PICT¶
Registry ID: pict
Categories: acceleration, specialized, physics-informed
Architecture: differentiable GPU-accelerated multi-block PISO CFD solver.
Method architecture¶
flowchart LR
X["Geometry, initial state<br/>and boundary conditions"] --> M["Multi-block mesh and<br/>differentiable operators"]
M --> P["Momentum predictor"]
P --> C["Pressure-correction<br/>PISO iterations"]
C --> U["Velocity and pressure<br/>update"]
U --> Y["Differentiable CFD state"]
Y -. next time step .-> P
classDef input fill:#ffffff,stroke:#167d9a,stroke-width:1.5px,color:#17324d;
classDef core fill:#e9f8fb,stroke:#167d9a,stroke-width:1.5px,color:#17324d;
classDef physics fill:#fff4e8,stroke:#c97932,stroke-width:1.5px,color:#5a3518;
classDef output fill:#eef8f0,stroke:#34855c,stroke-width:1.5px,color:#173f2b;
class X input;
class M,P,U core;
class C physics;
class Y output;
PICT is a differentiable solver framework rather than a stand-alone direct surrogate. The numerical scheme, mesh blocks, linear solvers, and differentiable components must be described as part of the method.
NAVIER-CFD library flow¶
flowchart LR
D["Case metadata and<br/>solver trajectories"] --> AD["Canonical solver adapter"]
AD --> S["CFDSample + solver context"]
S --> C["Task and case configuration<br/>mesh · fields · BCs · time step"]
G["ResearchContract<br/>compute and approval rules"] --> C
C --> H["ModelHub / external PICT adapter"]
H --> E["Approved differentiable<br/>solver execution"]
E --> P["State prediction and gradients"]
P --> V["Residual, conservation,<br/>accuracy and wall-clock metrics"]
V --> F["FigureLab + manifests"]
classDef data fill:#ffffff,stroke:#506784,stroke-width:1.3px,color:#17263a;
classDef config fill:#e8f6fb,stroke:#1580a0,stroke-width:1.3px,color:#17324d;
classDef execute fill:#eef8f0,stroke:#34855c,stroke-width:1.3px,color:#173f2b;
classDef verify fill:#fff4e8,stroke:#c97932,stroke-width:1.3px,color:#5a3518;
class D,AD,S,G data;
class C,H config;
class E,P execute;
class V,F verify;
Execution boundary
PICT execution is an explicit solver action. NAVIER-CFD's current read-only MCP tools do not launch it automatically.
Suitable tasks¶
Simulation-coupled learning, inverse problems, control, and learned CFD components.
Reference¶
Franz et al., PICT, Journal of Computational Physics, 2025.