{"id":"f8eee451-10de-4848-960f-941e47e7fed8","arxiv_id":"2504.13750","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"JANC is a JAX-based, GPU-accelerated, differentiable compressible reacting flow solver with a JIT-compatible block-structured AMR framework, reported to cut detonation simulation cost to about 1-2% of OpenFOAM's.","lead":"The paper presents JANC, a JAX-based solver for compressible reacting flows with a JIT-compatible adaptive mesh refinement framework and automatic differentiation. It reports that a single-GPU JANC run costs about 1 to 2 percent of a 384-CPU OpenFOAM run on a detonation-tube benchmark, and that adjoint optimization can recover an injection parameter from simulated rotating detonation snapshots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1–2% cost claim rests on 'cost per million steps' while OpenFOAM's time step and step count to the same physical time are never reported; the ratio is therefore not established.","rationale":"I agree with the reader's weakest-assumption: the efficiency comparison is the load-bearing pillar of the central claim. The paper's own validations (Sod shock tube, Cantera ignition delays, RDC inverse problem) support the solver's basic correctness and differentiability, and I am not raising an objection to those. But the quantitative central claim—1–2% of OpenFOAM cost—is a number that can change by an order of magnitude depending on a quantity (OpenFOAM's Δt) that the paper does not report. The internal CPMS arithmetic inconsistency and the lack of a matched-accuracy/grid-convergence check make this more than a stylistic omission. Since the reader's CONDITIONAL verdict already captures this (the claim needs additional support), my stress-test does not change the verdict; it sharpens the specific missing datum and the test that would resolve it.","tokens_in":21148,"tokens_out":6858,"duration_ms":63163,"concrete_test":"Ask the authors to report, for the core and AMR detonation benchmarks, the OpenFOAM time step size and total number of steps to t = 10 μs, and recompute CPMS per microsecond of simulated physical time; then repeat the comparison with matched solution error (e.g., equal resolution of the induction zone) rather than matched grid spacing. If the missing step count changes the JANC/OpenFOAM cost ratio by more than 50%, the headline 1–2% claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline efficiency figure is computed as cost per million solver steps (CPMS) from measured step rates, but the paper never states OpenFOAM's time step or the number of steps it took to reach the common physical time t = 10 μs. This is load-bearing because CPMS rewards a solver that takes smaller steps: JANC uses Δt = 0.5×10^-9 s in the uniform-grid case (20,000 steps to t = 10 μs), yet OpenFOAM's Δt is absent. If OpenFOAM's implicit-Euler scheme used only a 2× larger Δt, the ratio rises from about 1% to about 1.8%; at 10× larger Δt it becomes about 9%. Without this number, the 1–2% figure is not a time-to-solution comparison. The comparison also ignores accuracy: WENO5/RK3 is measured against MUSCL/first-order implicit Euler on the same 12.5 μm grid, so cost per million steps is not cost per simulation of comparable fidelity, and no grid-convergence or matched-error study is provided. The reported core-solver arithmetic is also internally inconsistent (0.5 step/s on 384 cores at $0.007/core-hr gives $1493 per million steps, not $1344), which further weakens confidence in the measured baseline. The central functionality may be sound, but the quantitative efficiency claim is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces JAX-AMR, a block-structured adaptive mesh refinement framework implemented in JAX, and JANC, a differentiable compressible reacting-flow solver built on it. The solver solves the 2D reactive Euler equations with detailed finite-rate chemistry using WENO5 spatial reconstruction, third-order TVD Runge-Kutta time stepping, and a point-implicit chemical source term. Validation is provided for the Sod shock tube against the analytical solution and for zero-dimensional ignition delay against Cantera. The main performance claim is that in a 2D detonation-tube benchmark, JANC on one A100 GPU has a cost per million solver steps of about 1% of that of OpenFOAM on 384 CPU cores, and 1.2-2.1% when AMR is enabled. The paper also demonstrates adjoint-based gradient optimization in an inverse problem for a rotating detonation combustor flow field.","tokens_in":21396,"tokens_out":5925,"duration_ms":55842,"significance":"If the efficiency claim were established, a fully differentiable, GPU-native, JIT-compiled reacting-flow solver with AMR would be a genuinely useful contribution to combustion science and to machine-learning-based design. The paper provides open-source code (MIT license), validates the core solver against independent benchmarks (Sod and Cantera), and demonstrates a working adjoint optimization pipeline that avoids the memory blow-up of naive reverse-mode differentiation. The AMR design, based on fixed-shape blocks with predeclared maximum block counts, is a sensible way to reconcile JAX's static-shape JIT requirement with dynamic refinement. These strengths are real, but the headline cost comparison is not currently supported by the data reported.","major_comments":[{"comment":"The cost metric is 'cost per million simulation steps' (CPMS), but the abstract and conclusions present this as 'computational cost reduced to 1% of that of OpenFOAM.' The paper never reports OpenFOAM's time step or the number of steps each solver needed to reach the common physical time t = 10 μs, while JANC's Δt = 0.5×10^-9 s implies 20,000 steps. If OpenFOAM's implicit-Euler scheme used a 10× larger Δt (which is plausible for an implicit scheme), its wall-clock cost to t = 10 μs would be roughly 10× lower than the CPMS-based estimate, changing the claimed ratio from about 1% to about 9%. Please report OpenFOAM's Δt, the step count to t = 10 μs, and the wall-clock time or dollar cost to reach that physical time for both solvers. Without these numbers, CPMS is not a time-to-solution comparison.","section":"§5.2.2, Table 3"},{"comment":"The reported OpenFOAM cost is internally inconsistent. At 0.5 step/s on 384 cores and a rental price of $0.007/core-hr, one million steps take 1e6 / 0.5 s = 2×10^6 s = 555.6 hr, at a total cost of 555.6 × 384 × 0.007 ≈ $1493, not $1344. The same calculation gives JANC: 1e6 / 5.0 s = 2×10^5 s = 55.6 hr at $0.24/hr = $13.3, which is consistent. Please recompute the OpenFOAM CPMS or provide the raw measured timings so the ratio is reproducible.","section":"§5.2.2, CPMS arithmetic"},{"comment":"The AMR comparison inherits the same time-to-solution problem: OpenFOAM's time step and step count to t = 10 μs are not reported, and the refinement thresholds are explicitly adjusted 'to ensure their refined areas and total number of refined grids are similar.' This makes the cost comparison dependent on a threshold choice whose sensitivity is not explored. In addition, the crossover-advance cost (0.83 s per 'large step (8Δt)') is not directly comparable to the synchronous-advance cost without specifying how the multi-level time stepping is composed. Please report wall-clock time or dollar cost to reach the same physical time for each configuration, with error bars or at least multiple runs.","section":"§5.2.3"},{"comment":"The efficiency comparison is presented without any grid-convergence or accuracy-matching study. JANC uses WENO5/RK3 and OpenFOAM uses MUSCL with a minmod limiter and first-order implicit Euler, so cost per step is not cost at comparable fidelity. The paper states that JANC shows 'superior shock-capturing ability' and 'clearer resolution,' but no quantitative error metric is provided. Please include a resolution study or a matched-error comparison (for example, comparing solutions on several grids and reporting the cost to achieve a given error level), or restrict the conclusions to computational speed at identical grid spacing without claiming higher-fidelity efficiency.","section":"§5.2"},{"comment":"The dynamic update strategy for N_i,max contains a logic error as written. In the 'elif' branch, the condition (N_i,valid + 1) < 2.5 N_i,max is satisfied for essentially all valid states; for example, with N_i,max = 100 and N_i,valid = 99, the condition 100 < 250 is true, so N_i,max would be halved to 50 even though all 99 blocks are in use. This would cause oscillation or an invalid state. Presumably the intended reduction condition is something like N_i,valid < N_i,max / 2.5. Please correct the pseudocode to match the actual implementation, or if the implementation has the same behavior, explain how it avoids the problem.","section":"§3.1.4, Algorithm 1"}],"minor_comments":[{"comment":"The performance units are inconsistent: §5.2.2 quotes OpenFOAM speed as 0.5 step/s, while §5.2.3 quotes it as 0.75 s/step. Please use one convention throughout.","section":"§5.2.2, §5.2.3"},{"comment":"Equations (4.3)-(4.5) have garbled symbol rendering in the submitted text (e.g., the NASA polynomial terms and the species enthalpy/specific heat definitions). Please ensure all equations are typeset correctly.","section":"§4.1"},{"comment":"The cloud rental prices are introduced and used without any sensitivity analysis. Given that these are spot-market estimates, please state that they are approximate and perhaps give a range.","section":"§5.2.2"},{"comment":"The claim of being 'the first' JAX-based block-structured AMR framework is strong. If the authors are not aware of prior work, suggest phrasing such as 'to our knowledge, the first' to avoid overclaiming.","section":"§1"},{"comment":"The near-linear speedup with GPUs is stated qualitatively; please include the measured speedup factors (e.g., 1.99×, 3.02×) in the text or figure.","section":"§5.2.2, Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The paper has a genuinely useful contribution: a working, open-source, differentiable reacting-flow solver with JAX-based AMR, validated on standard benchmarks. However, the headline efficiency claim rests on a cost metric that does not compare time-to-solution, the OpenFOAM CPMS number is arithmetically inconsistent, and Algorithm 1 contains a logic error. These are fixable in revision, but they are load-bearing, so a major revision is appropriate. The authors should also be asked to report OpenFOAM's step count and wall-clock times, and to provide a matched-accuracy or grid-convergence discussion before the 1-2% claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"JANC is a new JAX-based compressible reacting flow solver with a clever fixed-shape block AMR design that keeps JIT compilation happy, plus adjoint differentiation and an MIT-licensed codebase. That combination is new, at least relative to the cited literature, and the paper does several things right: the Sod tube and Cantera ignition-delay validations are independent external checks, the RDC inverse problem is a serious end-to-end demonstration of the adjoint machinery, and the authors are honest about current limitations (2D only, redundant blocks, no parallel AMR yet). The code release is a real asset; I can run this myself tomorrow.\n\nThe soft spot is the performance claim. The stress-test concern lands. Cost per million steps (CPMS) is not a time-to-solution metric unless the step sizes are reported, and OpenFOAM's time step is never given. The paper also has an internal arithmetic error: at 0.5 steps/s on 384 cores at $0.007/core-hr, the cost per million steps is about $1493, not the stated $1344. That undermines confidence in the OpenFOAM baseline. The AMR comparison adjusts refinement thresholds to equalize grid counts, which is acceptable for a spot check but not a rigorous method comparison, and there is no grid-convergence study or direct gradient-accuracy check. These are all fixable with more transparent reporting.\n\nThe core functionality, though, looks sound. The validations against Sod and Cantera mean the solver works in the tested regimes. The fixed-shape block trick with NAN sentinels is a useful contribution for anyone doing JAX-based AMR. The self-citation to the authors' previous point-implicit work is legitimate, not padding.\n\nFor a reader building differentiable CFD tooling, this paper is worth a look. I would send it to a computational physics or CFD journal with a request for a careful revision of the benchmark section, not a desk reject. If the authors provide step counts, a matched-accuracy time-to-solution comparison, and correct the OpenFOAM arithmetic, the performance story could be solid. The novel machinery alone justifies referee time. I would cite this for the AMR technique and the open-source implementation.","headline":"A genuinely new, open-source differentiable combustion solver with JIT-friendly block AMR, but the headline 1-2% cost claim rests on an apples-to-oranges benchmark and some arithmetic slips.","tokens_in":21966,"tokens_out":2562,"would_cite":true,"duration_ms":24999,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A JAX-based combustion solver with adaptive mesh refinement reduces detonation simulation cost to about 1 percent of OpenFOAM's, while keeping the entire numerical trajectory differentiable for machine learning.","keywords":["JAX","adaptive mesh refinement","open source","compressible reactive flow","combustion","computational fluid dynamics","automatic differentiation","GPU acceleration"],"falsifier":"Run both solvers on the detonation tube to the same physical end time, for example t = 10 microseconds, with accuracy-matched refinement, and record total wall-clock time and total monetary cost; if OpenFOAM reaches that physical time with no greater total cost, or if JANC requires substantially more steps because of its smaller time step, the 1 percent claim would not hold. A second check: if standard JAX automatic differentiation can compute the rotating detonation gradient on a 40 GB A100 within memory using gradient checkpointing, then the claimed need for adjoint optimization weakens.","tokens_in":1762,"feed_emoji":"🔥","tokens_out":2161,"duration_ms":71728,"temperature":0.7,"pith_summary":"This paper sets out to show that a combustion solver built entirely in JAX, with a block-structured adaptive mesh refinement that is compatible with JIT compilation, can be both dramatically cheaper to run and fully differentiable. The authors claim that in a two-dimensional detonation tube benchmark, the core solver on one A100 GPU costs 1 percent as much per million simulation steps as OpenFOAM running on 384 CPU cores, and that with AMR the cost drops to 1.2-2.1 percent. They further demonstrate an adjoint-based inverse problem in a rotating detonation combustor that converges to the correct injection equivalence ratio in minutes, where default automatic differentiation runs out of memory. If these claims hold, differentiable, GPU-native combustion simulation becomes practical for machine-learning-driven design and inverse modeling.","feed_headline":"GPU combustion solver cuts detonation cost to 1% of OpenFOAM's","feed_subtitle":"JAX-based adaptive mesh refinement plus adjoint optimization makes combustion simulation differentiable for machine learning.","key_machinery":"The load-bearing mechanism is JAX-AMR, a block-structured adaptive mesh refinement scheme in which every refinement layer is a fixed collection of same-shaped JAX arrays with a predeclared maximum block count; refinement is implemented by masked-grid selection, buffer extension, and interpolation between layers, while JIT compatibility is preserved by never changing array shapes at compile time. Invalid blocks are padded, a NAN block marks missing neighbors, and the maximum block count is updated only at irregular intervals, so the solver runs as a single compiled XLA program while the mesh still adapts. The other core mechanism is the point-implicit chemical source update combined with the adjoint method imported from neural ODE work, which makes the entire time trajectory differentiable without storing all intermediate states.","core_discovery":"The central discovery is that the obstacle to JAX-based AMR, namely that JIT compilation requires static array shapes while adaptive meshes change shape, can be removed by a multi-layer block structure with fixed block positions, fixed block shapes, and a predeclared maximum block count per layer, with invalid blocks padded and a dynamically updated block-count strategy to limit overhead. On top of this JAX-AMR framework, the paper builds JANC, a two-dimensional compressible reacting Euler solver with WENO5 reconstruction, Lax-Friedrichs flux splitting, TVD-RK3 time stepping, point-implicit chemical source advancement, and Cantera-supplied thermodynamics. In the benchmark detonation tube, JANC reaches 5.0 steps per second on one A100 GPU versus 0.5 steps per second for OpenFOAM on 384 AMD CPU cores, giving a cost per million steps of $13.30 versus $1344; with AMR the comparable figures are $12.00 versus $555.80 for synchronous advance and $6.70 for the recommended crossover advance. The paper also claims that replacing JAX's default reverse-mode differentiation with adjoint optimization makes a 400,000-grid, 100-step rotating detonation inverse problem solvable at 1.2 optimizer iterations per second, whereas standard automatic differentiation fails with an out-of-memory error.","pith_inferences":["I infer that the 1-2 percent cost figure is not a universal constant: it depends on cloud rental prices, the step counts each solver uses to reach the same physical time, and the AMR threshold choices, so a fairer comparison would report total wall-clock time to a fixed physical time with accuracy-matched refinement.","I infer that if JAX-AMR's dynamic block-count update strategy is optimized beyond the current doubling and halving heuristic, the redundant-block overhead will shrink and the efficiency gap over traditional CPU solvers could widen further.","I infer that the same fixed-shape block machinery could carry over to three-dimensional and multiphysics solvers, potentially making differentiable large-eddy simulation or reactive turbulence modeling feasible, not just Euler-equation detonation.","A testable extension I see is using JANC's adjoint gradients to calibrate Arrhenius parameters of a chemical mechanism directly from detonation speed or ignition delay data, which the current inverse example only hints at."],"forward_implications":["High-resolution detonation and combustion simulations could run on a single GPU at a cost roughly two orders of magnitude below a 384-core CPU cluster, making high-fidelity combustion studies accessible to groups without large HPC allocations.","Because the whole solver, including the time-advance loop, is differentiable through adjoint optimization, inverse problems such as inferring inlet conditions from flow snapshots can be solved with gradient-based optimizers in minutes instead of requiring hand-derived adjoint codes.","JAX-AMR provides a reusable AMR layer for other JAX PDE solvers, not only combustion, because JIT compatibility with dynamic refinement is the general bottleneck it removes.","The MIT-licensed release means the solver can be extended by the community; the paper lists three-dimensional extension, viscous and turbulent models, and immersed boundary methods as natural next steps.","The cost-per-million-steps metric shows JANC's throughput advantage, but the practical gain for a given study depends on how many steps each solver needs to reach the same physical time."],"supporting_citations":[{"why":"Supplies JAX's automatic differentiation, JIT compilation, and GPU or TPU execution that the entire solver is built on.","marker":"[16]"},{"why":"Establishes the prior art of fully differentiable compressible CFD solvers in JAX, which JANC extends to reacting flow with AMR.","marker":"[19]"},{"why":"Provides the block-structured AMR paradigm that JAX-AMR adapts to static-shape JIT constraints.","marker":"[21]"},{"why":"OpenFOAM is the mainstream open-source solver used as the performance baseline in the detonation tube comparison.","marker":"[23]"},{"why":"Supplies the adjoint method for ODE integration used to make JANC's trajectory differentiable with low memory.","marker":"[24]"},{"why":"Provides the NASA polynomial thermodynamic data and mechanism processing used for species properties.","marker":"[25]"},{"why":"Supplies the point-implicit chemical source advancement method used to handle stiff reactions.","marker":"[28]"},{"why":"Provides the 9-species, 19-reaction hydrogen-oxygen mechanism used in validation and detonation benchmarks.","marker":"[29]"},{"why":"Supplies the detonationFoam solver whose OpenFOAM results are compared against JANC.","marker":"[30]"}],"fun_headline_variants":["JANC: GPU detonation solver at 1% of OpenFOAM's cost","Differentiable JANC solver uses JAX AMR to beat OpenFOAM","JAX-based AMR enables 100x cheaper combustion simulations","JANC's adjoint optimization makes whole trajectory differentiable"],"cache_read_input_tokens":24064,"weakest_assumption_plain":"The load-bearing premise is that the OpenFOAM comparison is a fair baseline: it assumes the same physical simulation is being compared, that cloud rental prices represent real costs, and that OpenFOAM's step count and numerical choices do not distort the cost-per-million-steps metric.","fun_headline_variants_meta":{"raw":{"variants":["JANC: GPU detonation solver at 1% of OpenFOAM's cost","Differentiable JANC solver uses JAX AMR to beat OpenFOAM","JAX-based AMR enables 100x cheaper combustion simulations","JANC's adjoint optimization makes whole trajectory differentiable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00038,"raw_usage":{"total_tokens":2123,"prompt_tokens":1156,"completion_tokens":967,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":772,"completion_tokens_details":{"reasoning_tokens":890}},"tokens_in":772,"tokens_out":967,"duration_ms":7265,"temperature":1.0,"reasoning_tokens":890,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:01:12.593944+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both solvers on the detonation tube to the same physical end time, for example t = 10 microseconds, with accuracy-matched refinement, and record total wall-clock time and total monetary cost; if OpenFOAM reaches that physical time with no greater total cost, or if JANC requires substantially more steps because of its smaller time step, the 1 percent claim would not hold. A second check: if standard JAX automatic differentiation can compute the rotating detonation gradient on a 40 GB A100 within memory using gradient checkpointing, then the claimed need for adjoint optimization weakens.","supporting_citations":[{"cited_title":"Belbute-Peres, T","cited_arxiv_id":null,"evidence_quote":"Provides the block-structured AMR paradigm that JAX-AMR adapts to static-shape JIT constraints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adjoint method for ODE integration used to make JANC's trajectory differentiable with low memory."},{"cited_title":"Paszke, S","cited_arxiv_id":null,"evidence_quote":"Provides the NASA polynomial thermodynamic data and mechanism processing used for species properties."},{"cited_title":"Baniabedalruhman, Dynamic meshing around fluid- fluid interfaces with applications to droplet tracking in contraction geometries, Michigan Technological University, 2015","cited_arxiv_id":null,"evidence_quote":"Supplies the detonationFoam solver whose OpenFOAM results are compared against JANC."}],"review_version":1}