Reservoir simulation is one of the most computationally demanding tasks in petroleum engineering. Engineers run hundreds of thousands of grid cells through iterative pressure-saturation solves, waiting hours — sometimes days — for results that inform multi-million-dollar field development decisions.
Today, we're introducing BirSim: BIROVA AI's GPU-powered black oil reservoir simulator.
The Problem With Traditional Simulators
Legacy reservoir simulators were designed in an era when CPUs were the only option. Their solvers are fundamentally sequential — optimized for a small number of fast, general-purpose cores. Adding more CPUs gives diminishing returns. A model that takes 6 hours on 8 cores might only drop to 4 hours on 32 cores.
The result: reservoir engineers make fewer scenario runs. They simplify their models to fit available compute time. They accept wider uncertainty ranges because running enough realizations for robust P10/P50/P90 analysis is simply impractical.
BirSim was built to remove that constraint entirely.
What Is Black Oil Simulation?
The black oil model is the industry workhorse for reservoir simulation. It treats reservoir fluids as three phases — oil, water, and gas — with pressure-dependent PVT (pressure-volume-temperature) properties described by standard correlations.
The governing equations express:
- Darcy flow through porous media for each phase
- Mass conservation across all grid cells and timesteps
- Capillary pressure and relative permeability relationships
- Well equations coupling reservoir to surface conditions
These form a large, sparse, nonlinear system solved iteratively at every timestep. For a one-million-cell model running 3,650 timesteps (10 years, daily), that's billions of floating-point operations — a perfect workload for GPU parallelism.
Why GPU Acceleration Changes Everything
A modern datacenter GPU has over 10,000 CUDA cores running in parallel. A reservoir grid with 500,000 cells can distribute its pressure-saturation solve across those cores simultaneously, collapsing timestep wall-clock time from minutes to seconds.
BirSim's solver architecture is built from the ground up for GPU compute:
Fully Implicit Formulation (FIM). BirSim uses a fully implicit discretization scheme — coupling pressure and saturation unknowns at every timestep — which delivers the numerical stability required for large, heterogeneous models without restrictive timestep size limits.
CUDA-Accelerated Linear Algebra. The heart of every timestep solve is a large sparse linear system. BirSim uses cuSPARSE and cuBLAS under the hood, with a preconditioned iterative solver (ILU-preconditioned GMRES) tuned specifically for reservoir simulation sparsity patterns.
Adaptive Timestep Control. BirSim's timestepper monitors convergence behavior and automatically adjusts timestep size — larger steps in stable regions, smaller steps near wells or during rapid saturation changes — maximizing throughput without sacrificing accuracy.
Multi-GPU Scaling. For models exceeding single-GPU memory, BirSim partitions the reservoir grid across multiple GPUs with minimal communication overhead, enabling models up to 10 million active cells on a 4-GPU node.
Key Capabilities
| Feature | BirSim |
|---|---|
| Maximum active grid cells | 10 million |
| Phases | Oil, water, free gas, dissolved gas |
| Formulation | Fully implicit (FIM) and IMPES |
| Well types | Vertical, deviated, horizontal, multilateral |
| PVT input | Eclipse-compatible keywords (PVTO, PVTG, PVTW) |
| Grid types | Corner-point, Cartesian, unstructured |
| Multi-GPU | Up to 8 GPUs per simulation |
| Uncertainty runs | Native Monte Carlo ensemble mode |
| Output | Pressure maps, saturation maps, well performance, GOR, WCT |
Performance: Real Numbers
On a single NVIDIA A100 80GB GPU, BirSim benchmarks against a reference CPU cluster (32-core Intel Xeon) on identical black oil models:
| Model Size | CPU Cluster (32 cores) | BirSim (1× A100) | Speedup |
|---|---|---|---|
| 100,000 cells, 10 years | 22 min | 2.1 min | 10.5× |
| 500,000 cells, 10 years | 3.1 hrs | 18 min | 10.3× |
| 1,000,000 cells, 10 years | 6.8 hrs | 41 min | 9.9× |
| 5,000,000 cells, 5 years | 38 hrs | 3.9 hrs | 9.7× |
For uncertainty quantification workflows — running 100 realizations for P10/P50/P90 forecasting — BirSim's multi-GPU ensemble mode reduces a task that would take two weeks of CPU cluster time to under 4 hours.
Black Oil Is Just the Beginning
BirSim's black oil simulator is the first module in BIROVA AI's simulation stack. Compositional simulation (EOS-based, for gas condensate, miscible flooding, and EOR) is in active development.
More importantly, BirSim is not a standalone tool. It is the simulation engine inside the BIROVA AI platform — the compute layer that BIROVA AI's autonomous reservoir agents use to run, evaluate, and iterate on simulation scenarios without requiring manual engineer intervention.
When an AI agent asks "what happens to GOR if we reduce the BHP on Well A-3 by 200 psi?", it fires a BirSim run, parses the result, reasons about it, and feeds the insight back — in seconds, not hours.
Getting Access
BirSim is currently available in early access as part of the BIROVA AI platform. We are working with a small group of upstream operators and independent E&P companies to validate performance on real field models.
If you want to run your reservoir model on BirSim — bring your own Eclipse deck, your own PVT tables, your own grid — request a demo and our team will set up a benchmarking session with your data.
The era of waiting overnight for simulation results is over.