Generative models excel at synthesizing high-fidelity samples from complex data distributions, but they often violate hard constraints arising from physical laws or task specifications. A common remedy is to project intermediate samples onto the feasible set at each sampling step. However, repeated projection can disrupt the learned sampling dynamics because this feasible set is defined for clean samples rather than noisy intermediate states. Thus, recent studies first estimate clean samples and then project them to the feasible set, but this increases algorithmic complexity and accumulates errors across steps. To address these issues, this paper proposes Chance-constrained Flow Matching (CCFM), a training-free method that formulates constraint enforcement during sampling as stochastic optimization while retaining the hard-constraint feasibility guarantee of projection-based methods. By avoiding direct projection of noisy intermediate states onto the clean-sample feasible set, CCFM mitigates the distributional distortion. Unlike approaches that first estimate and project clean samples, CCFM avoids complex multi-stage procedures. Experiments show that CCFM outperforms current state-of-the-art constrained generative models in modeling complex physical systems governed by partial differential equations, molecular docking problems, and physics-informed motion generation, delivering superior feasibility and fidelity.
Figures & tables
Figure 1: Overview of chance-constrained flow matching (CCFM). Solid arrows follow the sampling path; dashed arrows indicate projection steps. In CCFM, early states remain unchanged ( xt1 ), intermediate states are minimally corrected ( xt2→xt2′ ), and the final state is guaranteed feasible ( x1′ ) in a high-probability region. Repeated projection distorts the sampling path. ECI sampling relies on the quality of one-step prediction and may project outside high-probability regions.
Method
W2 distance ( ↓ )
Branch-mass Error ( ↓ )
RP
1.026
0.231
ECI
2.695
0.152
Post-only projection
0.671
0.080
CCFM (ours)
0.412
0.0248
Figure 2: Two-dimensional example results over five runs.
#S
#Stp
Methods
L-RMSD (%)
A-RMSD (%)
PB Valid (%)
Running
<1A˚
<2A˚
<2A˚
All
+L <2 Å
+A <2 Å
Time (s)
1
2
CCFM
8.19
32.20
74.01
47.74
22.03
35.88
0.64
FlexDock
7.06
31.36
73.73
20.34
9.32
14.97
0.52
RP
6.78
31.07
73.73
19.49
9.32
14.69
0.73
10
2
CCFM
14.69
37.01
74.86
67.51
31.36
51.98
4.34
FlexDock
12.99
38.14
73.16
43.50
21.19
33.05
4.20
Table 1: PDBBind docking results. #S denotes the number of samples generated per complex, with the best pose selected for evaluation; #Stp is the number of inference steps. L-RMSD/A-RMSD report fractions of complexes below the given thresholds, PB Valid is the proportion passing PoseBusters checks, with +L and +A further requiring L-RMSD <2 Å or A-RMSD <2 Å, respectively.
Figure 3: Structures generated by CCFM (ours) and FlexDock. CCFM enforces the ligand–protein distance constraint, while FlexDock violates it.
Problem
Metric
CCFM
PCFM
DPDE
FFM
RP
RD
MMSE ×10−2 ( ↓ )
3.3
3.6
5.2
5.3
8.1
SMSE ×10−2 ( ↓ )
2.7
2.9
3.8
3.8
8.5
CV (IC) ×10−2 ( ↓ )
0
0
8.0
8.2
0
CV (CL) ×10−15 ( ↓ )
9.5
9.5
3.9 ×1010
3.9 ×1010
10.8
NS
MMSE ×10−1 ( ↓ )
0.8
1.7
1.4
1.5
1.0
SMSE ×10−1 ( ↓ )
0.8
1.6
8.8
9.0
0.9
Table 2: Comparison of model performance in solution generation tasks for 1D Reaction–Diffusion equation and 2D Navier–Stokes equation across metrics. A lower value indicates better performance. Bold indicates the best (smallest) value in each row.
Figure 4: MMSE (the darker the better) of Reaction–Diffusion solutions over physical time, with the horizontal axis representing time evolution and the vertical axis indicating errors at each point.
Figure 5: MMSE (the darker the better) of Navier–Stokes solutions over physical time, downsampled from 50 to 25 frames for clarity. Each frame corresponds to the error at a given physical time.
Unconstrained
RP
CCFM
Constraint Violation (MSE) ( ↓ )
278.46
0.8333
0.0100
FID ( ↓ )
16.985
27.184
26.767
Table 3: Comparison of model performance in the physics-informed motion task.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Table 9
Mean
Median
95th percentile
0.8622%
0.8544%
1.2987%
Appendix
Table 6: Normalized deviations from straight interpolation between the initial noise and clean sample.
Figure 6: Comparison of CCFM (solid lines) and FlexDock (dashed lines) under different sampling budgets and evaluation metrics.
Metric
CCFM(0.7)
CCFM(0.9)
CCFM(1.0)
CCFM(1.1)
CCFM(1.3)
PB Valid ( +L <2 Å )
29.85
30.16
30.79
30.48
30.48
PB Valid ( +A <2 Å )
58.19
57.85
58.19
58.19
58.88
Appendix
Table 7: Sensitivity analysis for probability of constraint satisfaction. The probability is determined by the function ϕn(t)=21tn , and we adjust n to observe the results. Results are reported using #S=1 and #Stp=10. PB Valid (+L <2 Å or +A <2 Å) denotes the percentage of physically valid samples that also satisfy the respective RMSD thresholds.
Methods
L-RMSD (%)
A-RMSD (%)
PB Valid (%)
<1A˚
<2A˚
<2A˚
All
+L <2 Å
+A <2 Å
CCFM
7.80±(0.31)
32.57±(0.57)
72.27±(0.69)
79.15±(0.41)
30.45±(0.68)
58.53±(0.49)
FlexDock
7.45±(0.43)
31.92±(0.81)
72.65±(0.81)
58.07±(1.24)
24.57±(1.13)
40.96±(1.38)
RP
7.80±(0.34)
32.19±(0.79)
72.26±(0.99)
56.48±(0.72)
27.68±(0.81)
42.93±(0.85)
Appendix
Table 8: PDBBind docking results with standard deviation (#S=1, #Stp=10).
Figure 7: Visualization for the generated complexes(6HHR, 6I65, 6I66, 6JSE, 6JSG, and 6OD6) from the PDBBind test dataset. Predicted protein structures are shown in pink, predicted ligands in teal, ground-truth proteins in dark blue, and ground-truth ligands in melon.
Problem
Metric
CCFM (ours)
PCFM
Reaction–Diffusion
Running Time (s) ( ↓ )
2.703
3.874
Navier–Stokes
Running Time (s) ( ↓ )
3.065
3.077
Appendix
Table 9: Comparison of running time between CCFM and PCFM on Reaction–Diffusion and Navier–Stokes equations.
Problem
Metric
CCFM (0.1)
CCFM (0.3)
CCFM (0.5)
CCFM (0.7)
CCFM (0.9)
Reaction–Diffusion
MMSE ×10−2 ( ↓ )
3.6
3.6
3.3
3.2
3.2
SMSE ×10−2 ( ↓ )
2.8
2.9
2.7
2.6
2.6
CV (IC) ×10−2 ( ↓ )
0
0
0
0
0
CV (CL) ×10−15 ( ↓ )
9.9
9.6
9.5
9.5
9.5
Appendix
Table 10: Sensitivity analysis for probability of constraint satisfaction. The probability is determined by the function ϕn(t)=21tn , and we adjust n to observe the results.
Figure 8: Comparison of generated solutions for the Reaction–Diffusion equation in terms of MMSE varying with physical time.
Figure 9: SMSE of Reaction–Diffusion solutions over physical time, with the horizontal axis representing time evolution and the vertical axis indicating the state at each time step.
Figure 10: Example solution fields of the Reaction–Diffusion equation generated using the models compared with the ground truth.
Figure 11: Comparison of generated solutions for the Navier–Stokes equation in terms of MMSE varying with physical time.
Figure 12: MMSE (the darker the better) of Navier–Stokes solutions over physical time with all 50 frames. Each frame corresponds to the system state at a given physical time.
Figure 13: SMSE (the darker the better) of Navier–Stokes solutions over physical time with all 50 frames. Each frame corresponds to the system state at a given physical time.
Figure 14: Point-wise averaged residuals (the lighter the better) of Navier–Stokes solutions over physical time with all 50 frames. Each frame corresponds to the system state at a given physical time.
Figure 15: Example solution fields of the Navier–Stokes equation generated using the models compared with the ground truth.
Problem
Metric
CCFM (ours)
PCFM
DiffusionPDE
FFM
RP
Navier–Stokes
MMSE ×10−1 ( ↓ )
0.79±(0.03)
1.67±(0.09)
1.44±(0.10)
1.56±(0.16)
1.08±(0.08)
SMSE ×10−1 ( ↓ )
0.81±(0.11)
1.53±(0.18)
9.21±(0.39)
9.57±(0.53)
0.97±(0.25)
CV (IC) ( ↓ )
0
0
0.12±(0.03)
0.13±(0.08)
0
CV (CL) ×10−15 ( ↓ )
3.9±(0.01)
3.4±(0.02)
1.2±(0.13)×1010
1.2±(0.12)×109
4.1±(0.04)
Appendix
Table 11: Results with standard deviations for solution generation on the 2D Navier–Stokes equation. A lower value indicates better performance. Bold indicates the best value in each row.
Figure 16: Generated trajectories instances for the physics-informed motion generation. In each frame, the black solid objects are generated using three constraint-handling approaches. The red circles represent the correct locations of objects under Moon gravity.
Deep generative models provide state-of-the-art performance across a wide array of applications, with recent studies showing increasing applicability for science and engineering. Despite a growing corpus of literature focused on the integration of physics-based constraints into the generation process, existing approaches fail to enforce strict constraint satisfaction while maintaining sample quality. In particular, training-free constrained sampling methods, while providing per-sample feasibility guarantees, introduce a fundamental mismatch between the training objective and the constrained sampling procedure, often leading to performance degradation. Identifying this training-sampling misalignment as a central limitation of current constrained generative modeling approaches, this paper proposes Constraint-Aware Flow Matching, a novel end-to-end framework that explicitly incorporates constraint projections into the training objective. By aligning the model's learned dynamics with the constrained sampling process, the proposed method mitigates distributional shift induced by projection-based corrections, enabling high-quality constrained generation. The proposed approach is evaluated on three challenging real-world benchmarks, illustrating the generality and efficacy of the method.
Jacob K. Christopher, James E. Warner, Ferdinando Fioretto
University of Virginia · NASA Langley Research Center
Flow matching models excel at generative modeling, and many downstream applications require their samples to satisfy prescribed constraints, such as observed measurements and physical laws. However, existing constrained samplers often face a trade-off: enforcing constraints can substantially displace samples from the pretrained data distribution. To address this trade-off, we introduce MintFlow, a training-free constrained sampling framework that formulates constraint enforcement as a minimal intervention on the pretrained flow trajectory. MintFlow seeks the minimal perturbation of an intermediate flow state such that its subsequent evolution under the pretrained flow field satisfies the target constraint. By minimally perturbing the flow state while keeping the pretrained flow field unchanged, MintFlow enforces the constraint while minimizing unnecessary deviation from the pretrained distribution. An adjoint formulation yields a closed-form expression for this perturbation, eliminating expensive iterative optimization. Furthermore, MintFlow adaptively selects the intervention time to balance the required perturbation magnitude with its amplification by the remaining flow. Across a range of tasks in generative vision and physical system modeling, MintFlow achieves competitive constraint satisfaction while preserving the pretrained generative distribution substantially better than state-of-the-art constrained methods.
Yesom Park, Kelvin Kan, Qifan Chen +3
Department of Mathematics, University of California, Los Angeles · Brookhaven National Laboratory
Generating scientific data under previously unseen equality constraints requires satisfying the prescribed constraints without destroying the physical structure learned from data. Existing inference-time approaches often enforce physical constraints by repeatedly correcting predicted endpoints, but such discrete corrections can distort learned generative structure and introduce unstable feedback. In this work, we propose Manifold Attracting Flow Matching (MAFM), which continuously evolves a persistent anchor through an ODE whose dynamics drive it analytically toward the constraint manifold. Specifically, its geometric correction preserves model-guided tangent motion while prescribing the normal motion to contract constraint violations toward the manifold. This construction separates model tracking from constraint attraction and provides theoretical guarantees for residual contraction and local feedback stabilization. Experiments across geometric and partial differential equation benchmarks demonstrate improved generation quality over variant constraint-guided baselines while maintaining low constraint residuals. MAFM provides a flexible framework for coordinating learned physical structure and explicit constraints through controllable sampling dynamics, without retraining the underlying generative model.
Guanyu Chen, Jinchen Wang, Shengze Xu +4
Zhejiang University · University of California, Los Angeles · The Chinese University of Hong Kong +1