AI-Enhanced Subsurface Infrastructure Mapping on AWS

This post was contributed by Santi Adavani from S2 Labs, Souvik Mukherjee from Empact-AI, Jacques Guigne from Kraken Robotics, and Ryan Qi, Vidyasagar Ananthan, and Srinivas Tadepalli from AWS.

Subsurface infrastructure mapping is the process of identifying and visualizing buried structures like pipelines, cables, storage tanks, and foundations that exist beneath the surface without excavation. This technology is critical for urban planning, utility maintenance, oil and gas operations, construction safety, and environmental protection. Without accurate subsurface maps, construction projects risk costly delays, dangerous utility strikes, and environmental damage. When Hurricane Ivan damaged an offshore oil platform in 2004, it left critical infrastructure buried under 35-45 meters of sediment, creating an invisible hazard that traditional mapping techniques couldn’t fully detect.

Through an innovative collaboration between S2 Labs, Empact AI, and Kraken Robotics, a breakthrough in subsurface infrastructure mapping has emerged on AWS. Their groundbreaking approach combines advanced magnetic imaging with physics-informed artificial intelligence to reveal underground structures with unprecedented clarity, especially in conditions where traditional methods fall short. This fusion of cloud computing power and AI is transforming how industries visualize and understand critical underground infrastructure, opening new possibilities for subsurface exploration.

Subsurface detection methods and limitations

Traditional subsurface imaging employs a range of geophysical techniques, each suited to specific materials and conditions. For instance, electromagnetic methods detect metal pipes and cables by their conductivity, while magnetometers measure Earth’s magnetic field variations to locate ferromagnetic materials like steel pipes. Ground-penetrating radar excels at imaging concrete structures and geological formations, and specialized frequencies can detect plastic utilities and water-bearing assets through their dielectric properties.

Simple manual interpretation of the imaging data involves analyzing 2D geophysical signals and making basic depth calculations – quick but approximate. There are two fundamental challenges: First, sometimes different arrangements of underground objects can produce the same results on our detecting tools, making it impossible to determine which configuration is actually correct without additional data. Second, real-world environments contain varying soil types, moisture levels, and material densities that change over short distances, creating complex signal patterns that traditional algorithms struggle to interpret accurately.

Looking at the magnetic survey results, we can see surface measurements that reveal varying magnetic intensities across a 50-meter area, as shown in Figure 1(a). When we process this data using traditional methods, we get a rough picture of what appears to be a pipe-like structure extending about 5 meters deep, but the image is somewhat blurry and lacks detail (Figure 1(b)). This is where our AI-based approach really shines – it provides a much clearer picture, showing a distinct pipe-like structure buried between 1 and 1.5 meters below the surface (Figure 1(c)). Therefore, the AI method pinpoints the location much more precisely while still staying true to the original magnetic measurements.

Figure 1. (a) Map view across a 50m survey section. (b) Conventional least-squares inversion method. (c) Deep learning-based inversion method.

Physics-informed deep learning solution

S2 Labs applies physics-informed AI and AWS high performance computing to solve complex engineering problems in oil and gas, manufacturing, healthcare, and biotechnology sectors, delivering solutions that maintain scientific accuracy while reducing computational time. S2 Labs collaborated with two specialized partners: Empact AI, providing 3D subsurface pipeline mapping, and Kraken Robotics, contributing high-resolution underwater imaging through their Synthetic Aperture Sonar systems. The collaboration integrates advanced sonar imaging, 3D subsurface analysis, and AI-driven pattern recognition using the AWS Cloud to identify and locate pipeline leak sources with enhanced accuracy and speed.

Our AI method combines the physics of magnetic fields with deep learning to better understand what’s buried underground. By teaching the AI with simulated data based on real-world structures like storage tanks and pipelines, we can train it to ‘read’ magnetic field measurements like a map. Using a special type of neural network called U-Net, the AI learns to translate these magnetic readings into clear pictures of underground structures, telling us not just where they are but also what they’re made of and what shape they take. If you’re interested in the technical nitty-gritty, check out the recent publication led by S2 Labs.

Model training

The physics-informed deep learning model was trained on AWS using a combination of high performance computing resources, data storage systems, and parallel processing services, as illustrated in the architecture diagram in Figure 2.

Using Amazon EC2 instances, we generated 202,000 3D susceptibility models (226,000 cells each) representing various underground scenarios – including pipes with different orientations, multiple pipe configurations, and storage tanks.

Models were then parameterized based on domain expertise and stored as NumPy files in Amazon S3 buckets. S2 Labs’s proprietary magnetostatic solver application was containerized and stored in Amazon ECR for consistent deployment across compute resources. The solver processed models sequentially from S3, saving responses back to S3.

We also implemented distributed computing using AWS Batch for data generation with Spot Instances for cost optimization. We used P4d instances, each of which provides eight NVIDIA A100 GPUs to compute magnetic responses across 1,800 measurement points at 2-meter intervals. The pipeline synchronized data between Amazon S3 and local storage, training a 2D U-Net architecture (500M parameters) over 110 epochs, achieving training and validation losses of 0.0018 and 0.0019. The computation required 100,000 CPU hours.

Figure 2. Architecture diagram for synthetic data generation and model training on AWS.

Workflow for inferencing large-scale magnetic surveys

Our magnetic survey processing pipeline employs a systematic four-stage workflow to handle large-scale surveys efficiently while maintaining high-quality reconstructions of subsurface structures, as shown in figure 3.

Stage 1 – Data Acquisition: Field data collection begins with magnetometer systems tailored to the survey environment – drone-mounted for aerial surveys, ground-based for terrestrial mapping, or underwater systems for marine applications. Surveys follow systematic grid patterns with consistent sensor heights and line spacing to ensure uniform coverage across the target area.

Stage 2 – Survey Domain Preparation: Instead of processing the entire survey area at once, we implement a modular approach by dividing the survey domain into smaller patches matched to the AI model’s training dimensions. Adjacent patches share overlapping regions, crucial for ensuring smooth transitions in the final reconstruction and avoiding edge artifacts.

Stage 3 – Parallel Processing Architecture: The workflow leverages parallel computing to process multiple patches simultaneously, significantly reducing computation time while maintaining consistency with the trained model’s parameters. This distributed approach makes efficient use of available computing resources through independent patch processing. As an example, our implementation can handle 400 m X 400 m X 60 m survey data in < 5 seconds.

Stage 4 – AI-Based Inference: Our trained AI model performs inference on each patch independently, reconstructing the subsurface magnetic susceptibility distribution from magnetic field measurements. The reconstructions are then seamlessly combined using weighted blending in overlapping regions, ensuring smooth transitions between adjacent patches for a cohesive final result. This modular workflow enables scalability for surveys of any size while maintaining consistent resolution and optimizing memory usage through efficient parallel processing, making it practical for real-world applications from infrastructure mapping to geological exploration.

Figure 3. Modular processing workflow for large-scale magnetic surveys.

Case study: detection of underwater oil & gas well conductors in Gulf of Mexico

Hurricane Ivan (2004) damaged an offshore oil platform in the Gulf of Mexico, burying well conductors under 35-45 meters of sediment. Initial acoustic imaging in 2022, while partially successful, was limited by gas-filled sediments masking critical areas. A high-resolution magnetometer array was deployed 3.5 meters above the seafloor to detect iron-rich conductors through hydrocarbon-saturated sediments.

The model we described in the previous sections successfully mapped buried conductors at 35-45m depth, revealing a main conductor bundle and a secondary debris segment 40m northeast of the well bay (as shown in figure 4). Results demonstrated superior differentiation of magnetic signatures in complex debris fields, verified against drilling points and acoustic imaging where available. This proved deep learning’s effectiveness where traditional acoustic methods fail.

Figure 4. Plan view (a) and oblique view (b) of relative susceptibility distribution.

Conclusions

Our work demonstrates how AI-enhanced magnetic imaging transforms subsurface infrastructure mapping across multiple sectors, from shallow onshore utilities to deep offshore well conductors. The physics-informed deep learning model was trained on AWS using a combination of high performance computing resources, data storage systems, and parallel processing services. Through real-world case studies, we’ve proven that deep learning can overcome traditional magnetic data interpretation limitations, successfully mapping structures 40m below seafloor that remained invisible to acoustic methods for 18 years. The technology’s impact spans oil & gas decommissioning, urban utility mapping, environmental protection, and marine operations. While these results are promising, further development opportunities include multi-physics integration, real-time processing, and enhanced resolution capabilities. For collaboration opportunities or implementation inquiries, please contact us at santi@s2labs.co or ryanqi@amazon.com.