Multiscale convection-diffusion-reaction (CDR) systems couple spatial transport with local stiff reaction kinetics, making their numerical simulation computationally demanding. Conventional neural surrogates require costly coupled CDR training data, and their parameter counts scale with the number of transported variables, limiting scalability to high-dimensional CDR systems. Here we propose a PDE-Structured Transport Operator (PSTO) that learns convection--diffusion dynamics solely from reaction-free data. By embedding the PDE structure into the neural architecture, PSTO generalizes to full CDR systems at inference without retraining. A single transport operator block is shared across all transported variables and refined through relaxed fixed-point iterations, performing an implicit-like correction in latent space. This improves stability during recursive multi-step prediction without adding trainable parameters. We evaluate PSTO on three benchmarks: a canonical 2D CDR system, a multiscale Robertson counterflow flame, and a multiscale hydrogen planar jet flame. In every case, PSTO reduces the mean relative L2 error after 100 coupled steps by at least 50% compared to the strongest neural baseline. Ablation studies on the fixed-point iterations and detailed error decomposition further confirm that PSTO achieves both high one-step accuracy and long-term stability in coupled rollouts. Time-step sensitivity analysis shows that PSTO maintains accuracy even with a tenfold larger time step, whereas baseline accuracies decay rapidly as the time step increases. This framework enables efficient, modular modeling of multiscale CDR dynamics by separating transport learning from the treatment of local kinetics.
Figures & tables
Figure 1: PSTO architecture. (a) A coupled Lie–Trotter transport–reaction prediction step with shared lifting map P , transport block, and projection map Q . The coefficient and reaction MLPs are trained independently, and their parameters remain fixed during PSTO training. Orange arrows indicate coefficient evaluation, black arrows indicate transport evolution, and red arrows indicate reaction coupling. The transport output follows Eq. ( 6 ), with the subtraction of z0 before projection and the addition of qn afterward omitted from the schematic. (b) The conditioned advection and diffusion branches and the relaxed fixed-point update defined in Eq. ( 8 ). The flux-divergence operation incorporates the variable-specific transport coefficient and density weighting in Eq. ( 5 ). The initial features z0 and the conditioning fields remain fixed throughout the L latent iterations.
Model
1
5
10
20
50
100
PSTO
0.012
0.056
0.110
0.211
0.477
0.893
FNO
0.013
0.061
0.116
0.239
0.910
2.82
U-Net
0.008
0.034
0.060
0.128
0.599
2.03
U-FNO
0.009
0.039
0.069
0.145
0.719
2.49
Table 1: Mean variable-wise relative L2 error (%) for the canonical CDR system, evaluated using Eq. ( 24 ). Columns indicate the number of coupled steps.
Figure 2: Canonical CDR fields after 100 coupled steps ( t−t0=0.025s ). Rows show u , v , and ϕ , and columns show the numerical reference, PSTO, FNO, U-Net, and U-FNO. All surrogate models start from the same initial state and include the same prescribed polynomial source terms.
Figure 3: Canonical CDR residual defined by Eq. ( 27 ). (a) Endpoint residual fields on a common color scale. (b) Spatial 95th percentile over 100 coupled steps. The diagnostic includes the full governing equations of the canonical CDR system.
Figure 4: Distribution shift between Robertson CD training states (blue) and CDR evaluation states (red), visualized through a joint t-SNE embedding of 100 snapshots per regime. The cross-regime MMD, computed in the pooled PCA representation, is 0.709.
Model
1
5
10
20
50
100
PSTO
0.364
0.914
1.04
1.13
1.30
1.50
FNO
0.640
2.05
2.75
3.84
7.09
12.4
U-Net
0.641
2.04
2.71
3.72
6.51
10.6
U-FNO
0.662
2.14
2.93
4.16
7.61
12.6
Table 2: Mean variable-wise relative L2 error (%) for Robertson counterflow, evaluated using Eq. ( 24 ). Columns indicate the number of coupled steps.
Figure 5: Robertson counterflow mass fractions after 100 coupled steps ( t−t0=0.01s ). Columns show YA,YB,YC,YD , and rows show the numerical reference, PSTO, FNO, U-Net, and U-FNO. All models use the same velocity field, inlet data, and reaction kinetics.
Model
1
5
10
20
50
100
PSTO
1.03
3.55
5.79
9.28
21.4
23.3
FNO
1.45
6.35
12.7
23.5
44.4
51.4
U-Net
1.74
7.57
12.6
20.9
40.9
50.5
U-FNO
1.49
7.26
12.6
21.2
41.0
48.5
Table 3: Mean variable-wise relative L2 error (%) for the hydrogen planar jet flame, evaluated using Eq. ( 24 ). Columns indicate the number of coupled steps.
Figure 6: Planar jet flame fields after 100 coupled steps ( t−t0=10−4s ). Columns show T , YH2O , u , and v , and rows show the numerical reference, PSTO, FNO, U-Net, and U-FNO. All surrogate models use the same initial state and Lie–Trotter transport–reaction coupling.
Figure 7: Iterative refinement and illustrative diffusion-mode analysis. (a) Normalized fixed-point residual rℓ/r0 evaluated at zℓ and (b) one-step relative error averaged over variables, evaluated on 64 Robertson CD validation pairs with trained weights held fixed. Lines indicate medians and bands represent interquartile ranges. (c) Amplification factors for explicit Euler, implicit Euler, and four relaxed iterations at ω=0.1718 , plotted against ξ=Δtλmode .
Figure 8: One-step error decomposition, transport sensitivity, and autoregressive error propagation for the planar jet flame. (a) Transport, reaction, and splitting contributions to the one-step error. The measured errors and summed contribution norms are shown for comparison. (b) Local transport sensitivity to small perturbations at successive reference states. The dashed line marks Λnloc=1 , indicating unchanged perturbation magnitude after one transport update. (c) Relative autoregressive errors en and recursive bounds Bn .
Figure 9: Computational performance for the planar jet flame over a physical horizon of 10−4s , corresponding to 100 coupled steps at Δt=10−6s for the neural models. (a) Coupled inference wall time averaged over five timed repetitions for the neural models, together with the runtime of the single OpenFOAM run. (b) Mean variable-wise relative L2 error at the common endpoint versus wall time. The neural models use a single RTX 4090 GPU, whereas OpenFOAM uses 128 CPU MPI ranks. OpenFOAM serves as the numerical reference and is therefore shown at zero error by definition.
Model
1
2
5
10
PSTO
5.96
9.16
21.0
23.4
FNO
19.5
33.7
68.7
74.7
U-Net
13.0
23.7
46.5
57.2
U-FNO
13.3
23.9
48.0
59.0
Table 4: Mean variable-wise relative L2 errors (%) for the hydrogen planar jet flame at Δt=10−5s , evaluated using Eq. ( 24 ). Column headings indicate the number of coupled steps. The corresponding physical horizons match those of steps 10, 20, 50, and 100 at Δt=10−6s in Table 3 . All four transport surrogates are coupled to the same reaction surrogate trained for the larger time interval.
Table 5: Mean coupled inference wall time (s) for the hydrogen planar jet flame over a common physical horizon of 10−4s . The two time steps require 100 and 10 coupled steps, respectively. Measurements follow the protocol in Section 4.4.3 , using a single RTX 4090 GPU, two warmup rollouts, and five timed repetitions.
Figure 10: Time-step dependence of coupled planar jet flame predictions over a common physical horizon of 10−4s , starting from the same initial state. (a) Coupled inference wall time at Δt=10−6s (left, opaque) and Δt=10−5s (right, translucent) within each model pair. Bars show mean wall times over five timed repetitions on a single RTX 4090 GPU. (b) Mean variable-wise relative L2 error at the common endpoint versus wall time. Colors identify the models, opacity distinguishes the time steps, and dashed lines connect the corresponding pairs.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure B.1: Initial and inlet conditions for the hydrogen planar jet flame. (a) Hydrogen mass fraction in the developed reacting state used to initialize coupled prediction. (b) Streamwise inlet velocity profile, with a faster central jet and slower air streams on both sides. The training and test compositions are specified in Section 4.3 .
Variable
PSTO
FNO
U-Net
U-FNO
u
1.44
5.64
3.26
4.25
v
1.10
2.28
2.12
2.53
ϕ
0.136
0.553
0.723
0.692
Appendix
Table C.1: Variable-wise relative L2 errors (%) for the canonical CDR system after 100 coupled steps.
Variable
PSTO
FNO
U-Net
U-FNO
YA
0.371
6.90
3.85
3.57
YB
3.70
7.80
6.61
6.70
YC
1.11
20.5
26.2
29.8
YD
0.821
14.5
5.58
10.5
Appendix
Table C.2: Relative L2 errors (%) in the four predicted species mass fractions for the Robertson counterflow benchmark after 100 coupled steps.
Variable
PSTO
FNO
U-Net
U-FNO
T
5.28
26.1
18.9
19.0
YH2
2.82
8.85
7.84
7.30
YH
36.0
85.8
73.9
71.6
YO
54.1
100
128
113
YO2
1.97
7.23
6.89
7.04
YOH
38.3
88.7
88.4
83.8
Appendix
Table C.3: Variable-wise relative L2 errors (%) in the twelve predicted fields for the planar jet flame after 100 coupled steps at Δt=10−6s , corresponding to a physical horizon of 10−4s .
Variable
PSTO
FNO
U-Net
U-FNO
T
6.37
29.7
22.1
32.1
YH2
6.63
11.4
7.34
35.8
YH
41.9
314
131
130
YO
50.5
106
93.1
100
YO2
2.02
12.5
10.3
9.18
YOH
34.2
111
86.3
89.3
Appendix
Table C.4: Variable-wise relative L2 errors (%) in the twelve predicted fields for the planar jet flame after 10 coupled steps at Δt=10−5s , corresponding to a physical horizon of 10−4s . All four transport surrogates are coupled to the same reaction surrogate trained for this time interval.
Figure D.1: Normalized pointwise absolute error in ϕ for canonical CDR after 100 coupled steps [Eq. ( D.1 )]. All models use the same error color scale.
Figure D.2: Normalized pointwise absolute error in YC for Robertson counterflow after 100 coupled steps [Eq. ( D.1 )]. All models use the same error color scale.
Figure D.3: Normalized pointwise absolute error in T for the planar jet flame after 100 coupled steps [Eq. ( D.1 )]. All models use the same error color scale.
Case
Model
Parameters
m
wnet
Larch
Canonical CDR
PSTO
197,841
16
8
4
FNO
1,640,223
16
20
4
U-Net
3,037,743
—
20
4
U-FNO
2,013,183
16
20
4
Robertson counterflow
PSTO
197,841
16
8
4
FNO
1,640,304
16
20
4
Appendix
Table E.1: Transport architecture settings and trainable parameter counts. The real and imaginary components of complex Fourier weights are counted separately. Frozen constitutive and reaction networks are excluded.
Case
Training pairs
Validation pairs
Epochs
Learning-rate schedule
Canonical CDR
1600
400
100
Linear warmup + cosine decay
Robertson counterflow
4000
1000
40
Linear warmup + cosine decay
Planar jet flame
4000
1000
40
Linear warmup + cosine decay
Appendix
Table E.2: Reaction-free CD datasets and training schedules for the four transport surrogate models.
Setting
Value
Optimizer
AdamW
Initial learning rate
3×10−4
Final learning rate
10−6
Warmup
100 optimizer updates, linear
Schedule after warmup
Cosine decay
Weight decay
10−5 on nonspectral parameters
Appendix
Table E.3: Optimizer settings shared across the three cases.
Case
λROI
λwave
λPDE
λmass
λneg
Canonical CDR
0.1
0.02
0.1
—
—
Robertson counterflow
0.1
0.02
0.1
1
1
Planar jet flame
0.1
0.02
0.1
1
1
Appendix
Table E.4: Case-specific loss weights shared by all four transport surrogate models.
Figure E.1: Reaction-free CD validation loss versus training epoch for (a) canonical CDR, (b) Robertson counterflow and (c) the planar jet flame. The checkpoint with the lowest validation loss is selected for coupled evaluation. Loss values depend on the fields and applicable loss terms in each case.