Automated Multi-Well Multi-Property Calibration of Basin and Petroleum System Models in a Massively Parallel High-Performance Computer Environment

Accurate basin and petroleum system models are crucial for understanding the complexities of petroleum systems, including predicting potential hydrocarbon accumulations, their timing, and migrating oil and gas phase behavior. These models rely on calibration with measured data including subsurface temperatures and thermal maturity indices. Calibrating this measured data is essential because the transformation of kerogen into hydrocarbons is driven by temperature-dependent kinetic reactions. Traditional calibration methods are often manual and time-consuming, but they are still used because of the high computational costs of optimization algorithms and/or limitations in existing simulators. This paper demonstrates that by leveraging a purpose-built simulator for high-performance computing environments and state of the art optimization algorithms, automated calibration of these models is not only feasible but also highly effective. Three components enable practical automated calibration of these models: 1) a simulator specifically designed to be run in a massively parallel high-performance computing environment is used, 2) an easy-to-use Python interface to parse and write input files and read output files for the simulator, and 3) the vast scientific Python libraries. To automatically calibrate basin models, an initial (uncalibrated) model is constructed. Uncertain parameters are defined either as numerical range, discrete values, or possible scenarios. Bayesian optimization with sampler based on Tree-structured Parzen Estimator is used for calibration. The automation workflow is applied on multiple real three-dimensional models to test the viability of the method. In one model, the uncertainty in basal heat flow and surface temperature histories is represented as a bulk shift in the initial maps used (±30%). Calibration data used are the borehole measured temperature and vitrinite reflectance with the first being weighted twice as the latter during error calculation. Results show that the optimization algorithm is able to calibrate the four wells at the same time. The optimization algorithm samples more realizations/solutions closer to the local minima in the response surface. This is because the optimization algorithm attempts to balance between exploitation (generating realizations near local minima) and exploration (generating realizations in areas with sparse response). In addition, in this case, the model is more sensitive to the basal heat flow history rather than the surface temperature history. Improvements that can be implemented to accelerate the optimization includes utilizing parallel simulations and halting the optimization once a predefined calibration threshold is achieved. Overall, the developed framework opens opportunities to apply different state of the art algorithms to aid in basin and petroleum system modeling. The novelty in the work lies in the fact that by combining multiple technologies, a practical workflow for automated calibration of realistic basin and petroleum system modeling is achievable. Thus, previous limitations, such as computational constraints, on such workflows is overcome.