Spectral And Sobolev Guided Residual Network

Wave 5.2R Stage 6 tests bounded coordinate residuals with raw circular coordinates, frozen Fourier features, and sinusoidal representation networks. Every branch reconstructs a complete periodic forward transmission-error curve from causal setpoint inputs.

The completed Stage 6 campaign promoted no candidate. This API remains an experimental, reproducible model surface and must not be interpreted as the accepted production predictor.

Bounded coordinate residual models for Wave 5.2R Stage 6.

class scripts.models.spectral_sobolev_guided_residual_network.SineLayer(input_size, output_size, *, omega_0, first_layer)[source]

Bases: Module

Apply one SIREN-compatible affine layer and sine activation.

Parameters:

  • input_size (int)

  • output_size (int)

  • omega_0 (float)

  • first_layer (bool)

__init__(input_size, output_size, *, omega_0, first_layer)[source]

Initialize one periodic layer with the prescribed weight scale.

Parameters:

  • input_size (int)

  • output_size (int)

  • omega_0 (float)

  • first_layer (bool)

Return type:

None

forward(input_tensor)[source]

Return the sinusoidally activated affine output.

Parameters:

input_tensor (Tensor)

Return type:

Tensor

class scripts.models.spectral_sobolev_guided_residual_network.BoundedCoordinateResidualNetwork(condition_input_size, harmonic_order_list, angular_sample_count, angular_architecture, residual_bound_list, *, rank=12, condition_hidden_size=64, angular_hidden_size=64, fourier_feature_order_list=None, siren_omega_0=30.0)[source]

Bases: Module

Add one bounded low-rank angular residual to a PF-A curve.

The network factorizes the residual into condition weights and shared angular basis functions. This evaluates a complete uniform curve with one matrix multiplication while keeping the analytical PF-A contribution and learned correction separately inspectable.

Parameters:

  • condition_input_size (int)

  • harmonic_order_list (list[int])

  • angular_sample_count (int)

  • angular_architecture (str)

  • residual_bound_list (list[float])

  • rank (int)

  • condition_hidden_size (int)

  • angular_hidden_size (int)

  • fourier_feature_order_list (list[int] | None)

  • siren_omega_0 (float)

SUPPORTED_ANGULAR_ARCHITECTURE_SET = {'coordinate_tanh', 'fourier_feature_tanh', 'raw_circular_tanh', 'siren'}
__init__(condition_input_size, harmonic_order_list, angular_sample_count, angular_architecture, residual_bound_list, *, rank=12, condition_hidden_size=64, angular_hidden_size=64, fourier_feature_order_list=None, siren_omega_0=30.0)[source]

Initialize one low-rank bounded coordinate residual.

Parameters:

  • condition_input_size (int) – Number of causal normalized setpoint inputs.

  • harmonic_order_list (list[int]) – PF-A coefficient reconstruction orders.

  • angular_sample_count (int) – Uniform samples in one complete cycle.

  • angular_architecture (str) – Angular basis implementation.

  • residual_bound_list (list[float]) – Training-only physical-unit bounds per angle.

  • rank (int) – Shared low-rank residual dimension.

  • condition_hidden_size (int) – Condition-network hidden width.

  • angular_hidden_size (int) – Angular-network hidden width.

  • fourier_feature_order_list (list[int] | None) – Frozen angular feature orders.

  • siren_omega_0 (float) – SIREN activation frequency scale.

Return type:

None

reconstruct_anchor_curve(anchor_coefficient_tensor)[source]

Reconstruct PF-A on the immutable angular grid.

Parameters:

anchor_coefficient_tensor (Tensor)

Return type:

Tensor

forward(condition_tensor, anchor_coefficient_tensor)[source]

Return PF-A, bounded residual, and complete curve prediction.

Parameters:

  • condition_tensor (Tensor)

  • anchor_coefficient_tensor (Tensor)

Return type:

dict[str, Tensor]