AI-Assisted CFD Simulation

Computational fluid dynamics and machine learning for methane detection and chamber design.

Overview

Detecting methane emissions with a static flux chamber is at least a seven-dimensional design problem. The relevant variables include well radius and height, chamber radius and height, outlet-hole radius, methane leak rate, and wind speed. This parameter space is too large to test experimentally at the resolution needed to identify an effective chamber design. Wind conditions and leak rates also cannot be controlled during field measurements, so parametric sweep simulations are needed to isolate the effect of each variable and evaluate interactions among them. The simulations also allow the system to be tested under extreme conditions that may be rare in the field but are important for assessing safety risk and detection performance.

Time-dependent simulations were automated in COMSOL Multiphysics through MATLAB LiveLink and run on high-performance computing resources, including the University of Southern California's computing cluster. Because each simulation is computationally expensive, a neural-network surrogate model was trained on the simulation results. An operator can enter a well's height and diameter, select an acceptable risk level, and receive a recommended chamber design within minutes.

Diagram showing the seven variables considered in the methane chamber design problem
The seven-dimensional design space includes well and chamber geometry, outlet size, methane leak rate, and wind speed.

CFD Simulation Workflow

The study used automated parameter sweeps to generate methane concentration data across 228 scenarios, 1,000 observation points, and 11 time steps during the first minute of each simulation.

  1. Define the chamber geometry, well geometry, outlet, methane leak rate, wind conditions, and observation locations.
  2. Automate time-dependent COMSOL Multiphysics sweeps through MATLAB LiveLink on high-performance computing resources.
  3. Extract methane concentrations at each observation point and time step.
  4. Use the simulation results to train a faster model for chamber design screening.
Workflow connecting COMSOL and MATLAB simulations to a GRU model and chamber design recommendation tool
COMSOL and MATLAB generate the training data, the GRU learns the effects of the design parameters, and the recommendation tool identifies suitable chamber dimensions.

Safety and Detection Categories

Each simulated observation was assigned to one of three categories: above the methane lower explosive limit (LEL), within the sensor's detection window, or below the detection range. These categories convert raw concentration predictions into outcomes directly relevant to safety and monitoring performance.

Diagram classifying methane concentrations as above the LEL, within the detection window, or below the detection range
Simulation results are categorized according to safety risk and whether the methane concentration can be detected reliably.

Neural-Network Surrogate Model

A gated recurrent unit (GRU) model was trained on the simulation outputs to approximate the relationships among design variables, environmental conditions, methane dilution, and the three outcome categories. This surrogate model allows many candidate designs to be evaluated far more quickly than running a new CFD simulation for every case.

GRU training and validation curves with predicted versus actual methane concentrations
Training and validation metrics, together with predicted-versus-actual comparisons, are used to evaluate model performance across all three concentration categories.

Design Recommendation Tool

The resulting tool accepts well radius and height as operator inputs and recommends chamber radius, chamber height, and outlet-hole radius. It can return either a conservative design that prioritizes risk reduction or a detection-optimized design that prioritizes monitoring coverage.

Related Publication

Challenges in Designing a New Smart Static Methane Detection and Monitoring Chamber: Application to Plugged and Unplugged Abandoned Wells, Environments, 13(7), 411 (2026).

DomainMethane transport and chamber design
ModelingCOMSOL Multiphysics
AutomationMATLAB LiveLink
ComputingHigh-performance computing at USC
Machine learningGRU recurrent neural network