acceptodds
Under review as a conference paper at ICLR 2027

DiffStent: Surrogate Adjoint Differentiation for Stent Deployment in Anisotropic Arteries

Abstract

We present DiffStent, a differentiable finite element framework for simulating arterial deformation during in silico stent deployment, where nearly incompressible, anisotropic vessel walls undergo large deformation and frictional contact with the stent. DiffStent combines the fiber-reinforced Holzapfel–Gasser–Ogden constitutive model with an implicit discrete adjoint that differentiates the converged mechanical equilibrium without unrolling the nonlinear solver, thus reducing the memory cost by three orders of magnitude. We identify three mechanisms that can lead to inaccurate, unstable, or uninformative gradients: nonsmooth switching, state mismatch, and gradient cancellation. To address these, we apply a contact surrogate in the backward pass while retaining the original physical model in the forward pass. We show superior gradient accuracy of our method over AutoDiff and finite differencing across constitutive benchmarks, idealized stented vessels, and patient-specific calcified coronary anatomies, and use the resulting sensitivities to quantify how tissue properties and calcification patterns govern lumen expansion and arterial wall stress. Finally, we demonstrate that DiffStent enables gradient-based system identification from retrospective clinical images of atherosclerotic coronary anatomy. Together, these results establish a practical approach for differentiating through nonlinear biomechanical simulations with stiff constitutive laws and nonsmooth contact, enabling patient-specific parameter inference and optimization in computational cardiovascular mechanics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Differentiable Physics, Optimization, Finite Element Method, Biomechanical Simulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over one review, above 85% of the others.
Reviewer_bopKOpenReview ↗
Rating
3
Confidence
1
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces DiffStent, a GPU-native differentiable finite element framework designed to simulate and differentiate arterial deformation during in silico stent deployment. The framework couples multibody frictional contact (balloon, stent, and multilayer arterial wall) with the nearly incompressible, anisotropic Holzapfel-Gasser-Ogden (HGO) constitutive model with fiber dispersion.

Critical strengths
  1. Simulating and optimizing endovascular interventions like stent deployment using gradient-based methods is a vital frontier in computational cardiovascular mechanics. DiffStent bridges a major gap by handling anisotropic, fiber-reinforced, nearly incompressible soft tissues under complex frictional contact.
  2. The implicit discrete adjoint approach drastically reduces memory complexity, making gradient computation independent of the number of forward simulation steps. The comparison against unrolled autodiff (Figure 5) clearly highlights the immense memory savings (avoiding multi-terabyte footprints).
Critical weaknesses
  1. Persistence of Gradient Discrepancies & Hyperparameter Sensitivity: As shown in Section 4.4 and the appendix, the proposed contact activation correction and chained plastic-history adjoint do not fully eliminate all gradient errors (e.g., intima stress sensitivities and media k1k_{1} on T3). Moreover, the optimal smoothing width ww is sensitive to operating conditions and does not universally transfer across testbeds. A deeper discussion on why these residual errors persist—and whether modern barrier-based methods (such as Incremental Potential Contact) could bypass them entirely—would strengthen the methodological narrative.
  2. Computational Cost Transparency. While the adjoint avoids unrolled memory scaling, the forward simulation itself is computationally demanding (taking anywhere from 12 minutes to 1 hour depending on the testbed), and graph-colored sparse probing/factorization adds non-trivial overhead. Providing a clearer end-to-end wall-clock time breakdown for gradient-based optimization loops (e.g., total time for the Adam iterations in Figure 9) would help practitioners assess practical feasibility.
Questions for authors

Please refer to weakness

1 comment

We thank the reviewer for their feedback and encouraging assessment. Responding to these comments made the manuscript clearer and more useful to practitioners. We have uploaded the revised manuscript and appendix above.

1. Persistence of Gradient Discrepancies & Hyperparameter Sensitivity: As shown in Section 4.4 and the appendix, the proposed contact activation correction and chained plastic-history adjoint do not fully eliminate all gradient errors (e.g., intima stress sensitivities and media on T3). Moreover, the optimal smoothing width is sensitive to operating conditions and does not universally transfer across testbeds. A deeper discussion on why these residual errors persist—and whether modern barrier-based methods (such as Incremental Potential Contact) could bypass them entirely—would strengthen the methodological narrative.

Answer: We thank the reviewer for this comment. The residual errors come from three sources, which we will discuss more explicitly in the revision. First, the forward solver uses explicit dynamic relaxation, so the recorded states are not in exact static equilibrium (residual forces of 5–9% in the wall and 20–40% in the stent at peak load), while the adjoint assumes they are in perfect equilibrium. Second, the adjoint holds the contact set and the set of yielded elements fixed. The activation correction smooths the contact stiffness but does not differentiate the contact-pair search, so it is a surrogate, and its best width depends on the local gap distribution, which changes with friction, penalty stiffness, and geometry. This is why one width does not transfer across testbeds. Third, part of the discrepancy comes from discretization and from the finite-difference reference. Refining the mesh moves the intima κ ratio from 0.76 to 0.95, and the T3 media and adventitia k1 misses are at low signal-to-noise (3–5).

We chose the explicit solver rather than an implicit solver because it is matrix-free, GPU-friendly, and robust under severe contact. We agree that a barrier method such as IPC is a promising remedy as its contact force is smooth across activation, and it is formulated with an implicit solver that converges to equilibrium, which would address the first two sources. It would not remove errors from gradient cancellation or from plastic yielding though. We are working on an implicit IPC-based version as future work and have added this discussion in Appendix M and N.

2. Computational Cost Transparency. While the adjoint avoids unrolled memory scaling, the forward simulation itself is computationally demanding (taking anywhere from 12 minutes to 1 hour depending on the testbed), and graph-colored sparse probing/factorization adds non-trivial overhead. Providing a clearer end-to-end wall-clock time breakdown for gradient-based optimization loops (e.g., total time for the Adam iterations in Figure 9) would help practitioners assess practical feasibility.

Answer: Below is the end-to-end wall-clock time breakdown for the stress-constrained pressure optimization on T3, measured on one RTX A6000. Each iteration consists of one forward deployment and one adjoint solve.

Table: Wall-clock time breakdown for a gradient-based optimization loop. Each iteration is one forward deployment and one adjoint solve.

ComponentTime
Forward deployment42.2 min
– Force evaluation (materials, pressure, springs)30.4 min
– Contact search9.0 min
– Time update1.5 min
– Contact forces0.7 min
– Overhead0.7 min
Adjoint (probing, factorization, solve)2.1 min
Iterations15
Total11.1 h

Each iteration is one forward deployment and one adjoint solve. The forward simulation dominates as the adjoint, including sparse probing and factorization, takes 2.1 min, about 5% of an iteration. Within the forward pass, force evaluation accounts for 72% of the time and contact search for 21%. For comparison, central finite differences on the four identified parameters would need eight extra deployments per iteration, about 5.6 h against 2.1 min for the adjoint. The practical bottleneck is therefore the explicit forward solver, not the gradient computation. We state this in the limitations, and reducing it is part of our ongoing work on an implicit solver.

Comments on the paper

6 comments
AuthorsOpenReview ↗

Summary of responses and modifications

We thank the reviewers for their help in significantly improving the quality of our manuscript. We are delighted that the reviewers recognized the importance and difficulty of the problem we address (bopK, 1UdY) and the memory advantage of our implicit adjoint over unrolled autodiff (bopK, 1UdY). We also appreciate that DiffStent was recognized for bridging a major gap in differentiable simulation of anisotropic, nearly incompressible tissue under frictional contact (bopK), and that our gradient checks, mesh comparisons, and reporting of failure cases were found detailed and useful (1UdY).

Both reviewers found the presentation of our initial manuscript difficult to follow. The paper draws on both computational mechanics and interventional cardiology, so we have reorganized and rewritten the introduction to define the technical and clinical concepts before the method uses them.

Alongside our individual responses, we summarize below the major changes to the manuscript. Each change is meant to address a specific reviewer concern, and together we believe they improve the informativeness and technical soundness of the paper. We highlight the following modifications, experiments, and analyses results from reviewer suggestions:

Clarified which results use the contact surrogate (1UdY). Section 4.4 now states that the surrogate is evaluated only in the ablation. Pressure optimization uses the endpoint adjoint and OCT fitting uses the chained adjoint, and neither uses the surrogate. The abstract now describes it as reducing contact-related gradient errors in some settings.

Characterized surrogate reliability and smoothing-width selection (1UdY, bopK). Sections 3.4 and 4.4 present the surrogate as an approximation with a single hyperparameter, w. Figure 6 and Appendix J show that its benefit depends on width, friction, contact stiffness, and deployment stage, and that a width tuned on T2 does not transfer to T3.

Qualified the gradient-accuracy comparison against AutoDiff (1UdY). The abstract and conclusion now claim an advantage over AutoDiff in memory and wall time, not in accuracy. Table 4 shows equal median error against the closed-form reference, and we removed the former Table 12, whose reference was not independent.

Replaced the synthetic identification with a fit to clinical OCT (1UdY). Section 4.6, Figure 10, and the new Appendix I fit three wall-stiffness constants to the lumen measured in a post-stent OCT pullback. The objective falls from 0.0326 to 0.0125. We present the fitted constants as effective parameters, not as identified tissue properties.

Reframed the pressure optimization (1UdY). Section 4.6 now states that this is a scalar, monotone problem that a bracketing search could also solve, and that it does not establish an advantage over scalar search. However, we have added a clinical case of material identification where gradient optimization will outperform scalar search.

Reported end-to-end wall-clock cost (bopK). The new Appendix L gives a time breakdown of an optimization loop. The adjoint takes about 5% of an iteration, so the bottleneck is the explicit forward solver.

Expanded the discussion of residual gradient errors (bopK). Appendix K separates contact switching from state mismatch and gradient cancellation. The new limitations and future-work sections discuss the explicit solver and an implicit solver with barrier contact.

All other points are addressed in the individual responses below. We greatly appreciate the feedback and welcome further suggestions to improve the manuscript.

We appreciate the reviewer’s careful reading of our work. Addressing these comments led us to precisely state and defend our claims with strengthened supporting evidence, and the manuscript is clearer as a result. The revised manuscript and appendix are available for download above.

1. Section 4.4 states that the subsequent experiments use the endpoint adjoint without either proposed treatment. Which reported results therefore demonstrate the benefit of the contact surrogate beyond the unmodified adjoint? Please reconcile this with the claims in the abstract and conclusion.

Answer: The benefit of the contact surrogate is shown in the ablation of Section 4.4, not in the later experiments. On stenosed T2 (Figure 6), it moves the FD/adjoint ratio for the sensitivity of mean media stress to intima κ toward 1 across different friction and penalty settings. It also improves the stiff-plaque case (1.91 → 1.19 at modulus ratio 152) and the T3 lumen-area sensitivities to intima κ and k1 (2.6 → 0.77 and 4.0 → 1.26, Appendix I).

However, the best width depends on the contact setting, and a width tuned on T2 does not transfer to T3. We therefore did not use it in Sections 4.5 and 4.6, so that those results do not depend on a tuned hyperparameter. On T3, seven of nine resolved ratios lie within 0.87–1.13 without the correction (Section 4.6).

We agree that the abstract and conclusion overstate the role of the surrogate. We have revised them to present it as a partial correction evaluated in an ablation, and the three pathologies as an analysis of where the adjoint is and is not reliable.

2. How was the contact-smoothing width selected without access to reference gradients? Given its condition-dependent effects and the sign reversal reported on T3, what existing evidence supports using the correction reliably on a new deployment case?

Answer: The width was not selected without reference gradients. We swept w from 0.1 to 5 µm and compared each value against finite differences (Figure 6). The 1 µm value used elsewhere was chosen from this sweep on T2. We have stated this clearly in the revision.

We agree that the current evidence does not support using the correction reliably on a new deployment case. A width chosen by matching finite differences may also compensate for errors unrelated to contact, such as state mismatch or discretization. Consistent with this, mesh refinement alone moves the same intima κ ratio from 0.76 to 0.95, and the T2 width produces a sign reversal on T3. For this reason, all results in Sections 4.5 and 4.6 do not use the contact surrogate. We therefore present the correction as a diagnostic, not as a general remedy. It shows that contact activation contributes to the gradient error, since changing only the adjoint’s contact stiffness changes the agreement with finite differences. A smooth barrier contact model with an implicit solver that removes the need for a generalized contact width is left for future work. As noted in our response to the previous comment, we have revised the abstract and conclusion so that the contact surrogate is no longer presented as a central contribution. The main contributions are the differentiable simulator and the implicit adjoint, which matches autodiff in accuracy at a fraction of the memory and time, together with the analysis of where its gradients are and are not reliable.

3. Please confirm whether material identification uses simulated full-node displacements rather than deformation measured from clinical images. If so, how should the abstract’s claim of “system identification from retrospective clinical images” be interpreted?

Answer: We thank the reviewer. We confirm that the identification experiment in the original submission uses simulated displacements at every tissue node, not deformation measured from clinical images. It is a synthetic inverse problem on patient-derived anatomy and does not support the abstract’s claim of identification from clinical images.

To close this gap, we have added an identification from clinical data. We reconstructed pre- and post-procedural coronary meshes from one patient’s OCT data and fitted wall stiffness parameters by minimizing the squared difference between simulated and OCT-derived radial displacements, sampled at 64 fixed angles on several cross-sections. The identification target is therefore image-derived deformation. A new figure shows the convergence of the lumen radius, the parameter trajectories, and the wall-clock time.

Regarding the remaining errors in the synthetic study: both fits converge to the repeat baseline of the loss, so the errors reflect the limited information in a displacement field at one load state more than gradient error. The same applies to the clinical fit, where the two intimal constants trade off. We therefore present the clinical estimates as effective parameters that reproduce the observed lumen, not as validated tissue properties.

We have revised the abstract and Section 4.6 to distinguish the synthetic validation from the OCT-based identification, which is on a single retrospective case.

4. What independent evidence establishes greater gradient accuracy than AutoDiff? Table 4 reports identical median errors against the closed-form reference, while Table 12 uses a relaxed adjoint reference. Please distinguish accuracy for the equilibrium problem from accuracy for the actual forward simulation trajectory.

Answer: We thank the reviewer for pointing this out. This distinction is important, and we have made it explicit in the revision. Table 4 is our only comparison against an independent ground truth: the closed-form solution on the well-relaxed T1 cube (200,000 steps). There the adjoint and autodiff have the same median error (2.5 × 10⁻⁵), and the adjoint’s worst-case error is about half that of autodiff (9.9 × 10⁻⁴ vs. 1.9 × 10⁻³). This shows that the adjoint matches autodiff. It does not establish greater accuracy. We agree with the reviewer that in former Table 12 the reference (the adjoint at a relaxed state) is not independent and favors the adjoint by construction. We have removed this table and the accompanying former Appendix K.

Autodiff differentiates the simulated trajectory and is exact for it by construction. The adjoint differentiates the static equilibrium that the trajectory approximates. The two coincide for a well-relaxed state, as in Table 4, and differ when the state is only partially relaxed. We target the equilibrium sensitivity because deployment is quasi-static, and the dynamics of the forward solver (mass scaling and damping) are numerical devices, not physics.

In the revised paper, we have replaced the claim of superior accuracy over autodiff with the following: the adjoint matches autodiff on the closed-form benchmark, at roughly 6,000× less memory and much lower gradient cost (0.4 s vs. 271 s). On full deployments, unrolled autodiff is not feasible at all (an estimated 7.5 TiB), so the practical advantage of the adjoint is that it makes these gradients computable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.