acceptodds
Withdrawn from ICLR 2027

DRiFT: Differentiable Grid-Based Rigid-Fluid Coupling for Learning and Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Intelligent agents, interacting with physical environments, require an accurate understanding of the consequences of their actions for efficient learning. Such agents are often trained inside simulated environments to alleviate over dependence on data, and gradients from such a simulation can help to train the agent. For this purpose, we present an end-to-end differentiable grid-based fluid simulation with strong two-way coupling with rigid bodies. In the forward pass, the solid-fluid boundary conditions are solved monolithically in a pressure linear system for stability. For backpropagation, we introduce a novel method of calculating and propagating gradients of combined fluid-solid state using adjoint method, which runs as fast as the forward solve. Our method, customized for coupling rigid bodies with inviscid fluids, is more suitable over general purpose methods like automatic differentiation, for use cases where performance is key for analyzing overall flow patterns and learning fluid properties. We demonstrate how our simulator can help to find an optimal force sequence to guide a pill through a complex fluid-filled blood vessel network, and how it can further generalize to arbitrary target states by helping to train a neural network controller. Additionally, we show the effectiveness of our differentiable simulator in isolation, in optimizing initial states which naturally evolve into desired target states. Finally, we analyze the accuracy, robustness and most importantly, the efficiency of our gradient computation method.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Differentiable Physics, Solid-Fluid Coupling

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_9QoiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces DRiFT, a differentiable Eulerian simulator for strongly coupled rigid-body and incompressible fluid dynamics. The framework employs a monolithic pressure projection to enforce fluid incompressibility and rigid-fluid coupling constraints, together with a manually derived discrete adjoint for efficient gradient computation. The authors demonstrate its applications to rigid-body control, initial-state optimization, and neural controller training. Experiments further investigate gradient accuracy and computational efficiency against finite differences and automatic differentiation frameworks.

Critical strengths
  1. Relevant research problem. Differentiable rigid-fluid coupling is relevant to physics-based learning and control. The paper considers a setting involving interactions between fluid pressure and rigid-body motion, where efficient gradient computation can be useful.

  2. Implementation of coupled simulation and adjoint differentiation. The paper implements a discrete adjoint approach for a monolithically coupled Eulerian fluid-rigid solver, supporting gradient computation in both 2D and 3D scenarios. While the underlying techniques are largely established, their integration provides a potentially useful framework for gradient-based optimization.

Critical weaknesses
  1. Insufficient evidence of the benefits of strong coupling.

The main contribution is the differentiable treatment of strongly coupled rigid-fluid dynamics. However, the experiments do not directly compare the proposed approach with weakly coupled differentiable solvers. Consequently, it remains unclear whether the monolithic coupling formulation provides substantial improvements in gradient accuracy, numerical stability, or control performance. Such comparisons are essential to justify the added complexity of the proposed method.

  1. Limited experimental complexity and evaluation.

Most experiments involve relatively simple rigid-body trajectory or initial-state optimization tasks. While these examples demonstrate the feasibility of the differentiable simulator, they provide limited evidence of its advantages in more challenging fluid-structure interaction problems. More systematic evaluations across different coupling strengths, rigid-body mass ratios, and longer optimization horizons would considerably strengthen the paper.

  1. Limited methodological novelty and baseline comparisons.

The framework builds upon established variational rigid-fluid coupling and discrete adjoint techniques. Although their integration is technically meaningful, the paper does not sufficiently distinguish its methodological advances from existing differentiable simulators. Comparisons primarily focus on computational performance against automatic differentiation frameworks rather than the benefits of the proposed coupling and differentiation strategy.

  1. Manuscript presentation and completeness.

The manuscript contains several unresolved cross-references, including references to "Tables 2 and ??", "Appendix ??", and "Figure ??". These issues suggest insufficient final proofreading and make some experimental descriptions difficult to follow. The presentation should be carefully revised to ensure clarity and reproducibility.

Questions for authors

Could the authors clarify how the geometry-dependent terms in the pressure adjoint are derived, particularly the derivatives of the pressure-system matrix and right-hand side with respect to the fluid level set?

How accurate are the computed gradients with respect to rigid-body positions, orientations, and fluid geometry? Can the authors provide independent finite-difference or Taylor-test validation?

How are Reynolds-number-dependent flow benchmarks interpreted within the inviscid fluid formulation, and can quantitative physical validation be provided?

Can the authors provide more systematic comparisons demonstrating the advantages of DRiFT for learning and control, beyond individual optimization examples?

Could the authors provide a direct comparison with weakly coupled differentiable rigid-fluid solvers under comparable settings? In particular, how does the proposed strong coupling formulation improve gradient accuracy, numerical stability, computational efficiency, and optimization performance? Such an evaluation would help clarify the practical benefits of strong coupling over simpler alternatives.

Reviewer_DRSpOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents DRiFT, a differentiable simulator for strongly two-way coupled fluid–rigid-body dynamics. The method builds upon an existing grid-based forward solver and derives its corresponding discrete adjoint to efficiently compute gradients through fluid–solid interactions. These gradients are used for optimization and control tasks, including guiding a capsule through a blood vessel network and optimizing initial states to reach desired targets.

Critical strengths

The paper develops an efficient discrete adjoint method for computing gradients in strongly coupled fluid–rigid-body systems. By deriving the backward computation for the coupled numerical solver, the proposed approach enables gradient-based optimization and control while reducing the computational overhead associated with general-purpose automatic differentiation. The authors conduct several experiments to validate the accuracy and computational efficiency of the proposed method.

Critical weaknesses
  1. Limited methodological novelty. The methodological novelty of this paper appears rather limited. The proposed approach primarily derives and implements the discrete adjoint of an existing fluid–rigid-body simulation solver, relying on well-established techniques such as semi-Lagrangian advection, variational pressure projection, and implicit differentiation. While the implementation may offer practical benefits, the underlying methodology is largely based on existing numerical techniques. Furthermore, the paper does not sufficiently distinguish its contributions from closely related differentiable fluid–structure interaction methods. The authors should more clearly explain which aspects of the proposed approach constitute genuinely new methodological contributions beyond adapting established adjoint techniques to a specific forward solver.

  2. Insufficient experimental evaluation. The experimental validation is not sufficiently comprehensive to support the paper's main claims. Although the authors compare gradient accuracy and computational efficiency against finite differences and general-purpose automatic differentiation frameworks, comparisons with closely related differentiable fluid–rigid-body simulators remain limited. Moreover, several downstream control experiments rely heavily on qualitative visualizations and loss curves, without systematic quantitative comparisons against alternative methods. More comprehensive evaluations of forward simulation accuracy, gradient accuracy, computational efficiency, memory consumption, and control performance would be necessary to convincingly establish the advantages of the proposed method.

  3. Presentation issues and typographical errors. The paper contains several presentation issues and typographical errors that should be corrected. For example, the caption of Table 3 includes a reference to a nonexistent figure.

Questions for authors

See above.

Reviewer_6LAgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This submission describes DRiFT, an Eulerian simulator with differentiable two-way coupling between fluids and rigid bodies. The claimed contribution is a manually derived discrete adjoint of a monolithic pressure projection, meant to deliver accurate gradients at far lower computational cost than automatic differentiation. Several optimization tasks are presented, and gradient accuracy as well as runtime are compared against existing approaches.

Critical strengths
  1. The problem is relevant, and a fast differentiable implementation of strongly coupled fluid-rigid-body dynamics would be genuinely useful. Covering both 2D and 3D examples adds value.

  2. Rather than a tape, the adjoint stores only states, which makes it usable at resolutions where the authors' AD setups exhaust their memory. In broad outline, the pressure adjoint matches what implicit differentiation of Eq. (5) would give, and source code is supplied.

Critical weaknesses

I have major reservations about the adjoint's mathematical correctness, the physical validity of the simulations, the experimental methodology, and the actual novelty of the contribution. Some of the paper's claims are contradicted outright by its own equations or numerical results. Weightiest of all, the experimental evaluation never establishes that the central method is correct.

  1. Working through the derivation, I was unable to reproduce Eq. (7) from the paper's own equations, and the problem seems to me more than notational. Appendix B differentiates the solution of the linear system in Eq. (5) with respect to the level-set geometry ϕ\phi. What that derivative should be depends on which operators in Eq. (5) depend on ϕ\phi, which is never stated in the paper. The ghost-fluid treatment at the free surface implies that GG does. Read as printed, G⊤G^\top then inherits that dependence, and since G⊤G^\top appears in both the system matrix AA and the right-hand side bb, implicit differentiation of Eq. (5) yields $A\frac{\partial p}{\partial\phi} =\frac{1}{\rho}\left(\frac{\partial G}{\partial\phi}\right)^{\top} M_F\ u_{t+1} -\frac{\Delta t}{\rho^{2}}G^{\top} M_F \frac{\partial G}{\partial\phi}p$. Appendix B retains just the second of these terms. The error is easy to exhibit concretely. Set G=ϕG=\phi and MF=J=MS=ρ=Δt=ut+1=1M_F = J = M_S = \rho = \Delta t = u_{t+1} = 1, v=0v=0 in Eq. (5). The exact solution is p=ϕ/(1+ϕ2)p = \phi/(1+\phi^2), so the exact ∂p/∂ϕ\partial p/\partial\phi at ϕ=1\phi=1 is zero but the formula printed in Appendix B evaluates to −1/4-1/4 there.

  2. The second issue is that in the forward step, advection reads st+1=W(ut)sts_{t+1}=W(u_t)s_t, where the interpolation weights composing WW depend on the velocity field. Differentiating a downstream objective through this update with respect to a velocity component ut,ku_{t,k} necessarily involves the quantity being transported: ∂L∂ut,k=(∂L∂st+1)⊤∂W∂ut,kst\frac{\partial \mathcal L}{\partial u_{t,k}}=\left(\frac{\partial\mathcal L}{\partial s_{t+1}}\right)^\top\frac{\partial W}{\partial u_{t,k}}s_t. Possibly this is fine for self-advection under some particular convention but the authors state explicitly that the same equations cover both velocity and level-set advection, and for the latter st=ϕts_t=\phi_t. As a general advection adjoint, the derivation is therefore wrong.

  3. When the collision response of Eq. (10) is differentiated with respect to position, the derivative through the contact normal picks up a negative prefactor. Equation (11) shows a positive one.

  4. The quaternion expansion contains a positive scalar dot product, which contradicts the multiplication rule stated directly above. The two Jacobian matrices given in Appendix D are themselves correct and only the expanded product carries the sign error. In Eq. (9), the reverse-mode formulas fail to transpose the forward Jacobians.

  5. The gradient checks in Section 6.3 rely mostly on perturbing the initial rigid-body velocity. However, derivatives with respect to rigid-body pose, fluid geometry, collision state, and control forces are left untouched. No result in the paper differentiates through the solver with respect to these quantities and checks the sensitivities against a reference but a check would be cheap by comparing each derivative against central finite differences of the forward simulation at several perturbation magnitudes. Given the inconsistencies above. Ideally one test would also exercise the coupled case, e.g. a floating body whose position perturbs the free surface, so that the ϕ\phi-derivative, the collision state, and the pressure solve are exercised together.

  6. Section 3.1 formulates inviscid, incompressible flow. Section 6.3 then validates the simulator on lid-driven cavity flow and cylinder wakes at Re=40Re = 40 and Re=100Re = 100 whose defining physics is viscosity. With no viscous term in the equations, a stated Reynolds number has no consistent meaning here. The benchmarks are being solved as a different PDE than the ones they were designed to test, and whatever damping plays the role of viscosity is numerical diffusion, which is not a physical model and is not controlled. Additionally, the cylinder case shows under-resolution. The radius is 0.1 m against a grid spacing of 0.04 m, so the obstacle spans five cells across its diameter. Vortex shedding, with its separation points pinned exactly at the cylinder surface, cannot be assessed meaningfully at that resolution. The text moreover calls these experiments steady-state validation while the Re=100Re = 100 case is precisely the unsteady shedding regime. Another gap is that both benchmarks hold the obstacle fixed. The two-way fluid–body coupling is the paper's contribution but no validation shows this.

  7. According to the abstract, the backward pass runs as fast as the forward pass. For the 3D dam-break experiment, Table 2 reports 3423.5 ms for the forward pass against 8597.8 ms for the backward pass, which is a factor of approximately 2.512.51. I also fail to see that DRiFT consistently beats automatic differentiation. In the 3D rigid-drop experiment, PyTorch's backward runtime is lower than DRiFT's. With the forward pass included, the PyTorch implementation is faster than DRiFT in two of the four settings where both ran (973 vs. 1370 ms for Rigid Drop 3D; 66 vs. 82 ms for Dam Break 140×80).

  8. The comparison with DiffFR involves two simulators with different discretizations, and equivalent physical accuracy is never established. From a fixed number of optimization epochs, the Appendix infers a runtime difference of roughly a factor of eight. Yet Figure 4c shows different loss trajectories, and the comparison never reports the time needed to reach a common objective threshold. With physical fidelity, accuracy, and optimization success unmatched, I have a hard time attributing the observed runtime difference to the proposed differentiation method.

  9. All experiments establish that gradient-based optimization brings down selected losses. They don't show that the proposed approach is more sample-efficient, or reaches better solutions than the alternatives as claimed. The abstract calls the method suited for "learning fluid properties" but no experiment identifies or learns a fluid property.

Questions for authors
  1. Which of the operators in Eq. (5) depend on ϕ\phi, and how must Eqs. (7) and (8) be changed accordingly? Which of the state-dependent operators, JJ, MFM_F, GG and CC included, are actually differentiated in the implementation, and how are their derivatives computed?
  2. Beyond the initial-velocity checks, has the complete adjoint been validated independently for position, orientation, geometry, collisions, and control forces? How do the authors reconcile the Dam Break 3D entry of Table 1 with the perturbation data in Table 3?
  3. Does the implemented solver contain viscosity? If not, what justifies the finite-Reynolds-number benchmarks?
  4. Relative to Batty et al. (2007), Takahashi et al. (2021), and existing implicit differentiation methods, what concretely is the methodological advance?
  5. Which results demonstrate advantages over properly controlled differentiation baselines and black-box optimization, including the setup of Ramos et al. (2022)?

Batty et al. “A Fast Variational Framework for Accurate Solid-Fluid Coupling”, ACM TOG 26(3) Takahashi et al. “Differentiable Fluids with Solid Coupling for Learning and Control”, AAAI 2021 Ramos et al. “Control of Two-way Coupled Fluid Systems with Differentiable Solvers”

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.