Skip to main content

Resultados relacionados

Resultados totales (Incluyendo duplicados): 3
Encontrada(s) 1 página(s)

DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication Otros (Other)

linuswalter/PINN-for-Subsurface-Flow

  • Walter, Linus
This software computes the diffusion of fluid pressure p(x,t) in a 1D domain based on the concept of Physics Informed Neural Networks (PINN) for a heterogeneous modeling domain., GNU General Public License 2 or later (GPL-2.0), Field-data assimilation to calibrate rock properties in numerical physics-based reservoir models of hydrogeological applications is challenging. Recently, Artificial Neural Networks (ANN) have emerged as a promising alternative to handle noisy data seamlessly. However, in hydrogeology, even though data are abundant over time, observation wells are sparse over space, which results in insufficient data to train ANN. Here, we propose Physics-Informed Neural Networks (PINN) to bridge the gap of sparse spatial data by imposing physical conditions. We test ANN and PINN on a synthetic dataset for fluid pressure diffusion p(x,t) through a low-permeable porous medium that hosts a high-permeability equivalent fracture material. By comparing the ANN and the PINN as a function of the number of observation wells, we find that the PINN model outperforms the ANN when having less than 14 wells. Adding noise to the training data reveals the advantage of PINN to be more robust for random measurement errors or ambient noise. We finally test the applicability and limitations of PINN to represent fractures as a thin equivalent material while having no observations inside the domain and find that accuracy can be maintained when reducing the thickness of the equivalent material at the expense of increasing the computation time. Given that hydrogeological applications count with a limited number of observation wells, PINN appears as a more suitable machine learning tool than purely data-driven ANN for reservoir modeling. Therefore, we consider this work a starting point for developing more realistic reservoir models with heterogeneous material distribution based on the PINN architecture., LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809., Peer reviewed
DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication Otros (Other)

linuswalter/WellPINN

  • Walter, Linus
This software computes the diffusion of fluid pressure p(x,y,t) in a 2D domain for a single injection well based on Physics Informed Neural Networks (PINN)., The documentation is located in the README.md file., Pumping wells are a key component for modeling subsurface fluid flow, where their accurate representation is essential for reliable reservoir characterization through history matching of flow rates and well pressures, as well as for simulating operational scenarios. Physics-informed neural networks (PINNs) have recently emerged as a promising method for reservoir modeling, offering seamless integration of observational data and governing physical equations. However, existing PINN-based studies still face major challenges in capturing fluid pressure inference near pumping wells, particularly in the early phase after injection start. In response, we introduce 'WellPINN', a modeling workflow that combines the outputs of multiple sequentially trained PINN models. The first model covers the entire reservoir, while the size of each subsequently trained model equals the equivalent well radius of the previously trained model. This workflow allows us to iteratively approximate the radius of the equivalent pumping well to the actual pumping well dimensions. Applied to a single pumping well in a two-dimensional domain, our forward model demonstrates how combining three PINNs can bridge a spatial scale spanning three orders of magnitude, from the reservoir boundary down to the well radius. The results showcase that our sequential training of superimposing networks around the pumping well is the first workflow that enables accurate inference of fluid pressure from pumping rates across the entire injection time, significantly advancing the potential of PINNs for inverse modeling and operational scenario simulations. Future research might enhance 'WellPINN' by employing it to an inverse problem. Additionally, further investigation could test its effectiveness for heterogeneous domains and multiple injection wells., LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809., Peer reviewed
DIGITAL.CSIC. Repositorio Institucional del CSIC
Publication Otros (Other)

linuswalter/Real-Time-Forecast-IBDP

  • Walter, Linus
[Short description what the software is about] This software computes a forecast for the target variables seismic event count and maximum seimsic moment magnitude based on operational data (fluid pressure, flow rate) and a seismic catalog based on a random forest machine learning algorithm., [Links or references to publications or other documentation] https://github.com/linuswalter/Real-Time-Forecast-IBDP, Geological carbon storage (GCS) can significantly reduce emissions from hard-to-abate industries. However, large-scale deployment is hindered by the risk of induced seismicity, which has led to multiple project shutdowns. Mitigating such events would require unprecedented reliable seismicity forecasting. We present an adaptive near-real-time workflow that forecasts seismic variables. Using a random forest trained on sequentially partitioned data, we forecast seismicity rate and maximum magnitude. Applied to the Illinois Basin Decatur Project, the model receives operational parameters and the seismic catalog as input. Our results identify well pressure, past event counts, and prior maximum magnitudes as key predictors. However, accuracy declines over time as fluid pressure diffuses into the far-field, highlighting the need to incorporate spatiotemporal pressure diffusion for improved long-term forecasts. Our framework is the first to offer near-real-time forecasting from the operation start with uncertainty quantification, providing a foundation for next-generation seismic hazard mitigation in geo-energy operations., LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809., File List: 01_IBDP_RF_Model.py 02_IBDP_RF_Postproc.py, Peer reviewed