Skip to main content

DRA2303 Airfoil Environments

This module provides DRA2303 airfoil CFD environments for reinforcement learning with two simulation scenarios at Ma=0.2 and Ma=0.7 with chord-based Re=400000.

Two actuation strategies share the DRA2303Base class:

DRA2303Jet Jet-based actuation. Actions lie in [−MAX_CONTROL, +MAX_CONTROL] and are passed directly to the CFD solver.

DRA2303SurfaceWave Actuated traveling surface wave parameterized by [amplitude, speed, wavelength]. Features two wave zones (pressure side and suction side) that are both actuated using the same wave parameters. All three parameters are strictly positive, which requires a per-action asymmetric action space and a non-zero reset; see the class docstring for config details.

Both environments receive the same solver force output [F_x, F_y, F_z, C_P, area] (nDim + 2 elements), where C_P is the power coefficient (0.0 for the jet case) and area is the wetted wall area used for normalisation. Forces and power are pre-normalised by the solver by dynamic pressure (q_∞ = ρ_∞ · ½ · U_∞²) and q_∞ · U_∞ respectively, so dividing by area yields the final dimensionless coefficients. Observation sizing and normalization for this extended force vector are handled in DRA2303Base.

DRA2303Base Objects​

class DRA2303Base(MaiaFlowEnv)

Base class for DRA2303 airfoil environments.

Uses the structured-grid m-AIA solver (MAIA_STRCTRD).

Supports two simulation scenarios:

  • Ma = 0.2 with chord-based Re = 400000
  • Ma = 0.7 with chord-based Re = 400000

Both actuation variants (jet and surface wave) receive the solver force output [F_x, F_y, F_z, C_P, area] (nDim + 2 elements). Forces F_x … F_z and C_P are pre-normalised by the solver using the dynamic pressure q_∞ = ρ_∞ · ½ · U_∞² (forces) and q_∞ · U_∞ (power), so dividing by the returned area gives the final dimensionless drag, lift, and power coefficients.

This class provides:

  • configure_observations() – sizes the force observation slot to nDim + 2 per boundary.
  • setup_normalization() – handles the C_P and area elements in the 'U_inf' normalization strategy.
  • get_reward() – unified reward for both actuation variants:

.. math::

R = -C_D - \omega \cdot C_P

where C_D = F_x / area and C_P = forces[3] / area. For the jet case C_P = 0, so the reward reduces to -C_D.

configure_observations​

def configure_observations() -> None

Configure the number of observations.

Overrides the base class to account for the extended solver force output [F_x, F_y, F_z, C_P, area] (nDim + 2 elements per boundary instead of nDim).

setup_normalization​

def setup_normalization() -> None

Set up observation normalization.

Overrides the base class to handle the extended force vector [F_x, F_y, F_z, C_P, area] (nDim + 2 elements per boundary).

For the 'U_inf' strategy, free-stream values computed from the isentropic relations in __init__ are used:

  • velocities scaled by U_T (total free-stream speed),
  • density by rho_inf,
  • pressure by the dynamic pressure q_inf = ½ · rho_inf · U_T².

Force-slot entries use loc = 0, scale = 1 since the solver already normalises them by q_inf (forces) / q_inf·U_T (power).

get_reward​

def get_reward() -> Tuple[float, Dict]

Compute the unified reward: :math:R = -C_D - \omega \cdot C_P.

The solver returns [F_x, F_y, F_z, C_P, area]. Forces and power are pre-normalised by q_∞ and q_∞ · U_∞ respectively; dividing by area yields the final dimensionless coefficients::

C_D = forces[0] / area C_P = forces[3] / area

For the jet environment C_P = 0 and the reward reduces to -C_D.

Returns:

Tuple of (reward, info_dict). info_dict contains 'forces', 'C_D', and 'C_P'.

DRA2303Jet Objects​

class DRA2303Jet(DRA2303Base)

DRA2303 airfoil with jet-based flow control.

Uses the same actuation setup as the :class:~hydrogym.maia.envs.Cube environment: actions lie in [−MAX_CONTROL, +MAX_CONTROL] (scaled by the base-class step()) and are passed directly to the CFD solver. The number of jet actuators is read from fvNoJets in the property file.

The solver returns [F_x, F_y, F_z, C_P, area] with C_P = 0.0, so the reward reduces to -C_D. Observation sizing, normalization, and reward computation are all provided by :class:``3.

Attributes:

  • ``4 int - Number of jet boundary conditions.

convert_action​

def convert_action(action: np.ndarray) -> np.ndarray

Pass the (already MAX_CONTROL-scaled) jet actuation to the solver.

Arguments:

  • action - Action array scaled by MAX_CONTROL.

Returns:

Action array for the CFD solver.

DRA2303SurfaceWave Objects​

class DRA2303SurfaceWave(DRA2303Base)

DRA2303 airfoil with actuated traveling surface wave.

The surface is driven by a traveling wave with three strictly positive control parameters:

============= ============================================= Parameter Description ============= ============================================= amplitude Wave amplitude speed Wave propagation speed wavelength Spatial period of the wave ============= =============================================

Two wave zones: This environment features two wave zones, one on the pressure side and one on the suction side of the airfoil. Both zones are actuated using the same wave parameters [amplitude, speed, wavelength].

Differences from :class:DRA2303Jet:

  • Action space – Per-action [lower_bound, upper_bound] read from maia.action_lower_bounds / maia.action_upper_bounds in the config. The MAX_CONTROL scaling in step() is not applied.
  • Reset – Uses [amplitude_init, speed_init, wavelength_init] from config (zeros would crash the CFD solver).
  • Reward – :math:5 (shared with :class:`DRA2303Jet` via :class:7; here C_P > 0 because the surface wave performs work on the fluid).

Required additional entries in the maia section of the YAML config:

.. code-block:: yaml

maia: amplitude_init: 0.05 speed_init: 1.0 wavelength_init: 2.0 action_lower_bounds: [0.01, 0.1, 0.5] action_upper_bounds: [0.2, 3.0, 10.0] omega: 0.1 wave_zone_ids: [10, 11] # pressure side, suction side BC IDs

set_observation_action_spaces​

def set_observation_action_spaces() -> None

Override to use a normalized [0, 1] action space per parameter.

The RL agent always operates in [0, 1]; physical values are recovered inside :meth:convert_action via the per-parameter affine mapping lower + action * (upper - lower).

convert_action​

def convert_action(action: np.ndarray) -> List[float]

Scale a normalized [0, 1] action to physical wave parameters.

Applies the per-parameter affine map::

physical[i] = lower[i] + action[i] * (upper[i] - lower[i])

The same physical parameters are applied to both wave zones (pressure side and suction side).

Arguments:

  • action - Normalized action array with values in [0, 1], corresponding to [amplitude, speed, wavelength].

Returns:

Physical actuation values for the CFD solver.

step​

def step(
    action: Optional[np.ndarray] = None
) -> Tuple[np.ndarray, float, bool, bool, Dict]

Advance the environment by one step.

Overrides the base class to skip MAX_CONTROL scaling. The incoming action is in [0, 1] and is converted to physical units by :meth:convert_action. The same physical wave parameters are applied to both wave zones (pressure and suction side).

Arguments:

  • action - Normalized action in [0, 1] for [amplitude, speed, wavelength].

Returns:

Tuple of (observation, reward, terminated, truncated, info).

reset​

def reset(seed: Optional[int] = None,
          options: Optional[Dict] = None) -> Tuple[np.ndarray, Dict]

Reset the environment with physically valid initial wave parameters.

Overrides the base-class reset to avoid setting zero control actions, which would crash the CFD solver for the surface-wave boundary condition. init_action is passed in physical units directly, bypassing the [0, 1] → physical scaling used during normal steps. The same initial parameters are applied to both wave zones.

Arguments:

  • seed - Optional random seed.
  • options - Optional reset options (unused).

Returns:

Tuple of (initial_observation, info).