GraphMDO: Dynamic Multi-Fidelity MDO Framework

GraphMDO bridges data engineering and MDO. It extracts topological data (solvers, variables, fidelity levels) to form an oriented graph, specifically utilizing GEMSEO for semantic formulation and execution. Execution is handled natively by GEMSEO and the Surrogate Modeling Toolbox (SMT), with optimization workflows currently centered on constrained Bayesian optimization (ax-platform) and DOE exploration through GEMSEO. The framework supports single- and multi-objective optimization over graph-derived design variables with explicit inequality constraints.

Key Features

• Native Graph Formulation: Uses FalkorDB to store problem definitions (variables, tools, dependencies) as a property graph.
• Dynamic Problem Construction: Automatically translates the graph topology into an executable GEMSEO MDO formulation.
• Typed Study Contract: A Pydantic StudySchema (with a published JSON Schema) describes the study, and validate_study() reports every problem with a stable code before any tool runs.
• Multi-Fidelity Surrogates: Integrates SMT for Co-Kriging and other surrogate models.
• Constrained Bayesian Optimization: Leverages Ax Platform for robust optimization, supporting GEMSEO objectives, inequality constraints, and mixed discrete/continuous parameters.

Project Architecture

The framework is divided into four sequential phases, each corresponding to a layer of abstraction:

1. Graph Layer — FalkorDB stores the Fundamental Problem Graph (FPG): typed variables, tools, and directed connections.
2. Core Layer — Python modules read the graph as a typed StudySchema, validate it, and translate it into executable GEMSEO constructs (disciplines, design space, topology).
3. Execution Service — A FastAPI microservice that manages a pool of GEMSEO OptimizationProblem instances and exposes an HTTP evaluation endpoint.
4. Optimization Layer — Bayesian (Ax Platform) or DOE (GEMSEO Sobol) drivers run over the GEMSEO MDO scenario.

The StudySchema (mdo_framework.schema) is the contract between the layers: variables are RangeVar, ChoiceVar, FixedParam or StateVar, and invalid studies are reported by validate_study() before any tool runs. See the Study Schema reference.

flowchart TD
    subgraph USER["👤 User"]
        U1["Define Variables\n& Tools"]
        U2["Provide Tool\nCallables / Registry"]
        U3["Read Results"]
    end

    subgraph GRAPH["🗄️ Graph Layer — FalkorDB"]
        G1["GraphManager\n(graph_manager.py)"]
        G2[("FalkorDB\nProperty Graph\nFPG")]
        G3["FalkorDB Client\n(client.py)"]
        G1 -- "add_variable / add_tool\nconnect_input_to_tool" --> G2
        G3 -- "Cypher queries" --> G2
        G1 -- uses --> G3
    end

    subgraph CORE["⚙️ Core Layer — mdo_framework.core"]
        C1["TopologicalAnalyzer\n(topology.py)\nresolve_dependencies()"]
        C2["GraphProblemBuilder\n(translator.py)\nbuild_problem()"]
        C3["GemseoComponent\n(components.py)\nGEMSEO Discipline wrapper"]
        C4["SurrogateComponent\n(surrogates.py)\nSMT Co-Kriging"]
        C5["LocalEvaluator\n(evaluators.py)"]
        C6["StudySchema + validate_study()\n(schema.py, validation.py)"]
        C6 --> C1
        C6 --> C2
        C2 --> C3
        C2 --> C4
        C3 --> C5
    end

    subgraph GEMSEO["📐 GEMSEO MDO Engine"]
        GS1["DesignSpace"]
        GS2["MDOScenario / DOEScenario"]
        GS3["OptimizationProblem"]
        GS2 --> GS3
        GS1 --> GS2
    end

    subgraph OPT["🔬 Optimization Layer — mdo_framework.optimization"]
        O1["BayesianOptimizer\n(optimizer.py)"]
        O2["AxOptimizationLibrary\n(ax_algo_lib.py)\nCustom GEMSEO Algorithm"]
        O3["Ax Client\n(ax-platform)"]
        O4["DOE Explore\nSobol sampler"]
        O1 --> O2
        O1 --> O4
        O2 --> O3
        O3 -- "suggest_next_trials()" --> O2
        O2 -- "complete_trial() / mark_failed()" --> O3
    end

    subgraph SVC["🌐 Execution Service — services.execution"]
        S1["FastAPI App\n(main.py)"]
        S2["SchemaProvider\n(schema cache + TTL)"]
        S3["ProblemPool\n(GEMSEO problem pool)"]
        S4["/evaluate endpoint"]
        S5["/health endpoint"]
        S1 --> S2
        S1 --> S3
        S1 --> S4
        S1 --> S5
        S4 --> S3
    end

    %% Cross-layer data flow
    U1 --> G1
    U2 --> C2
    G2 -- "get_study_schema()" --> C6
    C1 -- "design parameters\n& bounds" --> O1
    C5 --> GS3
    GS1 --> O1
    GS3 --> O2
    O2 -- "evaluate_functions(x)" --> GS3
    O3 -- "best_parameterization" --> O1
    O1 --> U3

    %% Remote path
    O1 -. "RemoteEvaluator\n(HTTP /evaluate)" .-> S4
    S3 -. "GEMSEO problem\ninstance" .-> S4

    style USER fill:#1e293b,stroke:#64748b,color:#f1f5f9
    style GRAPH fill:#0f172a,stroke:#3b82f6,color:#bfdbfe
    style CORE fill:#0f172a,stroke:#8b5cf6,color:#ddd6fe
    style GEMSEO fill:#0f172a,stroke:#06b6d4,color:#a5f3fc
    style OPT fill:#0f172a,stroke:#f59e0b,color:#fef3c7
    style SVC fill:#0f172a,stroke:#10b981,color:#a7f3d0

Installation

This project uses uv for dependency management.

1. Install uv (if not installed): See astral.sh/uv.

2. Clone and Install:

   git clone https://github.com/jultou-raa/GraphMDO.git
   cd GraphMDO
   uv sync
   

3. FalkorDB: Ensure you have a running FalkorDB instance (e.g., via Docker):

   docker run -p 6379:6379 -it falkordb/falkordb
   

Usage

1. Defining a Problem (Python API)

You can programmatically build your MDO problem graph. Variables are typed models from mdo_framework.schema: RangeVar (bounded design variable), ChoiceVar, FixedParam and StateVar (a value computed by a tool).

from mdo_framework.db.graph_manager import GraphManager
from mdo_framework.schema import RangeVar, StateVar, ToolNode

gm = GraphManager()
gm.clear_graph()

# Define Variables
gm.add_variable(RangeVar(name="x", lower=0.0, upper=10.0))
gm.add_variable(RangeVar(name="y", lower=0.0, upper=10.0))
gm.add_variable(StateVar(name="z"))

# Define Tool
gm.add_tool(ToolNode(name="MyTool"))

# Define Connections
gm.connect_input_to_tool("x", "MyTool")
gm.connect_input_to_tool("y", "MyTool")
gm.connect_tool_to_output("MyTool", "z")

Nodes stored without a kind by earlier versions are rejected with a LEGACY_NODE error: delete and recreate them with the typed API.

2. Running Optimization

Once the graph is populated, you can run the optimization workflow. You need to provide the actual Python functions corresponding to the tool names in the graph.

from mdo_framework.core.translator import GraphProblemBuilder
from mdo_framework.optimization.optimizer import BayesianOptimizer
from mdo_framework.core.evaluators import LocalEvaluator
from mdo_framework.core.topology import TopologicalAnalyzer
from mdo_framework.schema import ObjectiveSpec
from mdo_framework.validation import validate_study

# Define tool implementation
def my_tool_func(x, y):
    return x + y  # Simple example

# Registry maps graph tool names to Python callables
tool_registry = {
    "MyTool": my_tool_func
}

# Read the graph as a typed StudySchema and check it before running
schema = gm.get_study_schema()
report = validate_study(
    schema, objectives=[ObjectiveSpec(name="z")], registry=tool_registry
)
assert report.valid, report.errors

# Build GEMSEO Problem from the schema
builder = GraphProblemBuilder(schema)
prob = builder.build_problem(tool_registry)

# Resolve Topology mapping design variables automatically from the schema
analyzer = TopologicalAnalyzer(schema)
resolved = analyzer.resolve_dependencies(["z"])

# Run Optimization
evaluator = LocalEvaluator(prob, builder.variable_specs)
optimizer = BayesianOptimizer(
    evaluator,
    resolved.design_variables,
    [ObjectiveSpec(name="z")],
)

# 5 Sobol trials (n_init) + 10 Bayesian iterations (n_steps) = 15 evaluations,
# plus the start point x0 when evaluate_x0=True
result = optimizer.optimize(n_steps=10, n_init=5)
print(f"Best Result: {result['best_objectives']} at {result['best_parameters']}")
print(f"Stopped because: {result['stop_reason']}, feasible: {result['feasible']}")
print(f"Trial History: {result['history']}")

3. Running Tests

uv run pytest tests/                           # everything
uv run pytest -m "not e2e" tests/              # fast unit tests
OMP_NUM_THREADS=1 uv run pytest -m e2e tests/  # real Ax + GEMSEO runs (tests/e2e/)

Tests that pin an open bug are marked xfail(strict=True) with the issue number in the reason. The PR that fixes the bug removes the marker.

Contributing

1. Follow PEP 8 guidelines.
2. Ensure 100% test coverage for new features.
3. Use uv run pre-commit run --all-files before committing.

License

This project is licensed under the Mozilla Public License 2.0 (MPL-2.0). See the LICENSE file for details.