| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - geoscience |
| - porous-media |
| - neural-operator |
| - carbon-sequestration |
| - graph-neural-network |
| - reactive-infiltration |
| - pytorch |
| library_name: pytorch |
| thumbnail: banner.png |
| --- |
| |
| # Model Card — CHINO v1.0 |
|
|
| ## Model Description |
|
|
| **CHINO** (CHanneling Instability Neural Operator) is an attention-augmented graph |
| neural operator for predicting reactive infiltration instability in porous |
| media. It maps an initial CO2 concentration field to a predicted field at a |
| later time, capturing the large-scale dynamics of convective fingering in |
| deep saline aquifers. |
|
|
| | Field | Value | |
| |---|---| |
| | Model name | CHINO v1.0 | |
| | Architecture | Attention-augmented MeshGraphNet | |
| | License | Apache 2.0 | |
| | Framework | PyTorch | |
| | Parameters | 7,409,537 | |
|
|
| --- |
|
|
| ## Intended Use |
|
|
| CHINO is designed for researchers studying: |
|
|
| - Geological carbon capture and storage (CCS) |
| - Reactive infiltration instability (RII) in porous media |
| - Convective dissolution of CO2 in saline aquifers |
| - Neural operator methods for chaotic PDE systems |
|
|
| The model predicts individual realizations of the concentration field, |
| not ensemble statistics. It correctly captures large-scale spatial |
| structure (depth of CO2 penetration, broad convective pattern) but does |
| not reproduce exact finger positions, which are irreducibly stochastic at |
| high Rayleigh numbers. |
|
|
| --- |
|
|
| ## Physical Regime |
|
|
| | Parameter | Value | Physical interpretation | |
| |---|---|---| |
| | Rayleigh number Ra | 500, 1000, 1577 | Buoyancy-driven fingering intensity | |
| | Peclet number Pe | 317 | Advection vs. diffusion ratio | |
| | Da_S | 1.0 | Dissolution sink rate | |
| | Da_inj | 0.005 | CO2 injection rate | |
| | Domain | [0,10] x [0,1] | Non-dim. aquifer (500m x 2000m physical) | |
| | Grid | 200 x 40 | MAC staggered finite difference | |
| | t_max | 2.0 | 2000 years of geological sequestration | |
| |
| --- |
| |
| ## Architecture |
| |
| Each of the 6 processor blocks contains three sequential operations: |
| |
| 1. **Local edge update** (k=8 Moore + stride-4 edges, O(E)): captures |
| fine-scale concentration gradients at finger boundaries. |
| 2. **Global self-attention** (4 heads x 64 dim, O(N^2)): every node |
| attends to every other node simultaneously, representing the |
| instantaneous pressure coupling of the Darcy-Boussinesq system. |
| 3. **Node update MLP**: fuses local messages, global attention, and |
| sinusoidal time embedding. |
| |
| The attention matrix (N=8000 nodes, 4 heads) requires 1 GB of VRAM per |
| forward pass — negligible on modern GPUs. |
| |
| **Node input features (6):** c_in, S, x, y, t_in, Ra |
| |
| **Output:** c(x,y,t_out) >= 0 (Softplus activation) |
|
|
| --- |
|
|
| ## Training |
|
|
| | Detail | Value | |
| |---|---| |
| | Total epochs | 550 | |
| | Phase 1 (ep. 1-300) | Standard curriculum, w_phys=0.05 | |
| | Phase 2 (ep. 301-550) | Resumed, w_phys=0.3 | |
| | Loss | Anomaly relative L2 (0.7) + full L2 (0.3) + physics + BC | |
| | Optimizer | AdamW, lr=5e-4, cosine schedule | |
| | Hardware | NVIDIA RTX PRO 6000 Blackwell (95 GB) | |
| | Wall time | ~13 hours total | |
|
|
| --- |
|
|
| ## Performance |
|
|
| | Metric | Value | Seed | |
| |---|---|---| |
| | L2 at t=0.5 | 0.74 | Seeds 18-19, Ra=1577 | |
| | L2 at t=1.0 | 0.43 | Seeds 18-19, Ra=1577 | |
| | L2 at t=2.0 | 0.25 | Seeds 18-19, Ra=1577 | |
|
|
| At t=1.0, the model produces visible finger-like spatial structure that |
| corresponds spatially to the dominant fingers in the finite-difference |
| reference. At t=2.0, the broad swept pattern of merged fingers is |
| captured with correct left-right asymmetry. |
|
|
| --- |
|
|
| ## Limitations |
|
|
| - **Early-time prediction (t=0.5):** L2=0.74. Finger nucleation at short |
| times is controlled by sub-grid perturbations and is not predictable |
| deterministically. |
| - **Individual finger positions:** Not reproduced. The chaotic nature of |
| the Rayleigh-Taylor instability at Ra=1577 means finger positions are |
| sensitive to initial conditions at scales below the grid resolution. |
| This is a fundamental physical property, not a training failure. |
| - **Out-of-distribution Ra:** The model has not been tested above Ra=1577 |
| or below Ra=500. |
| - **2D only:** The current version is trained on 2D simulations. A 3D |
| extension is in development. |
|
|
| --- |
|
|
| ## Citation |
|
|
| If you use CHINO, please cite: |
|
|
| ```bibtex |
| @software{hier_majumder_2025_chino, |
| author = {Hier-Majumder, Saswata}, |
| title = {CHINO: CHanneling Instability Neural Operator}, |
| year = {2025}, |
| license = {Apache-2.0}, |
| url = {https://github.com/sashgeophysics/CHINO} |
| } |
| |
| @article{sun2020geological, |
| title = {Geological Carbon Sequestration by Reactive Infiltration Instability}, |
| author = {Sun, Yizhuo and Payton, Ryan L. and Hier-Majumder, Saswata and Kingdon, Andrew}, |
| journal = {Frontiers in Earth Science}, |
| volume = {8}, |
| pages = {533588}, |
| year = {2020}, |
| doi = {10.3389/feart.2020.533588} |
| } |
| ``` |
|
|
