Source-family Fault PiNet for spatial solver coupling

Source-family Fault PiNet for spatial solver coupling

The original synthetic material family in fault_pinet.py varies all four coefficients and does not represent the attached paper’s source cases. This separate experiment trains the same bounded [1,12,12,4] network on the material family a_source in [0.006,0.018], b=0.012, L=0.020 m, and f_ref=0.600. It preserves the original results/ directory.

This is a learned constitutive material-family representation. The prescribed source coefficient enters as a fixed descriptor, z=(a_source-0.012)/0.006. The network predicts all four effective coefficients; the implementation does not replace any neural output by the teacher value. The current slip rate and state are excluded from the network inputs. The existing rate-state potential, traction, aging update and admissibility arguments are therefore unchanged.

This experiment does not learn unknown friction physics or a space-time rupture operator. A spatial solver evaluates the network once for each fixed material field, then uses the resulting coefficients in its constitutive updates. A speedup of the rupture dynamics does not follow from this implementation.

Reproduce

python train_source_family.py --maxiter 5000 --seeds 10 11 12
python test_source_family.py

Outputs are in results_rupture/source_family/. source_model.npz is selected solely by the lowest whole-material-group validation loss across checkpoints and the three seeds. config.json records all groups, bounds, sample seeds, units, software versions and optimizer settings; metrics.json reports actual results. Training receives traction and state-step labels, not parameter labels. The teacher data are noiseless and exactly follow the prescribed constitutive prior.

There are 97 equally spaced material groups: 61 training, 16 validation and 20 test groups. Each training/validation group has 48 samples; each test group has 128 samples. The six actual central/outer source coefficients 0.0065, 0.007, 0.008, 0.016, 0.017, 0.0175 are held out from both training and validation. The signed slip-speed magnitude covers 1e-12 ... 10 m/s; the state covers 1e-3 ... 1e24 s, sampled uniformly in base-10 logarithm. Broad state coverage is necessary because inversion of the source’s initial traction at 1e-12 m/s produces very large kinetic-state values. These are internal variables, not literal geological ages.

The loss is the mean of ((mu_pred-mu)/0.1)^2 plus log(theta_next_pred/theta_next)^2. The frozen-speed step uses h=u*0.020/abs(V) with uniform u in [0.05,2]. Future-state supervision resolves the friction-only confounding between reference friction and slip distance.

Solver interface

from source_family_pinet import load_source_model, source_family_parameters
from fault_pinet import traction, aging_update

model = load_source_model()
parameters = source_family_parameters(model, a_field)  # shape a_field.shape+(4,)
tau = traction(velocity, theta, effective_pressure, parameters)
theta_next = aging_update(theta, velocity, dt, parameters[..., 2])

The last parameter axis is [a, b, L_m, f_ref]. The material descriptor must lie inside the trained range; the interface rejects values outside it rather than silently clipping them. source_truth is provided only for the analytic reference and must not be used in a learned-model solver branch.

For a native solver:

from source_family_pinet import export_source_field
export_source_field("neural_coefficients.npy", a_field, model=model)

The NPY file stores a little-endian float64 array in C row-major order, with the four coefficients contiguous at each site. example_coefficients.npy and source_case_coefficients.csv demonstrate the field shape and all six actual central/outer values. The exporter always evaluates the trained weights.

The source model is scalar. A vector dynamic-rupture solver must apply these scalar friction coefficients to its established isotropic friction law using slip-speed magnitude and correctly oriented resisting traction. Its bulk solver, spectral history, wave propagation, boundary conditions, state integration, convergence and sign conventions require independent validation.

Locked training results

Seed 12 was selected using validation, at iteration 4600. All three training runs ended with optimizer convergence status. The held-out traction relative L2 error is 2.1821918e-6; the RMSE at 120 MPa pressure is 0.0002151625 MPa. A 4001-point material grid has maximum absolute errors [3.6502322e-7, 2.2789395e-7, 3.8436675e-7 m, 3.4591396e-6] in [a,b,L,f_ref].

The initialization audit spans shear stresses 45–105 MPa and the entire material range. Its analytic initial states span approximately 0.144 ... 1.793e23 s, inside the training state range. The maximum absolute logarithmic initial-state ratio is 6.955978e-4. This checks a source initialization sensitivity; it does not establish trajectory or rupture-front accuracy.

Selected checkpoint SHA-256: 7307121720ef68469f97056bddc3abead11878cf73e21bf964af8286ee57405d.

The checkpoint round trip reproduces neural coefficients exactly. See metrics.json for complete seed-specific results and numerical precision.