GeoBrain is an open, modular, and extensible platform for Geoscientific Bayesian Reasoning with Artificial Intelligence, designed specifically for integrated subsurface modeling.
By combining differentiable physics, Bayesian inference, and deep learning, GeoBrain enables end-to-end workflows for subsurface characterization, from geomodeling and rock physics to geophysical simulation and inversion. It supports both explicit inversion of physical parameters and implicit inversion using neural network reparameterizations, enabling flexible and geologically consistent modeling.
- Differentiable Multiphysics Modeling: Seamlessly integrates geostatistics, reservoir simulation, rock physics, and geophysics in a computational graph.
- Automatic Differentiation: Enables inversion without hand-crafted gradients -- PyTorch autograd flows through the entire forward chain.
- Deep Neural Network Integration: Incorporates neural networks for prior modeling (Deep Image Prior), surrogate modeling, latent space representations, and network-based reparameterizations.
- Bayesian Gradient-Informed Samplers: Four samplers (SVGD, HMC, NUTS, LDS) for rigorous uncertainty quantification.
- 70+ Rock Physics Models: Effective medium theories, granular media models, fluid substitution, empirical relations, anisotropy, and resistivity.
- Plug-and-Play Architecture: Each physics module is self-contained and composable -- add new physics without rewriting core logic.
- Real-World Applications:
- CO2 storage site characterization at the Illinois Basin -- Decatur Project (IBDP).
- Joint seismic--resistivity inversion for the Sleipner CCS site in the Norwegian Sea.
Geostatistical Simulation |
Rock Physics Modeling |
Acoustic Wave Propagation |
Reservoir Flow Simulation |
Full Waveform Inversion |
Bayesian AVO Inversion |
git clone https://github.com/GeoBrain-Project/GeoBrain.git
cd geobrain
pip install -e ".[all]"Requirements: Python 3.9+, PyTorch 2.0+, NumPy >= 1.22, SciPy >= 1.8, Matplotlib >= 3.5
Some examples (Marmousi2, IBDP, Sleipner, etc.) require external data files. The data is included as zip files under examples/data/. Simply unzip them:
cd examples/data
unzip "*.zip"After extraction the directory should look like:
examples/data/
├── marmousi/ # Marmousi2 velocity & density SEG-Y files
├── svgd/ # IBDP pretrained model & seismic data
├── multiphysics/ # Sleipner joint inversion data
└── pvt/ # PVT tables for flow simulation
import torch
from geobrain.geomodel import Simulator, SimulationConfig
from geobrain.physics.rock import SoftSand, Gassmann, DensityModel, v_from_moduli
from geobrain.physics.wave import RickerWavelet, compute_reflectivity, create_conv_matrix
from geobrain import InverseProblem, SVGD
# 1. Generate geological model
config = SimulationConfig(shape=(100, 200), lh=15.0, lv=5.0,
mean=0.25, std=0.05, seed=42)
porosity = Simulator.create("fft_ma").simulate(config)
# 2. Rock physics: porosity -> elastic properties
K_m, G_m = 36.6, 44.0 # Mineral moduli (GPa)
K_fl, rho_fl = 3.06, 1.08 # Fluid properties
rho_m = 2.65 # Mineral density (g/cm3)
pressure = torch.ones_like(porosity) * 20.0
k_dry, g_dry = SoftSand()(K_m, G_m, porosity, 0.4, 7, pressure)
k_sat, g_sat = Gassmann()(k_dry, g_dry, K_m, K_fl, porosity)
rho = DensityModel()(porosity, rho_m, rho_fl)
vp, vs = v_from_moduli(k_sat, g_sat, rho)
# 3. Seismic forward modeling
refl = compute_reflectivity(vp[:-1], vs[:-1], rho[:-1],
vp[1:], vs[1:], rho[1:],
theta=[0], method='shuey')
wavelet, _ = RickerWavelet()(f0=25.0, dt=0.001)
W = create_conv_matrix(wavelet, refl.shape[-1])
seismic = refl @ W[len(wavelet)//2 : -(len(wavelet)//2)].T
# 4. Deterministic inversion
problem = InverseProblem(forward_fn=forward, observed=seismic, noise_std=0.01)
result = problem.create_inverter(initial_model=m0).run(problem.observed, max_epochs=200)
# 5. Bayesian uncertainty quantification
posterior = problem.as_posterior(log_prior=my_prior)
samples = SVGD(target=posterior).run(n_samples=50, n_steps=500)geobrain/
├── core/ # InverseProblem, registry system
├── geomodel/ # Geological model generation
│ ├── geostats/ # FFT-MA, SGS, variograms
│ ├── implicit/ # Differentiable implicit modeling
│ └── geogen/ # GAN, VAE, Diffusion generators
├── physics/ # Multiphysics simulations
│ ├── rock/ # 70+ rock physics models
│ ├── wave/ # Acoustic & elastic propagation, AVO, wavelets
│ └── flow/ # Differentiable reservoir simulation
├── optim/ # Deterministic inversion (L-BFGS, Gauss-Newton)
├── bayes/ # Bayesian samplers (SVGD, HMC, NUTS, LDS)
├── nn/ # Neural network components (Bayesian layers)
├── data/ # Data loading and transforms
├── io/ # SEG-Y, LAS, Eclipse grid I/O
├── vis/ # Visualization utilities
└── decision/ # Decision-making under uncertainty
examples/
├── 01_geomodel_sim.py ... 15_sleipner_joint_inv.py # Python scripts
├── notebooks/ # Matching Jupyter notebooks for all 15 examples
├── figs/ # Generated figures and animations
└── data/ # External datasets (see Installation)
We provide 15 end-to-end examples covering geomodeling, multiphysics simulation, deterministic inversion, and Bayesian inference. Each example is available as both a Python script (examples/) and an interactive Jupyter notebook (examples/notebooks/).
| # | Example | Description |
|---|---|---|
| 01 | geomodel_sim | Geostatistical simulation with FFT-MA |
| 02 | geomodel_cosim | Co-simulation, SGS & variograms |
| 03 | implicit_modeling | Implicit geological modeling |
| 04 | implicit_inversion | Karst cave detection via implicit model inversion |
| 05 | rock_physics | Rock physics modeling workflows |
| 06 | seismic_avo_forward | AVO forward modeling |
| 07 | acoustic_forward | 2D acoustic wave propagation (Marmousi2) |
| 08 | elastic_forward | 2D elastic wave propagation (Marmousi2) |
| 09 | acoustic_fwi | Acoustic full waveform inversion |
| 10 | flow_simulation | Oil-water two-phase flow simulation |
| 11 | flow_dynamic_wells | Dynamic well control & scheduling |
| 12 | seismic_inversion | Deterministic inversion (3 parameterizations) |
| 13 | bayes_avo_inversion | Bayesian AVO inversion (4 samplers) |
| 14 | ibdp_inversion_svgd | IBDP CO2 site -- latent-space SVGD |
| 15 | sleipner_joint_inv | Sleipner CO2 -- joint seismic-resistivity inversion |
| Method | Description |
|---|---|
| FFT-MA | Fast spectral-domain Gaussian simulation for large 2D/3D grids |
| SGS | Sequential Gaussian Simulation with Simple/Ordinary Kriging; conditional on well data |
| Co-simulation | Multi-variable correlated fields (e.g., porosity + sand volume) via CoSimConfig |
| Variogram | Composite variogram models (spherical, exponential, Gaussian) via VariogramModel |
| Implicit Modeling | Differentiable Geomodeling |
| VAE / GAN / Diffusion | Deep generative models for realistic geological facies generation |
| Category | Models |
|---|---|
| Effective medium | VRH, Hashin-Shtrikman, DEM, Self-Consistent, Critical Porosity, Hudson, Eshelby-Cheng |
| Granular media | Hertz-Mindlin, SoftSand, StiffSand, ContactCement, Walton, MUHS, PCM, Digby, Thomas-Stieber |
| Fluid substitution | Gassmann (forward & inverse), Wood, Brie, Batzle-Wang, Biot, CO2-Brine, Brown-Korringa |
| Empirical | Gardner, Han, Castagna, Krief, Raymer-Hunt, Storvoll, Japsen, Hillis |
| Anisotropy | Thomsen (VTI/HTI), Backus averaging, Bond transform |
| Permeability | Kozeny-Carman, Owolabi, Panda-Lake, Revil, Fredrich, Bloch, Bernabe |
| Resistivity | Archie's law |
| QI | Rock Physics Template (RPT), screening, cement estimation |
Built-in MINERALS and FLUIDS databases, plus RockPhysicsWorkflow presets for common scenarios.
| Component | Details |
|---|---|
| Propagators | 2D/3D acoustic & elastic FDTD solvers with gradient checkpointing |
| AVO | Zoeppritz (exact), Aki-Richards (4-term), Shuey (3-term + AVO attributes) |
| Wavelets | Ricker, Gaussian, Ormsby, Klauder — class-based and functional APIs |
| Survey | Source, Receiver, Survey geometry management; SeismicData container |
| Convolution | convolve_trace, convolve_gather, apply_wavelet, create_conv_matrix |
| Metrics | MSE, RMSE, MAPE, SNR |
- Two-phase oil-water reservoir simulation on structured grids
- Wells: injection/production with rate or BHP control; dynamic scheduling (add, shut-in, restart)
- PVT: pressure-volume-temperature tables (PVDO, PVTW) or analytic models
- Relative permeability: Corey or tabular (SWOF)
- Solver: fully implicit Newton solver with adaptive time stepping
- Differentiable: gradients flow through the simulator for history matching
| Feature | Details |
|---|---|
| Parameterizations | Explicit (direct), Latent (autoencoder), Network (Deep Image Prior) |
| Optimizers | Adam, L-BFGS (Wolfe line search), Gauss-Newton |
| Losses | MSE, NRMSE, L1, Huber |
| Constraints | Bound (clamp), Clip (projection), Sigmoid (differentiable) |
| Regularizers | L2, L1 |
| Bridge | InverseProblem unifies deterministic and Bayesian solvers |
| Sampler | Method | Best For |
|---|---|---|
| SVGD | Stein Variational Gradient Descent | Multi-modal posteriors, large ensembles |
| HMC | Hamiltonian Monte Carlo | High-dimensional with tuned step size |
| NUTS | No-U-Turn Sampler | Auto-tuned HMC |
| LDS | Langevin Dynamics (ULA/MALA) | Simple gradient-based sampling |
Distributions: Gaussian, GaussianMixture, Posterior. Kernels: RBFKernel, IMQKernel with adaptive bandwidth. Diagnostics: mmd, energy_distance.
- Bayesian layers:
LinearFlipout,Conv2dFlipout,Conv3dFlipoutwith KL divergence - Activations:
ClippedLinearActivationfor bounded output
| Format | Functions |
|---|---|
| SEG-Y | read_segy, write_segy, read_segy_volume |
| LAS | read_las, write_las (well log data) |
| Eclipse | read_egrid, read_init, read_restart, read_summary (reservoir grids) |
- Seismic:
plot_section,plot_wiggle,plot_gather - Fields:
plot_field,plot_slices,plot_comparison - Wells:
plot_logs(multi-track well log display) - Production:
plot_production,plot_watercut,plot_well_summary - Colormaps: geophysics-specific colormaps (Seismic, Velocity, etc.)
- Transforms:
Normalize,Clamp,Sigmoid,Logit,Log,Exp,Compose - Datasets:
SimpleTensorDataset,GeoDatasetfor PyTorch data loading
ValueOfInformation: Value of Perfect Information (VOPI) analysis for data acquisition planningClosedLoopManager: closed-loop reservoir management (observe → update → decide)
GeoBrain builds upon excellent open-source projects:
- Seismic Reservoir Characterization: Implementation inspired by SeReM - Seismic Reservoir Modeling.
- Rock Physics Module (
geobrain.physics.rock): Implementation inspired by rockphypy - Python Tool for Rock Physics. - Wave Propagation Module (
geobrain.physics.wave): Implementation inspired by deepwave - Differentiable Finite-Difference Wave Propagation, and ADFWI - Automatic Differentiation Full Waveform Inversion based on PyTorch. - Flow Simulation Module (
geobrain.physics.flow): Implementation inspired by ResSimAD.jl - Differentiable Reservoir Simulation in Julia. - Implicit Geological Modeling (
geobrain.geomodel.implicit): Implementation inspired by GemPy - Open-Source Implicit Geological Modeling in Python.
We thank the authors of these projects for their contributions to the open-source geoscience community.
If you use GeoBrain in your research, please cite:
@software{GeoBrain2026,
title = {GeoBrain: An End-to-End Differentiable Platform for Integrated Subsurface Modeling},
author = {Liu, Mingliang},
year = {2026},
url = {https://github.com/GeoBrain-Project/GeoBrain}
}This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
- Author: Mingliang Liu
- Email: mingliangliu@sdu.edu.cn