Operator learning approximates the solution map of a parametric partial differential equation from data, enabling fast prediction at new parameter values without repeated simulation. The dominant approach fits a single global low-rank basis, typically a proper orthogonal decomposition (POD), to all solution snapshots. This fails when the snapshots span qualitatively different solution shapes, from smooth diffusive profiles to near-discontinuous shocks. We propose TEMPO (Trajectory Expectation-Maximization Mixture of POD Operators), a two-phase framework for regime-aware operator learning. In the offline phase, TEMPO recovers latent regimes via expectation-maximization, fitting a dedicated linear or nonlinear basis to each discovered regime. In the online phase, a gating network and per-regime branch networks are trained jointly, routing new inputs to the regimes at inference. We test TEMPO on Burgers' flow, Sod shock tubes, Coanda flow, Darcy flow, Navier-Stokes flow and the heterogeneous Helmholtz equation. With 64 total POD modes, TEMPO gives 18.61% test error on Burgers', compared with 19.61% for global POD. It also gives lower errors on Coanda and Helmholtz.
Figures & tables
Figure 1: Training dataset D : solution trajectories s(i) on Ω×T for different κ(i)∈K .
Figure 2: Offline reconstruction using the TEMPO(POD) representation: M regimes with K modes each ( MK total), versus K global POD modes. Left: normalized error; right: relative error reduction.
Figure 3: Mean relative L2 test errors.
Figure 4: TEMPO(POD) regimes. UMAP (Uniform Manifold Approximation and Projection) colored by regime (left) and physical quantity (middle); POD coefficients (right). Diamonds mark centers; ellipses show two standard deviations.
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Train ν
Modes
ν=0.001
ν=0.1
ν=1
Mean
Time (ms)
Single-viscosity training (original split)
DeepONet (0.553M)
0.001
–
53.563
47.407
38.173
46.381
0.429
DeepONet (0.553M)
0.1
–
53.451
47.110
34.918
45.160
0.429
DeepONet (0.553M)
1.0
–
54.729
48.300
28.274
43.768
0.429
POD-DeepONet (0.468M)
0.001
32
27.266
29.697
63.896
40.286
0.123
POD-DeepONet (0.468M)
0.01
32
27.438
27.377
61.436
38.750
0.126
Appendix
Table A.1 : Relative L2 test errors (%) on Burgers’. Underlined: training viscosity. Bold: best in the common-test comparison.
Figure A.1 : Burgers’: normalized POD spectrum (left) and normalized squared NeuralPOD reconstruction residual (right).
Figure A.2 : First eight Burgers’ basis functions, scaled to unit peak magnitude: POD (top) and NeuralPOD (bottom).
Figure A.3 : Burgers’: reference solution, TEMPO prediction and absolute error for one initial condition. First three rows: POD at ν=0.001,0.1,1 . Last row: NeuralPOD at ν=0.1 .
Method
Modes
Error (%)
Time (ms)
TEMPO(POD) (11,866)
2×24
3.270
0.368(0.091)
TEMPO(NeuralPOD) (1.207M)
3×16
4.243
0.368(0.090)
DeepONet (0.529M)
–
9.405±0.939
0.106
POD-DeepONet ( K ) (0.037M)
24
3.989±0.002
0.110
POD-DeepONet (11,952)
48
3.369
0.157
NeuralPOD-DeepONet (1.168M)
48
6.527
0.135
Appendix
Table A.2: Relative L2 test errors (%) for Euler shock tubes.
Figure A.4 : Sod shock tubes: reference, TEMPO(POD) and TEMPO(NeuralPOD) predictions, and absolute errors. First three rows: ρ , u and p at r=0.1 . Last two rows: ρ at r=0.05 and 0.4 .
Method
Modes
Error (%)
Time (ms)
TEMPO(POD) (0.104M)
2×12
0.585±0.065
0.229(0.065)
TEMPO(NeuralPOD) (0.646M)
2×12
0.680
0.304(0.092)
DeepONet (0.528M)
–
15.870±1.526
0.125
POD-DeepONet ( K ) (0.035M)
12
1.032±0.129
0.101
POD-DeepONet (0.107M)
24
0.651
0.138
NeuralPOD-DeepONet (0.628M)
24
9.117
0.145
Appendix
Table A.3: Relative L2 velocity errors (%) for Coanda flow.
Figure A.5 : Coanda flow: reference, TEMPO(POD) and TEMPO(NeuralPOD) transverse velocities, followed by their absolute errors.
Figure A.6 : Darcy pressure at β=0.01 : reference, TEMPO(POD), TEMPO(NeuralPOD) and their absolute errors.
Method
Train Re
Test Re
Infer. (ms)
100
1000
3600
10000
Per-Re specialists (single-Re training)
DeepONet [ 4 ] (2.691M)
100
1.000
1.002
1.003
1.003
0.381
1000
1.000
1.000
1.000
1.000
0.381
3600
1.001
1.000
1.000
1.000
0.381
10000
1.001
1.000
1.000
1.000
0.381
Appendix
Table A.5 : Relative L2 test errors on 2D Navier–Stokes. Underlined: in-distribution; bold: best joint model per column.
Figure A.7 : Navier–Stokes flow at Re=100 : reference, TEMPO(POD) and TEMPO(NeuralPOD) predictions at t=1 (left) and t=2 (right). Rows show u and v .
Method
Modes
Error (%)
Time (ms)
TEMPO(POD) (6.704M)
4×16
3.605
0.553(0.095)
TEMPO(NeuralPOD) (8.123M)
4×16
5.832
0.470(0.185)
DeepONet (3.140M)
–
119.888
0.130
POD-DeepONet ( K ) (1.341M)
16
11.598
0.113
POD-DeepONet (6.579M)
64
14.865
0.154
NeuralPOD-DeepONet (7.935M)
64
55.165
0.159
Appendix
Table A.6: Relative L2 test errors (%) for Helmholtz.
Figure A.8 : Helmholtz: reference, TEMPO(POD) and TEMPO(NeuralPOD) predictions, and absolute errors for one wavespeed field at four frequencies.
Task ( Nval )
Basis
Modes
Best local LS
Prediction
Outside share (%)
Burgers’ (1200)
POD
2×32
14.7901
19.5692
54.44
NeuralPOD
3×16
23.3221
22.8662
80.65
Sod (408)
POD
2×24
3.1633
3.2113
88.74
NeuralPOD
3×16
4.1899
4.1909
89.06
Coanda (62)
POD
2×12
0.1396
0.8078
1.95
NeuralPOD
2×12
0.6915
0.8026
53.09
Appendix
Table A.7: Validation errors (%). Outside share is the fraction of squared relative prediction error outside the combined mode span (excluding mean fields).
Task
Before
After
Burgers’
26.601
21.534
Sod
5.967
4.243
Coanda
0.698
0.680
Navier–Stokes
9.131
8.956
Helmholtz
17.984
5.832
Appendix
Table A.8: Mean relative L2 test errors (%) of TEMPO(NeuralPOD) before and after additional training.
Model
Modes
Standard
Parallel heads
CUDA graph
TEMPO(POD) (1.397M)
2×32
0.405
0.235
0.184
Global POD (1.346M)
64
0.194
–
0.142
TEMPO(NeuralPOD) (2.917M)
3×16
0.534
–
0.210
Global NeuralPOD (6.449M)
32
0.131
0.139
0.104
Appendix
Table A.9: Burgers’ inference times (ms) with cached bases.
Figure A.9 : Offline EM convergence (top) and online training losses (bottom, first 200 epochs) for Coanda.
Benchmark
Basis
M
EM iters
Skip
Incr.
Full
1D Burgers’
POD
3
19
33%
35%
32%
1D Burgers’
NeuralPOD
3
13
49%
13%
38%
Appendix
Table A.10: Adaptive basis updates on Burgers’.
ν
Regime 1
Regime 2
Regime 3
0.001
0.32
0.28
0.40
0.1
0.32
0.26
0.42
1.0
0.33
0.26
0.41
πm
0.32
0.27
0.41
Appendix
Table A.11: Mean regime weights on Burgers’.
Task
TEMPO(NeuralPOD)
TEMPO(POD)
Burgers’
+11.79
+28.21
Navier–Stokes
+0.70
+0.70
Appendix
Table A.12 : Increase in mean relative L2 error (percentage points) when the parameter is replaced by its training mean.
Task/model
Known parameter
Estimated parameter
Burgers’, TEMPO(POD), M=4
0.32978
0.33283
Navier–Stokes, TEMPO(POD)
0.09648
0.09724
Navier–Stokes, TEMPO(NeuralPOD)
0.08956
0.09039
Appendix
Table A.13: Mean relative L2 errors with known and estimated physical parameters.