flepimop2-op_system

flepimop2-op_system provides the flepimop2 system adapter for op_system.

It packages the flepimop2.system.op_system provider so flepimop2 can load compiled models with spatial or trait-structured fields, nonlocal interactions, compartmental dynamics and reaction channels. The provider carries coordinate and operator metadata to an engine; that numerical consumer owns discretization, operator staging and time/event integration.

Read a compiler result

The illustrated core example shows held forcing, independent replicate shapes and a generated drift/population table. It evaluates a fixed state; an engine owns time integration.

Bound system steppers accept and return Flepimop2's namespace-neutral Array protocol. The provider performs no state conversion: NumPy, JAX, and other supported array namespaces are selected by the state supplied by the numerical engine. module: op_system selects this system provider; it does not select an array namespace. There is no array_backend configuration field. See the terminology guide for legacy compiler keyword compatibility.

The wheel ships py.typed for downstream type checking. Repository quality checks run strict mypy on the adapter in its separate environment; the root compiler's mypy exclusion does not bypass that check. just build-all and CI's provider-typing job additionally check the installed wheel's public types and verify that an invalid state argument fails strict downstream typing.

System options preserve normalized axis metadata for engine providers: axis_order, axis_sizes, numeric axis_coords, original axis_labels, and declared axis_types. Consumers can therefore distinguish categorical, ordinal, and continuous semantics without inferring them from coordinate values.

Time-indexed parameter tables select interpolation through the inline spec's time_interpolation field: linear is the default, and previous holds the value at each coordinate until the next, including the new value at an exact boundary. Both policies clamp beyond the endpoints. All bound flat, PyTree, block, and reaction steppers share the policy and preserve the state's array namespace.

The provider also exposes time_axis, time_interpolation, time_coordinates, and forcing_breakpoints through system.option(...). Coordinates and breakpoints are immutable tuples. Hold-mode tables publish all coordinates after the first as possible forcing changes; linear tables and unused time axes publish no breakpoints. A single-coordinate time table is constant and has no forcing changes. Parameter requests still include the complete declared time axis, so the parameter producer supplies the full table at run time.

An engine can use these options to configure forcing boundaries independently of its output grid. Exact breakpoint SSA also requires every other external time dependency in the propensity to be constant between those boundaries. The provider exposes the schedule; the engine chooses how to consume it.

It also provides the sparse_table parameter module, which assembles a dense array for a routing transition's matrix parameter (for example eta[time, imm:i, imm:j]) from a declared support. Each entry is a number or any nested parameter configuration, sampled with the request's leading axes:

parameter:
  eta:
    module: sparse_table
    indices: [imm, imm]
    entries:
      - index: [x0, x3]
        value: {module: fixed, value: 0.2}
      - index: [x5, x7]
        value: 1.0

Off-support entries are zero. SparseTableParameter.support(axes) returns the support positions for inference.

For an erfc mixing kernel on numeric coordinates, declare axes: [age, age] and params: {scale: 1.0, sigma: 2.0}. The provider constructs the float64 host matrix C[i, j] = scale * erfc(abs(age[i] - age[j]) / sigma) once when the system is built. It uses Python's complementary error function directly to preserve small positive tail values; SciPy is not a runtime dependency. Kernel declarations without axes retain their existing scalar params.value behavior.

The provider source headers grant GPL version 3 or any later version (GPL-3.0-or-later). The standalone core uses MIT. The shared citation records both existing grants; preserve the relevant license when reusing code.