AdjointFlow: Accelerating Scientific Inverse Problems with Physics-Based Transport
Authors: Ethan Hanold, Nicholas M. Boffi, Qin Li, Leonardo Zepeda-Núñez
Organizations: University of Wisconsin–Madison, Madison, WI, USA · Carnegie Mellon University, Pittsburgh, PA, USA · Google Research, Mountain View, CA, USA
Inverse problems are ubiquitous across scientific applications ranging from seismic imaging to astrophysics to nondestructive testing, yet classical solvers often struggle with nonconvexity and repeated evaluations of expensive physical models. Diffusion models can amortize this computation offline, but largely data-driven conditional models underutilize known physics and suffer from ill-conditioned transport maps when the target distribution is concentrated. Plug-and-play (PnP) methods explicitly leverage the physical model to guide a learned prior, improving data consistency but reintroducing costly iterative computations. We instead propose AdjointFlow, a framework built on stochastic interpolants that injects the known physics into the generative transport itself. We design a data-dependent, adjoint-based initialization to reduce transport distance to the target, modify the transport velocity to remain regular as the endpoint distributions concentrate, and parameterize the network to build the initialization directly into the learned velocity. Together, these enable accurate reconstruction at substantially lower cost than generative models starting from Gaussian noise. Experiments on inverse scattering, seismic imaging, and black hole imaging show that AdjointFlow consistently produces more accurate reconstructions than conditional diffusion models while being up to five orders of magnitude faster at inference than PnP methods. Optional inference-time refinement that queries the physical model can further improve data consistency, matching or exceeding the accuracy of PnP methods without retraining.
Figures & tables
Figure 1: Overview. Traditional diffusion-based approaches start from uninformative Gaussian noise and can suffer from ill-conditioned transport fields when the target posterior becomes concentrated, such as when more informative measurements are acquired. Conversely, AdjointFlow starts from a data-dependent initialization – for example, one constructed from the adjoint F∗y as shown above – reducing transport distance relative to Gaussian noise and producing favorable learning conditions for the solver. AdjointFlow further leverages a noise-free velocity which remains regular near concentrated endpoints.
Figure 2: Diagram of the different settings and modalities of inverse problems we consider. Top: a diagram of the data acquisition setting, including a representative example of the parameters m represented by the wavespeed (in the case of scattering problems) or brightness (in the case of black holes). Bottom: an example of the measurement data used for the reconstruction. Note that the full aperture scattering is measured in the frequency domain, while the geophysical scattering is measured in the time domain.
Figure 3: Model architecture. A data encoding network w extracts domain-specific features c=w(y) to condition the target-prediction network u(x,t;c) , which then forms a velocity x˙t=u(x,t;c)−x0 to be integrated, with initial state x0=a(y) at t=0 , to produce an estimate x1≈m† at t=1 . Red boxes denote learned components.
Figure 4: Qualitative results on nonlinear inverse scattering for (a) 10h Overlapping Squares , (b) 3–5–10h Triangles , (c) Shepp–Logan , and (d) Brain Scans . Within each panel, columns show distinct test examples; rows show, from top to bottom, the ground-truth perturbation, the adjoint-based initialization, the AdjointFlow reconstruction, and its absolute reconstruction error. The wavelength marker corresponds to 8 points per wavelength (PPW) in (a)–(c) and 5 PPW in (d).
Shepp–Logan
3–5–10h Triangles
10h Overlapping Squares
Method
RRMSE ↓
MELR ↓
SD ↓
RRMSE ↓
MELR ↓
SD ↓
RRMSE ↓
MELR ↓
SD ↓
Traditional
FWI
52.041
241.893
7.658
28.637
159.894
0.809
38.777
281.057
4.773
Deterministic
SwitchNet
4.305
8.071
1.122
15.084
9.377
0.570
20.102
24.295
2.643
WideBNet
3.843
13.255
1.059
17.263
10.582
0.608
14.327
44.182
1.996
Table 1: We compare performance on three synthetic datasets ( Shepp–Logan, 3–5–10h Triangles, 10h Overlapping Squares ) using RRMSE (%), MELR (×10−2) , and Sinkhorn divergence (SD) metrics (see D for definitions, bold indicates best performance for a given metric). All Sinkhorn divergence values, including those of the baselines, are recomputed using an updated library version and therefore differ from those reported in [ 132 ] . We find that AdjointFlow outperforms all models, including conditional diffusion, across the board.
Brain Scans
Method
RRMSE ↓
MELR ↓
SD ↓
PnP Diffusion Prior
DAPS
12.572
18.700
0.278
RED-Diff
2.873
7.370
0.107
Conditional Diffusion
EquiNet CDM
5.661
3.494
0.157
Table 2: We compare AdjointFlow and RED-Flow to PnP-DP and CDM baselines for nonlinear inverse scattering on the Brain Scans dataset, with bold indicating the best performance on each metric. We report RRMSE as a percentage and MELR ×10−2 . EquiNet CDM results are obtained from a model we trained using the pipeline from [ 132 ] . CDM † indicates that we evaluate the diffusion model at the same number of function evaluations as AdjointFlow (namely, 8 sampling steps), thus demonstrating the advantage of AdjointFlow when compute is limited.
Figure 5: Qualitative comparison of reconstructions on the Brain Scans dataset for nonlinear inverse scattering. For CDM † , we evaluate the diffusion model at the same number of NFE (number of function evaluations) as AdjointFlow (namely, we fix 8 sampling steps). AdjointFlow clearly performs far better in this iso-compute comparison, demonstrating its superior efficiency.
Figure 6: Comparison of the efficiency–accuracy tradeoff for nonlinear and linear inverse scattering. All timings were measured on the same NVIDIA RTX A6000 GPU. (a) AdjointFlow achieves a strong balance between performance and efficiency, while RED-Diff and RED-Flow offer small accuracy improvements at substantially greater inference cost; the seconds-per-sample axis uses a logarithmic scale. (b) AdjointFlow achieves competitive accuracy at the lowest cost, though the efficiency gain is smaller than in the nonlinear problem.
Figure 7: Qualitative comparisons of reconstructions from AdjointFlow and diffusion baselines on linear inverse scattering with 60 receivers.
Number of receivers
360
180
60
Method
PSNR ↑
SSIM ↑
Meas err (%) ↓
PSNR ↑
SSIM ↑
Meas err (%) ↓
PSNR ↑
SSIM ↑
Meas err (%) ↓
Traditional
FISTA-TV
32.126 (2.139)
0.979 (0.009)
1.23 (0.25)
26.523 (2.678)
0.914 (0.040)
2.65 (0.30)
20.938 (2.513)
0.709 (0.103)
6.05 (0.65)
PnP Diffusion Prior
DDRM
32.598 (1.825)
0.929 (0.012)
1.04 (0.26)
28.080 (1.516)
0.890 (0.019)
1.57 (0.39)
20.436 (1.210)
0.545 (0.037)
3.04 (0.92)
DDNM
36.381 (1.098)
0.935 (0.017)
0.78 (0.22)
35.024 (0.993)
0.895 (0.027)
0.58 (0.16)
29.235 (3.376)
0.917 (0.022)
0.28 (0.07)
Table 3: We present the results for linear inverse scattering with three different levels of measurement sparsity and noise level σy=10−4 following [ 138 ] (bold result indicates best). “Meas err” quantifies relative agreement between predicted and observed measurements, following InverseBench’s normalization convention. Across all three levels of data sparsity, RED-Flow achieves the best reconstruction PSNR, while also having either the best or close to the best SSIM and competitive data consistency.
Figure 8: Comparisons of AdjointFlow against different diffusion-based methods on the seismic imaging task.
Method
RRMSE ↓
PSNR ↑
SSIM ↑
Data misfit ↓
Traditional
Adam
33.3 (8.6)
9.968 (2.083)
0.305 (0.120)
115.14 (52.10)
Adam †
8.9 (2.1)
21.273 (2.045)
0.679 (0.073)
15.89 (10.16)
L-BFGS †
7.0 (2.3)
23.398 (2.749)
0.704 (0.077)
9.18 (6.47)
PnP Diffusion Prior
DPS
25.0 (15.4)
14.111 (6.820)
0.491 (0.161)
155.08 (92.17)
Table 4: We present the results for seismic imaging from InverseBench [ 138 ] (bold result indicates best non-oracle method). “Data misfit” quantifies agreement between predicted and observed measurements, following InverseBench’s normalization convention. For AdjointFlow we report performance using 1 sampling step, as we found this to give the best performance based on validation data. † indicates “oracle models” initialized from ground-truth data blurred with a Gaussian filter ( σ=20 pixels), which requires access to the unknown ground truth and therefore serves only as an oracle baseline.
Figure 9: Comparisons of AdjointFlow against different diffusion-based methods for black hole imaging with 3% observation ratio. Even with the simple ring-shaped initialization, AdjointFlow properly matches the shape of both simpler ring-like black holes as well as those with a more exotic accretion disk.
Figure 10: We compare accuracy (RRMSE) and efficiency (number of sampling steps) of different stochastic interpolants on the four nonlinear inverse scattering datasets. The best performance on each dataset is labeled with a diamond. We do not use TP parameterizations for this ablation. For SDE sampling we choose ϵt=σ2γt/2 in ( 12 ) to cancel the explicit 1/γt factor in the score correction. For γt=1−t , the noise-free-velocity ODE uses a deterministic initialization, while the other samplers use a Gaussian initialization. We observe that sampling with x˙t=vt(xt) performs consistently well across datasets and only requires a small number of sampling steps for near-optimal performance.
RRMSE (%) ↓
Method
Shepp–Logan
3–5–10h Triangles
10h Overlapping Squares
Brain Scans
Uncoupled SI + bt
0.972 (0.395)
0.590 (1.599)
1.808 (2.403)
4.711 (1.250)
Uncoupled SI + bt + TP
1.097 (0.504)
1.356 (2.871)
2.649 (2.537)
4.887 (1.282)
Coupled SI + vt
1.031 (0.265)
1.228 (1.473)
2.345 (2.159)
4.233 (1.020)
Coupled SI + vt + DD-TP ( AdjointFlow )
0.379 (0.158)
1.108 (2.861)
1.143 (1.686)
3.421 (0.809)
Table 6: Comparison of reference distribution and model parameterization on the four nonlinear inverse scattering datasets, with mean and standard deviation of RRMSE across the test set shown. We fix the number of sampling steps at 8 for all models. Without TP, we find that a coupled interpolant does not meaningfully improve results on these datasets, demonstrating the practical necessity of the DD-TP parameterization. Bold indicates the best result on each dataset.
Initialization
W2↓
PSNR ↑
SSIM ↑
Meas err (%) ↓
60 receivers
Constant
127.0 (16.3)
13.007 (2.830)
0.436 (0.208)
38.05 (8.67)
Training mean
157.1 (20.9)
17.025 (2.675)
0.631 (0.156)
14.74 (2.31)
Adjoint
110.2 (18.4)
25.341 (6.155)
0.910 (0.066)
5.50 (1.85)
Pseudoinverse
53.5 (10.2)
25.460 (4.969)
0.920 (0.054)
4.70 (0.99)
180 receivers
Table 7: We present the results of AdjointFlow for linear inverse scattering with four choices of initialization a(y) , three different levels of measurement sparsity, and noise level σy=10−4 following [ 138 ] (bold result indicates best for each receiver count). All models share the architecture and training procedure of Section 4.2 and differ only in a(y) ; the pseudoinverse corresponds to AdjointFlow in Table 3 . Within each block, the first two initializations are data-independent and the last two are data-dependent. W2 is the empirical transport distance W2(δa(y),δm)=∥a(y)−m∥2 between singular endpoints ( E.3 ), averaged over the test set.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
RRMSE (%) ↓
Noise Level
Method
Shepp–Logan
3–5–10h Triangles
10h Overlapping Squares
Brain Scans
0% Noise
AdjointFlow
0.379
1.108
1.143
3.421
EquiNet CDM
1.323
1.590
1.744
5.661
10% Noise
AdjointFlow
2.844
1.141
2.015
21.598
EquiNet CDM
3.355
1.564
2.675
21.889
20% Noise
AdjointFlow
7.800
1.677
5.763
38.648
Appendix
Table 8: Model performance on the four nonlinear inverse scattering datasets with increasing levels of noise added to the data, from 10% to 40%. We fix the number of sampling steps for AdjointFlow at 8, and compare to the results of the EquiNet CDM from [ 132 ] with 256 sampling steps. EquiNet CDM results for brain scans are obtained from a model we trained using the pipeline from [ 132 ] .
Figure 11: Effect of measurement noise on nonlinear inverse scattering. The panels show 10h Overlapping Squares (top left), 3–5–10h Triangles (top right), Shepp–Logan (bottom left), and Brain Scans (bottom right). Within each panel, columns correspond to 0%, 10%, 20%, and 40% added noise, as defined in the text. Rows show the measurement data, adjoint-based initialization, AdjointFlow reconstruction, reconstruction within the boxed region of interest (ROIs), reconstruction error, and error within that region.
Figure 12: Effect of finite-difference stencil order on nonlinear inverse scattering for (a) 10h Overlapping Squares , (b) 3–5–10h Triangles , and (c) Shepp–Logan . Each panel shows the ground-truth perturbation at left. The remaining columns compare the 8th-order and 4th-order results and their signed difference: measurement data at the highest probing frequency in the top row, and the corresponding adjoint back-projections in the bottom row.
Shepp–Logan
3–5–10h Triangles
10h Overlapping Squares
Stencil Order
RRMSE ↓
MELR ↓
SD ↓
RRMSE ↓
MELR ↓
SD ↓
RRMSE ↓
MELR ↓
SD ↓
4th Order Stencil
0.726
0.788
0.624
1.112
0.617
0.258
1.242
1.506
0.525
6th Order Stencil
0.382
0.479
0.576
1.122
0.626
0.258
1.150
1.411
0.515
8th Order Stencil
0.379
0.474
0.576
1.108
0.607
0.258
1.143
1.401
0.514
Appendix
Table 9: AdjointFlow performance on the three synthetic inverse scattering datasets when the test data is generated using different finite-difference stencil orders (training data was generated with an 8th order stencil). We report RRMSE as a percentage and MELR ×10−2 , and fix the number of sampling steps for AdjointFlow at 8.
Stencil Order
RRMSE (%) ↓
PSNR ↑
SSIM ↑
Data misfit ↓
2nd order
9.7 (1.9)
20.515 (1.509)
0.676 (0.062)
112.39 (29.17)
4th order
9.7 (1.9)
20.530 (1.516)
0.676 (0.062)
115.63 (30.89)
Appendix
Table 10: AdjointFlow performance on seismic imaging using test data generated with 2nd and 4th order spatial finite-difference stencils (training data was generated with a 4th order stencil). We report performance using 1 sampling step, as we found this to give the best performance based on validation data.
Figure 13: Effect of spatial finite-difference stencil order on seismic imaging. The ground-truth wavespeed map is shown at left. The remaining columns compare the 4th-order and 2nd-order results and their signed difference. The top row shows single-shot receiver measurements, and the bottom row shows the corresponding adjoint-based initializations. Each color bar gives the scale of its associated image.
Training Samples
Method
Shepp–Logan
3–5–10h Triangles
10h Overlapping Squares
Brain Scans
RRMSE ↓
RRMSE ↓
RRMSE ↓
RRMSE ↓
1,000
AdjointFlow
1.912
4.356
7.933
5.476
EquiNet CDM
2.085
8.475
12.454
9.485
2,000
AdjointFlow
1.242
3.136
5.856
4.829
EquiNet CDM
1.644
5.454
7.893
7.995
4,000
AdjointFlow
0.853
2.582
4.014
4.331
Appendix
Table 11: Model performance on four inverse scattering datasets as the number of training samples increases from 1,000 to 21,000. We fix the number of sampling steps for AdjointFlow at 8, and compare to the results of the EquiNet CDM from [ 132 ] with a fixed budget of 256 sampling steps. EquiNet CDM results for brain scans are obtained from a model we trained using the pipeline from [ 132 ] .
Reference Distribution
Shepp–Logan
3-5-10h Tri
10h Squares
Brain Scans
Cells ( Nrec=60/180/360 )
Seismic Imaging
Black Holes
Gaussian
113.1 (3.0)
111.6 (18.7)
113.1 (3.8)
112.8 (8.6)
180.7 (11.4)
180.1 (18.3)
89.7 (11.9)
Physics-based
94.1 (16.0)
77.4 (22.9)
68.6 (7.2)
95.7 (24.3)
53.6 (10.1) / 24.0 (4.8) / 22.8 (4.7)
169.6 (29.0)
48.9 (11.3)
Appendix
Table 12: Comparison of Wasserstein-2 distances estimated numerically between the reference and target distributions when the base distribution is a standard Gaussian x0∼N(0,I) vs. when the base distribution ρ0=δa(y) is concentrated at the physics-based approximation. We report the averages across each training dataset with standard deviation shown in parentheses.
Model
Problem
∣ϕ∣
∣θ∣
EquiNet-CNN
Nonlinear inverse scattering
156k
375k
ADM (PnP)
Nonlinear inverse scattering
0
26.8M
EquiNet-CNN
Linear inverse scattering ( Nrec =60/180/360)
20k/145k/549k
2.73M
ADM (PnP)
Linear inverse scattering
0
26.8M
ADM (PnP)
Seismic Imaging
0
26.8M
ADM (conditional)
Seismic Imaging
550k
27.8M
Appendix
Table 13: Neural network parameter counts by problem. As in Algorithm 1 , θ denotes the main score/velocity/target-prediction network parameters and ϕ the conditional branch parameters. PnP models have no conditional branch ( ∣ϕ∣=0 ).
Problem
Ntrain
Ntest
Dimension of m
Dimension of y
Forward Eval. (sec)
Gradient Eval. (sec)
Nonlinear inverse scattering
21000
500
80×80
80×80×2×3
2.310
2.455
Linear inverse scattering
10000
100
128×128
20×Nrec×2
0.004 / 0.005 / 0.006
0.011 / 0.012 / 0.015
Seismic Imaging
50000
10
128×128
16×129×1001
4.004
6.261
Black hole imaging
50000
100
64×64
185/885/4007
0.001 / 0.002 / 0.008
0.005 / 0.019 / 0.111
Appendix
Table 14: Dataset parameters and forward/gradient evaluation timings (medians of 50 evaluations after warm-up). FWI measurements are Fourier-transformed and sub-sampled to 16×33×20 (Section 4.3 ). Linear inverse scattering uses Nrec∈{60,180,360} receivers; black hole observation ratios are 3% , 10% , and 100% . Inverse scattering uses one Quadro RTX 6000 GPU; seismic and black hole imaging use Intel Xeon Gold 6248R CPUs on the same node, with 16 single-threaded workers and one CPU thread, respectively, following [ 138 ] .
Figure 14: Further qualitative comparisons of AdjointFlow with different diffusion-based methods on nonlinear inverse scattering for the Brain Scans dataset.
Figure 15: Further qualitative comparisons of AdjointFlow against diffusion baselines for linear inverse scattering with 60 receivers.
Figure 16: Further qualitative comparisons of AdjointFlow against diffusion baselines for seismic imaging. Note that while other methods occasionally produce accurate reconstructions, only AdjointFlow consistently avoids cycle-skipping.
Figure 17: Further qualitative comparisons of AdjointFlow against diffusion baselines for black hole imaging with 3% observation ratio.