonstrates early preparation for probabilistic geophysical-hydrological interpretation, uncertainty-aware subsurface characterisation, and remediation decision support.
The objectives of this project are to:
- Generate synthetic TEM-type response data for a simplified layered subsurface system.
- Apply Monte Carlo probabilistic inversion to estimate plausible subsurface parameter ranges.
- Compare prior and posterior uncertainty for key parameters.
- Evaluate scenario sensitivity for aquifer depth, conductive layer properties, and measurement noise.
- Develop AI-assisted surrogate models for rapid parameter screening.
- Translate posterior uncertainty into simplified groundwater and contaminant pathway implications.
- Demonstrate how uncertainty-aware interpretation can support conceptual site model development, monitoring design, and preliminary remediation decision support.
The project consists of six main components:
A simplified layered subsurface model was used to generate synthetic TEM-type response curves. The model includes a near-surface layer, an intermediate conductive/low-permeability layer, and a deeper aquifer-like zone.
Monte Carlo sampling was used to test many possible subsurface parameter combinations. Synthetic observations were compared with candidate model responses to identify posterior parameter ranges.
Key estimated parameters include:
- Layer 2 resistivity
- Aquifer resistivity
- Aquifer boundary depth
Prior and posterior distributions were compared to show how synthetic TEM-type observations reduce uncertainty in subsurface interpretation.
Several scenarios were tested to evaluate how uncertainty changes under different subsurface and measurement conditions:
- Baseline scenario
- Deeper aquifer scenario
- More conductive layer scenario
- Higher measurement noise scenario
Machine learning surrogate models were trained to estimate subsurface parameters from synthetic TEM-type response curves. The optimized target-specific AI surrogate models were trained using 20,000 synthetic TEM-type response curves with variable noise levels. The final optimized approach used target-specific Extra Trees regression models.
Final optimized AI surrogate performance:
| Target Parameter | R2 Score |
|---|---|
| Layer 2 resistivity | 0.752 |
| Aquifer resistivity | 0.670 |
| Aquifer boundary depth | 0.621 |
The AI surrogate models are interpreted as rapid screening tools, not replacements for physics-based or probabilistic inversion.
Posterior uncertainty ranges were translated into simplified hydrological indicators, including hydraulic conductivity classes, horizontal travel time, vertical screening time, and groundwater pathway concern categories.
Selected screening results:
| Indicator | P05 | P50 | P95 |
|---|---|---|---|
| Layer 2 hydraulic conductivity, K (m/s) | 2.15e-09 | 4.52e-08 | 7.00e-07 |
| Aquifer hydraulic conductivity, K (m/s) | 2.89e-06 | 6.09e-05 | 7.56e-04 |
| Aquifer boundary depth, z2 (m) | 12.46 | 24.66 | 33.97 |
| Horizontal travel time over 100 m (years) | 0.10 | 1.42 | 29.95 |
| Vertical screening time to aquifer (years) | 0.96 | 14.20 | 221.19 |
These results are intended to demonstrate how uncertainty-aware geophysical interpretation can support early-stage contaminated site assessment, monitoring well placement, and remediation planning.
synthetic-tem-hydrogeophysical-interpretation/
|
|-- README.md
|-- requirements.txt
|-- notebooks/
| |-- Synthetic_TEM_Hydrogeophysical_Interpretation_Project.ipynb
|
|-- data/
| |-- synthetic_TEM_type_data.csv
| |-- posterior_uncertainty_summary.csv
| |-- scenario_uncertainty_comparison.csv
| |-- final_ai_surrogate_r2_comparison.csv
| |-- groundwater_contaminant_screening_summary.csv
|
|-- figures/
| |-- figure_1_synthetic_TEM_response.png
| |-- figure_2_best_fit_response.png
| |-- prior_posterior_layer2_resistivity.png
| |-- target_specific_ai_predicted_vs_true_rho2.png
| |-- groundwater_contaminant_risk_distribution.png
|
|-- reports/
| |-- mini_project_report.md
| |-- optimized_ai_surrogate_final_summary.md
| |-- groundwater_final_interpretation.md
| |-- prior_posterior_interpretation.md
|
Important output files include:
- posterior_uncertainty_summary.csv
- scenario_uncertainty_comparison.csv
- final_ai_surrogate_r2_comparison.csv
- groundwater_contaminant_screening_summary.csv
- groundwater_pathway_concern_distribution.csv
- monitoring_remediation_implication_table.csv
- optimized_ai_surrogate_final_summary.md
- groundwater_final_interpretation.md
This project used:
- Python
- NumPy
- Pandas
- Matplotlib
- scikit-learn
- Monte Carlo sampling
- Random Forest regression
- Extra Trees regression
- Screening-level hydrogeological interpretation
To run the project locally:
git clone https://github.com/ndimassaputro/synthetic-tem-hydrogeophysical-interpretation.git
cd synthetic-tem-hydrogeophysical-interpretation
pip install -r requirements.txt
jupyter notebook notebooks/Synthetic_TEM_Hydrogeophysical_Interpretation_Project.ipynbThis project has several important limitations:
- The TEM-type response is synthetic and simplified.
- The workflow does not represent a full electromagnetic forward model.
- The inversion is a simplified Monte Carlo uncertainty screening, not a calibrated field inversion.
- Hydraulic conductivity estimates are rule-based and used only for conceptual interpretation.
- The groundwater and contaminant pathway screening is not a formal risk assessment.
- Real site application would require field TEM data, borehole control, lithological information, groundwater levels, contaminant concentration data, and calibrated hydrogeological modelling.
Possible future improvements include:
- Applying the workflow to real TEM datasets.
- Using a physics-based electromagnetic forward modelling package.
- Coupling geophysical inversion outputs with groundwater flow and contaminant transport models.
- Testing Bayesian inversion or Bayesian Evidential Learning approaches.
- Developing AI surrogate models for computational acceleration in probabilistic inversion.
- Integrating borehole, geochemical, and monitoring data for stronger conceptual site model development.
Aigner, L., Michel, H., Hermans, T. and Flores Orozco, A. (2025) ‘Stochastic inversion of transient electromagnetic data to derive aquifer geometry and associated uncertainties’, Geophysical Journal International, 242(2), ggaf236. https://doi.org/10.1093/gji/ggaf236
Auken, E., Christiansen, A.V., Jacobsen, B.H., Foged, N. and Sørensen, K.I. (2005) ‘Piecewise 1D laterally constrained inversion of resistivity data’, Geophysical Prospecting, 53(4), pp. 497–506. https://doi.org/10.1111/j.1365-2478.2005.00486.x
Bai, P., Vignoli, G. and Hansen, T.M. (2021) ‘1D stochastic inversion of airborne time-domain electromagnetic data with realistic prior and accounting for the forward modeling error’, Remote Sensing, 13(19), 3881. https://doi.org/10.3390/rs13193881
Binley, A., Hubbard, S.S., Huisman, J.A., Revil, A., Robinson, D.A., Singha, K. and Slater, L.D. (2015) ‘The emergence of hydrogeophysics for improved understanding of subsurface processes over multiple scales’, Water Resources Research, 51(6), pp. 3837–3866. https://doi.org/10.1002/2015WR017016
Puzyrev, V. (2019) ‘Deep learning electromagnetic inversion with convolutional neural networks’, Geophysical Journal International, 218(2), pp. 817–832. https://doi.org/10.1093/gji/ggz204
Puzyrev, V. and Swidinsky, A. (2021) ‘Inversion of 1D frequency- and time-domain electromagnetic data with convolutional neural networks’, Computers & Geosciences, 149, 104681. https://doi.org/10.1016/j.cageo.2020.104681
Rubin, Y. and Hubbard, S.S. (eds.) (2005) Hydrogeophysics. Dordrecht: Springer. https://doi.org/10.1007/1-4020-3102-5
Wang, L., Dai, Y., Liu, W. and Zhou, S. (2024) ‘Deep learning transient electromagnetic inversion for seawater intrusion’, Journal of Geophysics and Engineering, 21(6), pp. 1810–1821. https://doi.org/10.1093/jge/gxae107
Nurwahid Dimas Saputro
Master's student in Civil and Environmental Engineering
Ehime University, Japan