JAX-WAVE: A Differentiable Physics-Supervised Framework for Inverse Wave Design in Heterogeneous Media
Authors: Honghui Du, Binyao Guo, Yushu Pan
Organizations: Department of Civil, Environmental, and Geo- Engineering, University of Minnesota, Minneapolis, MN, USA. · Department of Computer Science, Columbia University, New York, NY, USA.
Designing a prescribed wave field in a heterogeneous medium requires an initial pressure distribution that compensates for the medium's propagation response. Repeated numerical inversion solves this task afresh for each target and environment. We introduce JAX-WAVE for amortized inverse wave design: a shared neural network maps desired terminal fields and known wave-speed maps directly to bounded initial pressures. Training through a differentiable finite-difference time-domain solver uses the physical design objective without requiring optimized control labels. Linear propagation and symmetric pressure bounds imply an exact sign-reversal relation for the regularized optimum. A shared-weight projection enforces this relation, preserves feasibility, and satisfies an exact identity for the paired-sign objective. We characterize the discrete propagator and its weighted adjoint to connect learning, numerical inversion, and optimality assessment. On an independent confirmation set, three trained models achieve unseen-material errors of 25.24% for geometric targets and 4.77% for reachable targets. Matched-capacity comparisons show improved reconstruction with the sign parameterization. Within a 5 ms median online budget, the learned map reduces mean error by 45% relative to accelerated numerical inversion on the heterogeneous benchmark; longer numerical solves recover greater accuracy. The results establish a reusable inverse-design procedure and show how its accuracy and computational value depend on target structure and query workload.
Figures & tables
Figure 1: JAX-WAVE training and prediction. The desired terminal field and known speed map condition a bounded initial-pressure prediction. During training, the FDTD solver propagates this control and supplies the field loss; gradients pass through the solver to the network. During prediction, the trained network generates the control directly. A subsequent propagation can evaluate the resulting field.
Known materials
Unseen materials
Model
Geometry
Reachable
Geometry
Reachable
Compact ( w=16 )
30.77±0.048
8.17±0.005
30.64±0.066
8.23±0.042
Selected ( w=64 )
25.32±0.039
4.69±0.023
25.24±0.050
4.77±0.023
Table 1: Independent confirmation: mean and sample standard deviation across three training seeds, each with 60,000 updates, with 512 cases per split/class. Both models are symmetrized; errors are percentages. Costs on A40 are mean time per Adam update and per-model inference latency medians; batch-16 time covers the whole batch. Measurement procedures are given in Supplementary Material S1.
Figure 2: Preselected unseen confirmation cases, one per target/material combination. Columns show speed, target, selected initial pressure, compact and selected terminal fields, and selected absolute residual. Both models are symmetrized; seed-zero relative errors are percentages. Signed pressure shares one scale per row. Residuals are normalized by each target’s peak absolute amplitude and use a separate common scale; speeds range from 0.65 to 1.5.
Known materials
Unseen materials
Update time
Model
Geometry (%)
Reachable (%)
Geometry (%)
Reachable (%)
(ms)
Plain ( w=64 )
27.89±0.043
6.58±0.055
27.70±0.053
6.67±0.043
21.83±0.04
Symmetrized ( w=64 )
25.31±0.038
4.68±0.020
25.22±0.052
4.76±0.024
38.07±0.17
Table 2: Paired comparison of the sign parameterization: both models have 4,832,705 parameters, 65,536 training pairs and 60,000 updates. Mean and sample standard deviation across three seeds; 512 cases per split/class. This additional paired analysis reuses the previously evaluated confirmation cases. Update costs use the common A40 measurement protocol in Supplementary Material S1.
Figure 3: Paired validation mean squared relative field error for width-64 models. The update panel reports the mean and sample standard deviation across three seeds, including initialization. The time panel shows each run against its synchronized A40 optimizer time, defined in Section 3.2 .
w
Parameters
e2
Geometry (%)
Reachable (%)
Update (ms)
64
4,832,705
0.03219
24.22
4.59
38.01
96
10,870,945
0.03063
23.66
4.13
66.86
128
19,323,777
0.02997
23.42
3.96
93.96
Table 3: Exploratory expansion beyond width 64 on the original 256 validation cases. Each model is trained from scratch with seed zero, 65,536 pairs, and 60,000 updates. Errors are means of relative terminal-field errors; e2 selects checkpoints. A40 update costs follow the measurement procedure in Supplementary Material S1.
Smooth media
Six inclusions
Model
Geometry
Reachable
Geometry
Reachable
Compact ( w=16 )
38.92±0.200
13.60±0.096
31.72±0.112
9.79±0.058
Selected ( w=64 )
35.17±0.187
9.63±0.166
26.69±0.028
6.02±0.076
Table 4: Transfer of the confirmation models to smooth and six-inclusion media. Errors (%) report three-seed mean and sample standard deviation; each family has 64 geometric and 64 reachable cases. No adaptation or selection uses these cases.
Figure 4: Free evolution of two width-64 controls for unseen geometric/layered and reachable/inclusion cases, selected independently of prediction quality. Each row uses one pressure scale; speed ranges from 0.65 to 1.5.
Control
Target
n=64
n=129
n=259
Compact
Geometry
30.58±0.050
30.79±0.043
30.83±0.045
Selected
Geometry
25.37±0.038
25.80±0.027
25.89±0.035
Numerical
Geometry
14.15
29.18
31.16
Compact
Reachable
7.17±0.112
7.14±0.108
7.14±0.108
Selected
Reachable
4.30±0.056
4.32±0.058
4.34±0.056
Numerical
Reachable
0.21
0.94
1.05
Table 5: Fixed-control refinement on the same unseen reference cases, 16 per target class. Neural errors (%) report mean and sample standard deviation across the three frozen confirmation seeds. Errors are evaluated at shared coarse-grid nodes. Numerical controls are independently certified on the coarse grid and then held fixed.
Method
Geometry (%)
Reachable (%)
Pooled (%)
Reduction (%) [95% CI]
Projected gradient
45.65
31.94
36.51
48.5 [44.3,52.6]
FISTA
44.03
29.25
34.18
45.0 [40.4,49.4]
Restarted FISTA
44.03
29.25
34.18
45.0 [40.4,49.4]
JAX-WAVE
31.87
12.25
18.79
—
Table 6: Independent heterogeneous-media benchmark within a 5 ms median online computation budget on A40. Lower mean terminal-field error is better. The 32 geometric and 64 reachable cases comprise six equally weighted conditions. Neural errors average three separately evaluated seeds; numerical budgets use timing-only iteration selection. The final column reports relative error reduction by JAX-WAVE against each baseline, with paired 95% bootstrap confidence intervals (CI).
Figure 5: Accuracy versus control-generation time on the independent heterogeneous-media benchmark. Panels separate smooth geometric and reachable targets. Numerical curves use uniform iteration counts; stars show the three frozen networks. The dashed line marks 5 ms. Table 6 selects iteration counts per condition using timing alone. Measurements use synchronized A40 execution under the control-generation timing protocol in Section 3.4 , which excludes terminal-field verification.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Method
5 ms
10 ms
20 ms
JAX-WAVE
18.79
18.79
18.79
Projected gradient
36.51
24.48
17.64
FISTA
34.18
18.71
11.84
Restarted FISTA
34.18
18.72
11.84
Appendix
Table S1: Mean field error (%) on the same 96 heterogeneous confirmation cases. JAX-WAVE predicts once without correction; extra budget is used only by the numerical methods.
Figure S1: Relative field-error reduction by JAX-WAVE over each numerical baseline on the 96 heterogeneous cases. Positive values favor the network. Bars show paired 95% bootstrap intervals using descriptor-ID resampling stratified by material family. The 5 ms comparison is primary; 10 and 20 ms are secondary.
Figure S2: Terminal fields for descriptor zero at contrast 1.5 in each confirmation target family (all layered media). Columns show the target, JAX-WAVE seed zero, FISTA within 5 ms, and the 4,096-step reference. Relative L2 errors are percentages; each row shares a pressure scale. Selection uses descriptor ID, while aggregate comparisons retain all heterogeneous cases and three networks.
Figure S3: Accuracy–latency tradeoffs on eight fixed cases in one A40 session. Points and bars show median and interquartile range of per-case latency medians. The right panel enlarges the three overlapping neural results. Each solver curve uses 8/16/32/64 zero-start FISTA updates from left to right. New-material dense costs add each case’s assembly and solve times before aggregation; cached costs exclude assembly. Compilation, transfers, training and final-forward verification are excluded.
Method
Error (%)
Time (ms)
IQR (ms)
Frozen compact
16.59
0.402
0.402–0.403
Frozen selected
12.28
3.253
3.242–3.267
Paired plain
14.39
2.199
2.180–2.201
Paired symmetrized
12.09
3.265
3.258–3.272
Conservative FISTA, 16
17.79
12.049
12.008–12.068
Conservative FISTA, 32
9.65
23.903
23.858–23.910
Appendix
Table S2: Same-session A40 control-query comparison on eight stratified cases. Error is the mean over those cases; latency is the median of their synchronized repeat medians, with across-case IQR. Neural rows use seed zero and batched signs where applicable. Cached methods exclude assembly.
Known materials
Unseen materials
Method
Geometry
Reachable
Geometry
Reachable
U-Net
33.72±0.147
10.95±0.070
33.92±0.151
11.20±0.050
Symmetrized U-Net
30.17±0.144
8.20±0.063
30.55±0.075
8.26±0.070
Spatial odd
31.30±0.004
8.22±0.057
31.80±0.036
8.24±0.045
Appendix
Table S3: Mean relative field error (%) across three training seeds, with sample standard deviation of the per-seed mean. Each run uses the same data and 60,000 updates.
Figure S4: Unseen-material confirmation error distributions. Each line is one frozen training seed evaluated on the same 512 cases per target class; colors distinguish compact and selected models. Dashed lines mark the 90th percentile; annotations average the three per-seed percentiles.
Parameters
Known materials
Unseen materials
Method
Geometry
Reachable
Geometry
Reachable
U-Net
302,705
33.89
11.02
34.09
11.21
Wide U-Net
472,701
32.46
10.06
32.74
10.27
Sign augmentation
302,705
34.77
11.13
34.97
11.40
Spatial affine
508,176
32.23
9.52
32.56
9.70
Spatial odd
477,904
31.30
8.16
31.77
8.20
Appendix
Table S4: Mean relative field error (%) on held-out targets. Each material split contains 512 geometric and 512 reachable targets. This six-model comparison uses seed zero, the same training data and 60,000 updates. Bold entries identify the lowest error in each column.
w
Training pairs
Updates
Parameters
e2
Geometry
Reachable
16
16,384
60,000
302,705
0.04897
29.55
7.90
16
65,536
60,000
302,705
0.04858
29.47
7.60
24
65,536
60,000
680,425
0.04206
27.52
6.54
32
16,384
60,000
1,209,057
0.04070
27.05
6.39
32
65,536
60,000
1,209,057
0.03833
26.33
5.87
48
65,536
60,000
2,719,057
0.03421
24.95
5.05
Appendix
Table S5: Symmetrized-model screening on 256 validation cases, seed zero. e2 selects checkpoints; the final columns report mean relative field errors (%). Independent confirmation is reported in the main paper.
Target
λ
Error (%)
Control RMS
Neural
Numerical
Neural
Numerical
Geometry
2×10−5
29.40
13.57
0.3085
0.4488
Geometry
2×10−4
29.42
13.83
0.3070
0.4300
Geometry
2×10−3
29.62
15.64
0.2934
0.3757
Reachable
2×10−5
7.92
0.04
0.1949
0.2030
Reachable
2×10−4
7.92
0.25
0.1941
0.2008
Appendix
Table S6: Regularization sensitivity on the same 64 reference cases, 32 per target class. Neural results use the symmetrized network trained at each regularization with seed zero; numerical references use the matching objective. Errors are percentages; RMS denotes nodal control root mean square.
Smooth media
Six inclusions
Method
Geometry
Reachable
Geometry
Reachable
U-Net
41.57
16.67
35.23
12.50
Symmetrized U-Net
39.03
13.50
31.86
9.75
Spatial odd
39.62
12.73
33.23
9.53
Appendix
Table S7: Mean field error (%) in the exploratory material-family probe. Each family contains 64 geometric and 64 reachable targets. All three models retain their original seed-zero weights.
Control
Target
64
129
259
Symmetrized
Geometry
30.72
30.95
31.00
Spatial odd
Geometry
31.93
32.15
32.20
Numerical
Geometry
14.15
29.18
31.16
Symmetrized
Reachable
7.20
7.16
7.16
Spatial odd
Reachable
7.15
7.15
7.16
Numerical
Reachable
0.21
0.94
1.05
Appendix
Table S8: Mean relative error (%) under refined propagation of fixed controls on unseen materials. Each target class contains 16 cases. Errors are evaluated at shared coarse-grid nodes. Numerical controls are obtained for the coarse-grid regularized objective.
Method
Error (%)
Time (ms)
IQR (ms)
U-Net
20.56
0.409
0.407–0.413
Symmetrized U-Net
16.57
0.420
0.417–0.424
Conservative FISTA, 8
30.36
6.076
6.060–6.089
Conservative FISTA, 16
17.79
11.991
11.956–12.015
Conservative FISTA, 32
9.65
23.802
23.770–23.915
Backtracking FISTA, 8
20.14
9.659
9.493–9.848
Appendix
Table S9: Compact architecture-study models and numerical solvers on eight stratified cases in the same A40 session. Time is the median of eight within-case medians, each from five synchronized repeats; the IQR describes across-case variability. Cached methods exclude matrix construction.
Basque Center for Applied Mathematics (BCAM), Bilbao, Spain · University of the Basque Country (UPV/EHU), Leioa, Spain · Ikerbasque: Basque Foundation for Science, Bilbao, Spain