Optimizing Groundwater Flow Modeling with AI

Modeling groundwater flow in three-dimensional fractured crystalline media presents a significant computational challenge. The strong spatial heterogeneity induced by fractures necessitates detailed simulations, known as Discrete Fracture-Matrix (DFM) models. While accurate, these simulations are extremely expensive in terms of computational resources, especially when repeated evaluations are required. This high cost limits the applicability of such models in scenarios demanding rapid analysis or frequent iterations.

To address this issue, research is moving towards adopting innovative methodologies that leverage artificial intelligence. The goal is to employ frameworks like Multilevel Monte Carlo (MLMC), where numerical homogenization is used for upscaling sub-resolution fracture effects, facilitating the transition between different accuracy levels and reducing the overall computational load.

The Surrogate Model Architecture and its Performance

Central to this optimization strategy is the development of a surrogate model, designed to predict the equivalent hydraulic conductivity tensor (Keq). This model operates on a voxelized 3D domain, representing tensor-valued random fields of matrix and fracture conductivities. Fracture characteristics, such as size, orientation, and aperture, are sampled from distributions informed by natural observations, ensuring a faithful representation of geological reality.

The surrogate's architecture is particularly interesting: it combines a 3D convolutional neural network (CNN) with feed-forward layers. This hybrid configuration allows the model to effectively capture both local spatial features and global interactions within the domain. Three surrogate models were trained on data generated by DFM simulations, each calibrated for a different fracture-to-matrix conductivity contrast. Evaluations showed high accuracy, with normalized root-mean-square errors below 0.22 across most test cases, confirming the reliability of the method.

Computational Efficiency and the Role of GPUs

The practical applicability of these surrogate models was demonstrated by comparing numerically homogenized conductivities with surrogate predictions in two macro-scale problems: computing equivalent conductivity tensors and predicting outflow from a constrained 3D domain. In both scenarios, surrogate-based upscaling maintained the necessary accuracy while substantially reducing computational cost.

The most significant benefit lies in computational efficiency: model inference, when performed on a GPU, achieved speedups exceeding 100x compared to conventional methods. This data is crucial for decision-makers and infrastructure architects, as it highlights the potential of GPUs not only for training complex models but also for high-speed inference in scientific and engineering applications. The ability to execute complex simulations an order of magnitude faster can transform research and development cycles, enabling quicker iterations and more in-depth analyses.

Implications for On-Premise Deployments and TCO

The drastic reduction in computation time achieved through GPU inference has direct implications for deployment strategies, particularly for organizations evaluating self-hosted or on-premise solutions. The ability to accelerate the execution of complex models by over 100 times means that results can be obtained much faster with the same hardware infrastructure, or the same timings can be achieved with less powerful hardware or in reduced quantities. This translates into a positive impact on the Total Cost of Ownership (TCO), reducing operational and potentially capital expenditures.

For CTOs, DevOps leads, and infrastructure architects, the choice of hardware optimized for inference, such as GPUs, becomes a key factor in planning local stacks for AI/LLM workloads. Data sovereignty and compliance often require air-gapped or self-hosted environments, where the efficiency of local hardware is paramount. Models like the one described, offering high accuracy with extremely fast inference on GPUs, represent a concrete example of how algorithmic innovation, combined with the right hardware, can unlock new possibilities for research and engineering, maintaining data control and optimizing resources.