Mosaic: A Benchmark Suite for Differentiable Physics Solvers

arXiv:2606.27895 · physics.comp-ph, cs.LG · Submitted 2026-06-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Mosaic: A Benchmark Suite for Differentiable Physics Solvers".

Mira: Here's the gist of the paper. Differentiable partial differential equation (PDE) solvers underpin solver-in-the-loop ML training, gradient-based optimal control, and inverse problems, yet the practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented <ref:2606.27895#pg2>.

Kai: Mosaic introduces an extensible benchmarking framework for differentiable PDE solvers that standardizes access to solver gradients <ref:2606.27895#pg4>.

Mira: The Problem of Opacity The practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented <ref:2606.27895#pg2>.

Kai: Integration effort, computational cost, gradient accuracy, and numerical conditioning vary widely across solvers and are discoverable only by trial and error <ref:2606.27895#pg2>.

Paper summary: Mira: So looking at "Mosaic: A Benchmark Suite for Differentiable Physics Solvers", what it really does is provide this organized way to test different solvers against a standard set of challenges in fluid dynamics, structural mechanics, and heat transfer. It claims that this benchmarking framework surfaces the practical differences between solvers by standardizing access to solver gradients.

Kai: The authors argue that the key finding isn't necessarily about one single solver being perfect, but rather about how different solvers behave when you compare them across these defined metrics like runtime and conditioning.

Lev: And they point out that even if all the solvers produce gradients, they often converge to the same final solution, which means gradient quality might not be the limiting factor for getting a good answer on its own.

Mira: So for us, this suggests we should focus our attention more on how well a solver handles numerical stability and conditioning during long optimization runs rather than just chasing perfect pointwise gradient accuracy at every single step.

Kai: And the paper’s implication is that reverse-mode automatic differentiation seems to offer the best effort to capability tradeoff when working with these complex physics simulations right now.

Lev: So, in simple terms, this work gives us a common testing ground so we know what to expect from different solvers before we commit resources to building the actual ML applications on top of them.

Conclusion: Kai: So we've been looking at Mosaic, and I think the title itself really tells you what this is about—it's a whole system for testing how these math solvers work when you want to use them in machine learning.

Mira: Right. It sets up this standardized way to check different physics solvers against the same set of problems. It’s like setting up a universal lab bench for whatever solver you pick later on.

Kai: Exactly, and it's about making the messy process of getting gradients from these solvers much more predictable for training models with them. The authors are just providing this infrastructure to help people stop guessing how good a solver is going to be.

Lev: From an engineering side, what’s interesting is that they aren't just picking one solver and showing you their results; they’re testing the gradient mechanism itself in isolation first. That means we get a clear view of the fundamental limitations of each approach before we even try to fit it into a larger pipeline.

Mira: And that leads into the core idea—the problem is that right now, figuring out if a solver gives you accurate gradients or if it's numerically stable is totally undocumented. It’s all trial and error for every single solver you look at.

Kai: So what this paper does is create this benchmark suite to standardize those tests—measuring things like how much time it takes, the accuracy of the gradient, and how well the math holds up under pressure. It’s about surfacing these hidden costs.

Lev: And that points toward a crucial detail from their findings—they found that for many problems, if a solver can produce *any* gradient at all, they often end up converging to the same final answer regardless of how much accuracy or conditioning differs between them.

Mira: That’s interesting because it means chasing perfect gradient accuracy isn't always the bottleneck for getting a good result in these physics applications. It suggests that other things are more important.

Kai: So where does this leave us? If we can stop wasting time guessing solver performance and start focusing on stability and conditioning, what kind of real-world systems could we actually build with this standardized testing ground?

Andrin Rehmann, Heiko Zimmermann, Dion Häfner

Pasteur Labs

physics.comp-ph, cs.LG

Submitted: 2026-06-26

Updated: 2026-11-09

Comments: 37 pages, 18 figures, 11 tables. NeurIPS 2026 Evaluations and Datasets Track. Code available at https://github.com/pasteurlabs/mosaic

Code: https://github.com/pasteurlabs/mosaic

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 87/100

The gist: Differentiable partial differential equation (PDE) solvers underpin solver-in-the-loop ML training, gradient-based optimal control, and inverse problems…

Key concepts

Solver Gradients
These are the mathematical derivatives calculated by a PDE solver to determine how small changes in input parameters affect the solution. Mosaic tests how accurately and reliably different solvers can produce these gradients for various physical problems.
Mosaic Infrastructure
This is an extensible benchmarking suite that standardizes testing. It packages solvers into containers with a uniform API, allowing researchers to test the solver's gradient mechanism without worrying about the underlying programming language or automatic differentiation strategy.
Jacobian Conditioning
This measures how sensitive the solver's output is to small changes in its inputs. Poor conditioning means small errors in gradients can lead to large errors in the final solution, which directly impacts training stability and convergence during machine learning.
Rollout Length
This refers to how many steps or iterations a process takes when calculating a gradient or performing an optimization step. The paper shows that Jacobian conditioning often degrades as rollout length increases, meaning longer processes can introduce more numerical instability.

Terminology

Summary

Differentiable partial differential equation (PDE) solvers underpin solver-in-the-loop ML training, gradient-based optimal control, and inverse problems, yet the practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented <ref:2606. p. 0

Where the paper says it

haefner@simulation. science Abstract Differentiable partial differential equation (PDE) solvers underpin solver-in-the-loop ML training, gradient-based optimal control, and inverse problems, yet the practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented.

27895#pg2>. Mosaic introduces an extensible benchmarking framework for differentiable PDE solvers that standardizes access to solver gradients <ref:2606. p. 0
Where the paper says it

Integration effort, computational cost, gradient accuracy, and numerical conditioning vary widely across solvers and are discoverable only by trial and error. We introduce Mosaic, an extensible benchmarking framework for differentiable PDE solvers that standardizes access to solver gradients.

27895#pg4>.

The Problem of Opacity

The practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented <ref:2606.27895#pg2>. Integration effort, computational cost, gradient accuracy, and numerical conditioning vary widely across solvers and are discoverable only by trial and error <ref:2606. p. 0

Where the paper says it

science Abstract Differentiable partial differential equation (PDE) solvers underpin solver-in-the-loop ML training, gradient-based optimal control, and inverse problems, yet the practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented. Integration effort, computational cost, gradient accuracy, and numerical conditioning vary widely across solvers and are discoverable only by trial and error.

27895#pg2>. Solvers range from turnkey applications to composable frameworks and low-level kernel toolkits, each demanding different expertise and integration effort <ref:2606. p. 1
Where the paper says it

Yet the practical cost of obtaining correct, usable gradients from any of these solvers is opaque. Solvers range from turnkey applications to composable frameworks and low-level kernel toolkits, each demanding different expertise and integration effort.

27895#pg3>. Runtime overhead, gradient accuracy, and numerical conditioning have not been systematically measured, and solver-specific constraints surface only after a practitioner has committed to a particular solver <ref:2606. p. 1
Where the paper says it

Solvers range from turnkey applications to composable frameworks and low-level kernel toolkits, each demanding different expertise and integration effort. Runtime overhead, gradient accuracy, and numerical conditioning have not been systematically measured, and solver-specific constraints (periodic-only boundary conditions, incompatible penalization schemes, stability limits) surface only after a practitioner has committed to a particular solver.

27895#pg2>.

The Mosaic Infrastructure

Mosaic is an extensible benchmark suite that tests solver gradients in isolation, on standalone optimization problems without learned components where finite differences provide ground truth and the solver’s gradient mechanism is the only variable under test <ref:2606.27895#pg3>. The primary contribution is the benchmark infrastructure itself, not the empirical results from any single instantiation <ref:2606. p. 1

Where the paper says it

We introduce Mosaic, an extensible benchmark suite that tests solver gradients in isolation, on standalone optimization problems without learned components, where finite differences provide ground truth and the solver’s gradient mechanism is the only variable under test. The primary contribution is the benchmark infrastructure itself, not the empirical results from any single instantiation.

27895#pg4>. Mosaic consists of three artifacts: a containerized solver interface, an evaluation protocol, and a library of benchmark task templates <ref:2606. p. 3
Where the paper says it

Each panel shows the control variable, the physical process, and the optimization objective. 2 Design principles Mosaic consists of three artifacts: a containerized solver interface, an evaluation protocol, and a library of benchmark task templates.

27895#pg4>. Each solver is packaged as a containerized component (Tesseract) exposing a uniform gradient API regardless of language or automatic differentiation (AD) strategy <ref:2606. p. 0
Where the paper says it

We introduce Mosaic, an extensible benchmarking framework for differentiable PDE solvers that standardizes access to solver gradients. Each solver is packaged as a containerized component (Tesseract) exposing a uniform gradient API regardless of language or automatic differentiation (AD) strategy, enabling researchers to evaluate, compare, and build on non-trivial physical solvers.

27895#pg2>.

Standardized Evaluation Protocol

The evaluation protocol measures gradient accuracy, computational cost, Jacobian conditioning, setup compatibility, and end-to-end optimization convergence <ref:2606.27895#pg4>. Setup compatibility is recorded by checking whether the solver produces a gradient, fails numerically, or is structurally unable to represent the task <ref:2606. p. 4

Where the paper says it

Setup compatibility. For each solver–task pair, we record whether the solver produces a gradient, fails numerically, or is structurally unable to represent the task (Section 4.

27895#pg4>. Gradient accuracy is assessed by comparing AD gradients against central finite differences at solver-specific optimal perturbation size <ref:2606. p. 6
Where the paper says it

jl uses Zygote reverse-mode AD yet still requires over 470 shared lines of gradient-specific code, illustrating that AD framework support does not always translate into turnkey gradients. 2 Gradient quality To assess whether AD gradients are correct, we compare each solver’s gradients against central finite differences at solver-specific optimal perturbation size.

27895#pg4>. Performance metrics include primal and VJP wall-clock time, their ratio, and peak (V)RAM consumption measured on each solver’s intended hardware target <ref:2606. p. 4
Where the paper says it

Performance. Primal and VJP wall-clock time, their ratio, and peak (V)RAM consumption, measured on each solver’s intended hardware target across three to four problem sizes (grid resolutions or cell counts).

27895#pg4>.

Empirical Findings on Solver Selection

The evaluation demonstrates that solver choice can determine whether gradient-based workflows succeed or fail, independent of the optimizer <ref:2606.27895#pg4>. On well-conditioned problems, pick the fastest solver <ref:2606. p. 9

Where the paper says it

Memory scales likewise, and several solvers exceed GPU limits before gradient quality becomes relevant. On well-conditioned problems, pick the fastest solver.

27895#pg4>. Reverse-mode AD offers the best effort-to-capability tradeoff <ref:2606. p. 9
Where the paper says it

Because gradient quality is not a differentiator in these regimes, the selection criterion reduces to forward-solve speed and integration effort. Reverse-mode AD offers the best effort-to-capability tradeoff.

p. 9
Where the paper says it

Because gradient quality is not a differentiator in these regimes, the selection criterion reduces to forward-solve speed and integration effort. Reverse-mode AD offers the best effort-to-capability tradeoff.

27895#pg4>. The optimal perturbation size shifts by orders of magnitude across solvers and with rollout length, reinforcing that conditioning is likely a better predictor of training stability than pointwise gradient accuracy <ref:2606. p. 9
Where the paper says it

While more complex tasks may reveal cases where gradient quality becomes limiting, the pattern on these benchmarks suggests that conditioning is likely a better predictor of training stability than pointwise gradient accuracy. Rollout length is the critical scaling parameter.

27895#pg4>.

Key Observations on Gradient Quality

Gradient quality is often good enough; the real bottlenecks lie elsewhere <ref:2606.27895#pg4>. On every task where they can run, all solvers that produce gradients converge to equivalent optima despite order-of-magnitude differences in accuracy and conditioning <ref:2606. p. 9

Where the paper says it

Gradient quality is often good enough; the real bottlenecks lie elsewhere. On every task where they can run, all solvers that produce gradients converge to equivalent optima despite order-ofmagnitude differences in accuracy and conditioning.

p. 9
Where the paper says it

Gradient quality is often good enough; the real bottlenecks lie elsewhere. On every task where they can run, all solvers that produce gradients converge to equivalent optima despite order-ofmagnitude differences in accuracy and conditioning.

27895#pg4>. Rollout length is the critical scaling parameter, and Jacobian conditioning degrades with rollout length across all solvers <ref:2606. p. 9
Where the paper says it

Rollout length is the critical scaling parameter. Jacobian conditioning degrades with rollout length across all solvers (Figure 5), with projection-based solvers deteriorating faster than LBM or spectral methods.

27895#pg4>.

Limitations of the Benchmark

The evaluation rests on three assumptions: that central finite-difference gradients at optimal perturbation size are a reliable reference <ref:2606.27895#pg4>. That out-of-the-box solver configurations are representative of practitioner experience <ref:2606. p. 9

Where the paper says it

Limitations. The evaluation rests on three assumptions: that central finite-difference gradients at optimal perturbation size are a reliable reference (Figure 14), that out-of-the-box solver configurations are representative of practitioner experience, and that standalone optimization tasks are a meaningful proxy for gradient quality in broader pipelines.

27895#pg4>. And that standalone optimization tasks are a meaningful proxy for gradient quality in broader pipelines <ref:2606. p. 9
Where the paper says it

Limitations. The evaluation rests on three assumptions: that central finite-difference gradients at optimal perturbation size are a reliable reference (Figure 14), that out-of-the-box solver configurations are representative of practitioner experience, and that standalone optimization tasks are a meaningful proxy for gradient quality in broader pipelines.

27895#pg4>. Gradient quality may differ on untested setups, and results are pinned to specific solver versions <ref:2606. p. 9
Where the paper says it

The evaluation rests on three assumptions: that central finite-difference gradients at optimal perturbation size are a reliable reference (Figure 14), that out-of-the-box solver configurations are representative of practitioner experience, and that standalone optimization tasks are a meaningful proxy for gradient quality in broader pipelines. Gradient quality may differ on untested setups; the template mechanism is the mitigation.

27895#pg5>.

Future Directions

Natural extensions include additional domains (multiphysics, compressible flow), solver backends (Enzyme/LLVM-level AD), and hybrid physics-ML tasks <ref:2606.27895#pg4>. The infrastructure is in place, and we invite community contributions <ref:2606. p. 9

Where the paper says it

Natural extensions include additional domains (multiphysics, compressible flow), solver backends (Enzyme/LLVM-level AD), and hybrid physics-ML tasks. The infrastructure is in place, and we invite community contributions.

27895#pg4>. Mosaic’s uniform interface makes each backend a drop-in differentiable simulator for solver-in-the-loop or hybrid physics-ML workflows <ref:2606. p. 9
Where the paper says it

What Mosaic enables beyond this evaluation. Mosaic’s uniform interface makes each backend a drop-in differentiable simulator for solver-in-the-loop or hybrid physics-ML workflows, and the template mechanism lets the benchmark grow with the community.

27895#pg4>. The template mechanism lets the benchmark grow with the community <ref:2606. p. 9
Where the paper says it

What Mosaic enables beyond this evaluation. Mosaic’s uniform interface makes each backend a drop-in differentiable simulator for solver-in-the-loop or hybrid physics-ML workflows, and the template mechanism lets the benchmark grow with the community.

27895#pg4>.

Solver Specific Constraints

Specific limitations were identified during benchmarking, such as PICT's viscosity not being differentiable in PISOtorch diff <ref:2606.27895#pg5>. JAX-CFD is structurally excluded from the drag optimization task because its spectral pressure solve requires periodic boundary conditions, which are incompatible with the non-periodic geometry of that task <ref:2606. p. 9

Where the paper says it

5%, because the projection prevents non-solenoidal components from corrupting the curvature model. The 2D drag minimization (F2) is the hardest deployment test, and also the task where structural incompatibilities thin the solver field most: only three of seven backends support the non-periodic, obstacle-laden setup required for drag optimization.

27895#pg5>. INS.jl is excluded because Zygote. p. 1
Where the paper says it

Adjoint methods for PDE-constrained optimization date back to Pironneau [43] and Jameson [27], with discrete adjoint formulations following in [19, 40]. Modern AD frameworks now differentiate solver code automatically: JAX source transformation [12], PyTorch operator overloading [41], Julia source-to-source AD via Zygote [26], LLVM-IR-level AD via Enzyme [39], NVIDIA Warp kernel-level AD [37], and

jl cannot differentiate through the diffusion term in IncompressibleNavierStokes. p. 1
Where the paper says it

Adjoint methods for PDE-constrained optimization date back to Pironneau [43] and Jameson [27], with discrete adjoint formulations following in [19, 40]. Modern AD frameworks now differentiate solver code automatically: JAX source transformation [12], PyTorch operator overloading [41], Julia source-to-source AD via Zygote [26], LLVM-IR-level AD via Enzyme [39], NVIDIA Warp kernel-level AD [37], and

jl due to its ChainRulesCore rule returning NoTangent for the viscosity argument <ref:2606.27895#pg5>. XLB exhibits an intrinsic O(Ma 2) compressibility error inherent to the D2Q9 BGK lattice Boltzmann method <ref:2606. p. 4
Where the paper says it

Success is defined as reaching a final objective within 1% of the best solution achieved across all solvers within the budget. 4 Included solver backends The solver set (Table 2) is chosen to maximize diversity along three axes: AD strategy (autodiff, tape-based discrete adjoints, hand-derived/kernel adjoints), numerical method (projection-based, lattice Boltzmann, finite-volume, spectral, finite elements with both direct and iterative linear solvers), and solver abstraction level.

27895#pg5>.

Optimization Strategy Insights

L-BFGS+proj and Adam+proj variants on F3 apply a spectral Helmholtz projection to the gradient before each optimizer update, which is specific to second-order optimizers <ref:2606.27895#pg5>. This projection restricts the search direction to the divergence-free subspace and is exact to machine precision on the periodic NS state space <ref:2606. p. 9

Where the paper says it

, 2026 10 gradient onto the divergence-free subspace before each L-BFGS update (L-BFGS+proj) drops the IC error below 0. 5%, because the projection prevents non-solenoidal components from corrupting the curvature model.

27895#pg5>. The benefit of this projection is specific to L-BFGS, as Adam’s per-coordinate second-moment scaling accumulates variance estimates from both solenoidal and irrotational gradient directions without a cross-coordinate curvature model to exploit the constraint geometry <ref:2606. p. 9
Where the paper says it

, 2026 10 gradient onto the divergence-free subspace before each L-BFGS update (L-BFGS+proj) drops the IC error below 0. 5%, because the projection prevents non-solenoidal components from corrupting the curvature model.

27895#pg4>.

Physical Accuracy Validation

All solvers that support the periodic TGV domain pass both checks with errors below 10−2 <ref:2606.27895#pg5>. For F2 cylinder-flow forward accuracy, three solvers (JAX-CFD, INS. p. 0

Where the paper says it

II ··· Standardized interface Tesseract apply(x) y vjp(x, v) g Benchmark tasks forward solve optimized inflow (via gradient) Evaluation results Fwd VJP Grad Opt. PhiFlow XLB PICT JAX-CFD INS.

jl, Warp-NS) produce invalid results on this geometry due to boundary condition incompatibilities <ref:2606.27895#pg5>. Solvers that fail on this geometry are omitted <ref:2606. p. 0
Where the paper says it

Diverse solvers are wrapped behind a uniform containerized interface, enabling cross-solver comparison on shared benchmark tasks involving different physical domains. arXiv:2606.

27895#pg5>. The divergence RMS measures how strongly the solver violates the incompressibility constraint, and for a well-posed incompressible solver this should be near machine precision ≲ 10−6 <ref:2606.27895#pg5>.

Cost and Scaling Behavior

Absolute solver cost varies by orders of magnitude at the same resolution <ref:2606.27895#pg4>. The fastest backends (Exponax, JAX-CFD) support only periodic boundaries and no solid obstacles <ref:2606. p. 6

Where the paper says it

3 Computational cost and scaling Absolute solver cost varies by orders of magnitude at the same resolution (Table 3; full scaling curves in Section B. 3), but speed and generality are entangled: the fastest backends (Exponax, JAX-CFD) support only periodic boundaries and no solid obstacles, while solvers that handle the full F2 drag task (XLB, PICT) are up to ∼200× slower.

27895#pg4>. Cost comparisons are therefore meaningful only within the set of solvers that can represent a given task <ref:2606. p. 6
Where the paper says it

3), but speed and generality are entangled: the fastest backends (Exponax, JAX-CFD) support only periodic boundaries and no solid obstacles, while solvers that handle the full F2 drag task (XLB, PICT) are up to ∼200× slower. Cost comparisons are therefore meaningful only within the set of solvers that can represent a given task (Section 4.

27895#pg4>. The VJP/forward ratio is roughly constant across problem sizes for source-transformation AD (JAX), while for structural and thermal domains, it sits close to 1 at moderate sizes because the adjoint of a symmetric elliptic system shares the same stiffness matrix as the forward solve <ref:2606. p. 8
Where the paper says it

Among solvers that remain within budget, the cost of obtaining a gradient depends on the AD strategy. Source-transformation AD (JAX-based solvers) maintains a nearly constant VJP/forward ratio across problem sizes, while tape-based adjoints (FEniCS, Firedrake) scale more steeply (Section B.

27895#pg4>.

Convergence and Optimization Performance

Adam stagnates above 40% IC error across all solvers <ref:2606.27895#pg4>. L-BFGS converges to substantially lower loss and below 6% IC error in fewer iterations <ref:2606. p. 8

Where the paper says it

Adam stagnates above 40% IC error across all solvers (Figure 4): the loss decreases, but the recovered IC differs noticeably from the target. L-BFGS converges to substantially lower loss and below 6% IC error in fewer iterations.

27895#pg4>. MMA, by contrast, converges to compliance ≈0. p. 3
Where the paper says it

H Heat transfer (slab) Conductivity inversion 128 5 Negative control: smooth physics isolates AD bugs S Structural mech. (cantilever) Compliance minimization (SIMP) 2 048 5 Well-studied topology optimization with known optimal F2 Incompr.

00099 in only 40 iterations, matching Adam’s solution in under 2% of its iteration budget <ref:2606. p. 4
Where the paper says it

We run the optimizers with each solver’s gradients on each benchmark task for a fixed iteration budget (500 iterations for F2 and F3; 2 000 for H; 2 500 for S) and report convergence curves and final objective value. Success is defined as reaching a final objective within 1% of the best solution achieved across all solvers within the budget.

27895#pg4>. This shows that for smooth, well-conditioned physics, gradient quality does not limit solution quality <ref:2606. p. 2
Where the paper says it

Tasks are intentionally simpler than state-of-the-art challenges: complexity in the physics would confound the gradient properties we aim to isolate, so any failure can be attributed to the solver’s gradient mechanism rather than to an ill-posed optimization problem. The heat domain (H) additionally serves as a negative control: its smooth, well-conditioned physics means that any gradient failure on H is attributable to an AD bug rather than to chaotic or ill-conditioned dynamics.

27895#pg4>.

Gradient Verification Across Domains

All differentiable solvers reach cosine similarity > 0.999 at their optimal perturbation size ε, confirming gradient correctness across domains <ref:2606. p. 6

Where the paper says it

999 between AD and central finitedifference gradients at their optimal perturbation size 𝜀, confirming AD correctness; per-solver minimum FD errors are reported in the Error (FD) column of Table 3. The optimal 𝜀 shifts by orders of magnitude across solvers and with rollout length and viscosity; see Section B.

27895#pg5>. The U-shaped error curve reflects the truncation/roundoff tradeoff, with the minimum shifting rightward and rising as the rollout grows, reflecting increasing Jacobian ill-conditioning <ref:2606.27895#pg5>. Solvers with rapidly decaying singular value spectra lose useful gradient signal faster <ref:2606. p. 8
Where the paper says it

5s high ν, T=0. 5s low ν, T=1s low ν, T=2s Figure 5 | Normalized Jacobian singular value spectra (𝜎𝑖/𝜎0) for six solvers on the 3D periodic NS problem at four rollout lengths.

27895#pg5>.

The Role of Rollout Length

Even solvers with correct gradients at short horizons can degrade at longer rollouts, consistent with the Jacobian conditioning analysis <ref:2606.27895#pg5>. The optimal perturbation size ε also drifts with horizon, reinforcing that a fixed FD step used for validation can give misleading results at longer rollout lengths <ref:2606.27895#pg5>. Faster spectral decay in projection-based solvers is attributable to singular vectors associated with non-solenoidal velocity modes which incompressible flow forbids by construction <ref:2606. p. 6

Where the paper says it

Conditioning varies dramatically across solvers at the same resolution, reflecting the numerical method rather than the AD strategy: the same governing equations discretized with lattice Boltzmann versus pressure projection yield radically different spectra. The sharp spectral drop-off in projection-based solvers is attributable to singular vectors associated with non-solenoidal velocity modes, which incompressible flow forbids by construction [10, 20].

27895#pg5>.

Final Conclusion on Practicality

Gradient quality is often good enough; the real bottlenecks lie elsewhere <ref:2606.27895#pg4>. On every task where they can run, all solvers that produce gradients converge to equivalent optima despite order-of-magnitude differences in accuracy and conditioning <ref:2606.27895#pg4>. Conditioning is likely a better predictor of training stability than pointwise gradient accuracy <ref:2606. p. 9

Where the paper says it

On every task where they can run, all solvers that produce gradients converge to equivalent optima despite order-ofmagnitude differences in accuracy and conditioning. While more complex tasks may reveal cases where gradient quality becomes limiting, the pattern on these benchmarks suggests that conditioning is likely a better predictor of training stability than pointwise gradient accuracy.

27895#pg4>. On the negative-control domains (H, S), all solvers produce equivalent results regardless of AD strategy or numerical method <ref:2606. p. 9
Where the paper says it

On well-conditioned problems, pick the fastest solver. On the negative-control domains (H, S), all solvers produce equivalent results regardless of AD strategy or numerical method.

27895#pg4>. Reverse-mode AD offers the best effort-to-capability tradeoff <ref:2606.27895#pg4>. The uniform interface makes each backend a drop-in differentiable simulator for solver-in-the-loop or hybrid physics-ML workflows <ref:2606. p. 9
Where the paper says it

What Mosaic enables beyond this evaluation. Mosaic’s uniform interface makes each backend a drop-in differentiable simulator for solver-in-the-loop or hybrid physics-ML workflows, and the template mechanism lets the benchmark grow with the community.

27895#pg4>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Create solver-in-the-loop ML training pipelines by using Mosaic to evaluate which solver fits a problem, as Mosaic’s uniform interface makes each backend a drop-in differentiable simulator for solver-in-the-loop or hybrid physics–ML workflows.

  2. Stress test failure modes before committing to a particular stack by using Mosaic's standardized evaluation protocol, as it allows researchers to stress–test failure modes before committing to a particular stack.

  3. Build solver selection tools that predict training stability by analyzing Jacobian conditioning and rollout length, as conditioning predicts which solvers will struggle with second-order methods and long-horizon problems.

  4. Automate the creation of new solver backends by following the documented procedure, as Contributing a new solver requires implementing the three files described above for an existing domain.

Abstract

Differentiable partial differential equation (PDE) solvers underpin solver-in-the-loop MLtraining, gradient-based optimal control, and inverse problems, yet the practical cost of obtaining correct, usable gradients from a given solver on a given problem is largely undocumented. Integration effort, computational cost, memory, gradient accuracy, and numerical stability vary widely across solvers and are discoverable only by trial and error. We introduce Mosaic, an extensible benchmarking framework for differentiable PDE solvers that standardizes access to solver gradients. Each solver is packaged as a containerized component (Tesseract) exposing a uniform gradient API regardless of language or automatic differentiation (AD) strategy, enabling researchers to evaluate, compare, and build on non-trivial physical solvers. Our evaluation of 14 solvers across fluid dynamics, structural mechanics, and heat transfer demonstrates that the benchmark surfaces practically relevant differences: order-of-magnitude variation in computational cost and memory, alongside setup incompatibilities and instabilities that rule out solvers for specific tasks. Despite this variation, all solvers that produce gradients converge to similar optima on our tasks, indicating that the practical barriers are memory limits, numerical stability, and setup compatibility rather than gradient accuracy alone. Mosaic is open-source and available at github.com/pasteurlabs/mosaic.

Sources

Related papers