{"id":"34a55a37-33e4-4d9b-aa23-5633b73e9a89","arxiv_id":"2412.12360","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FDTDX is an open-source JAX-based FDTD framework with memory-efficient time-reversal automatic differentiation, enabling GPU-accelerated inverse design of 3D photonic nanostructures at billion-cell scales.","lead":"Researchers introduce FDTDX, an open-source software package for designing 3D photonic nanostructures using FDTD simulations with automatic differentiation in JAX. It aims to make large-scale inverse design accessible on GPUs, demonstrating a 92% efficient silicon waveguide bend and a stitching device robust to 2 micrometer fabrication misalignments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The accuracy of the time-reversal gradient is the load-bearing risk: without a numerical gradient check, the optimized device efficiencies (92% bend, stitching robustness) may not reflect the actual objective landscape.","rationale":"I agree with the reader's weakest_assumption. The gradient-accuracy concern is the most load-bearing issue because the paper's main contribution is the time-reversal automatic differentiation method; if gradients are inaccurate, the optimization results and the framework's central value proposition are undermined. The forward validation against Meep does not cover the backward pass. The absence of a gradient check is not a question of internal consistency or consensus, but a direct missing support for the key technical claim. The reader's verdict of CONDITIONAL is appropriate; my proposed test would confirm whether the condition is satisfied. I do not see reason to change the verdict, so I mark UNCHANGED. I also note the paper's open-source release and forward validation are real mitigating evidence, but they do not substitute for a gradient check.","tokens_in":11438,"tokens_out":3374,"duration_ms":31173,"concrete_test":"Run FDTDX's gradient for a small 3D simulation (e.g., the Section 3.2 waveguide bend geometry, or a simplified 3D structure with PML) and compare against central finite differences of the forward objective. Choose a scalar objective such as Poynting flux at the output detector, and compute the FDTDX gradient and the finite-difference gradient for 10 to 20 randomly selected design parameters. Accept the implementation if the relative error is below 1% for all tested parameters; if any errors exceed 10%, the inverse time-stepping is not reliable. Additionally, directly test the inverse update by saving the boundary slices, applying the forward FDTD update for one or more steps, then applying the inverse update and measuring reconstruction error in the interior; it should be at machine precision without PML and small with PML.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 implements reverse-mode AD by inverting the FDTD update using time-reversibility, storing only six 2D slices at the PML interfaces because the PML update is non-invertible. The correctness of this backward pass is assumed, not demonstrated. The forward solver is validated against Meep in Appendix A, but that validates only the forward propagation, not the gradient. The central claim, 'enables gradient-based optimization over large simulation volumes', depends on these gradients being accurate; the reported 92% bend efficiency and 2 µm-robust stitching are the product of this gradient, and if the inverse update is unstable or inaccurate, these results are unsupported. The paper provides no finite-difference comparison, no independent adjoint comparison, and no reconstruction-error test for the inverse update. This is particularly important because the Yee update with sources, material averaging, and interpolation (mentioned in Section 2.1) may break exact reversibility; the boundary saving scheme needs verification that it recovers the full interior fields.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FDTDX, an open-source JAX-based FDTD simulation and inverse-design framework. The central methodological claim is a memory-efficient reverse-mode automatic differentiation that exploits the time-reversibility of Maxwell's equations, storing only six two-dimensional slices at the PML boundaries instead of full three-dimensional field histories. The paper also introduces a relational object API for scene specification and demonstrates multi-GPU scaling. Results include a speed comparison against Meep, Ceviche, OmniSim, and Tidy3D; an optimized silicon waveguide bend with reported 92% efficiency; and a polymer waveguide stitching device robust to translations up to 2 µm. The forward solver is validated against Meep in Appendix A.","tokens_in":11539,"tokens_out":5334,"duration_ms":52843,"significance":"If the gradient computation is correct, FDTDX is a potentially valuable open-source contribution: it provides native GPU execution, automatic differentiation, multi-GPU scaling, and a user-friendly constraint API, all of which address real bottlenecks in photonic inverse design. The forward-solver comparison against Meep in Appendix A is a positive external check, and the two-stage constrained optimization and stitching-robustness experiments are nontrivial demonstrations. However, the central numerical claims depend on the correctness of the reverse time-stepping gradients, which the paper does not validate, and the speed comparison is not a controlled software comparison. These issues must be addressed before the paper's main claims can be accepted.","major_comments":[{"comment":"The reverse time-stepping gradient is never numerically validated. Appendix A validates only the forward solver against Meep, whereas the optimized efficiencies reported in Sections 3.2 and 3.3 are produced by gradients computed through the inverse update. No finite-difference comparison, no independent adjoint comparison, and no reconstruction-error test for the inverse update are reported. This is load-bearing because Eqs. (1)-(2) describe a lossless, sourceless continuum system, while the actual discrete update includes TFSF source injection, CPML, and material averaging/interpolation, any of which can break exact time reversibility. Please add a small-scale gradient accuracy test (e.g., against finite differences or an independent adjoint) and a test that the six saved 2D PML-boundary slices are sufficient to reconstruct the interior fields in the backward pass, with a quantitative error measure.","section":"Section 2.2"},{"comment":"The speed comparison is not a controlled software comparison. FDTDX is run on one or four NVIDIA H100 GPUs, Meep and Ceviche on a 16-core AMD EPYC CPU cluster, OmniSim on a consumer laptop CPU, and Tidy3D on undisclosed remote hardware. The text's claim that \"FDTDX outperforms all other open-source simulation software by a large margin\" and the reported roughly 10x speedup over Meep therefore conflate hardware differences with software performance. Please either benchmark the open-source solvers on comparable hardware (including a CPU-only FDTDX configuration, if necessary) or explicitly restrict the claim to \"on the hardware used in this study\" and remove the general outperformance claim.","section":"Section 3.1, Table 2"}],"minor_comments":[{"comment":"The abstract contains a duplicated word: \"random translational offsets offsets of up to 2 µm.\"","section":"Abstract"},{"comment":"The y-axis label in Fig. 4(a) reads \"Attenuation [dB]\" with positive plotted values, while the text reports \"an attenuation of -0.36dB, or equivalently an efficiency of 92%.\" Please clarify the sign convention (insertion loss versus transmission in dB) so the reader can interpret the optimization curves correctly.","section":"Section 3.2, Figure 4"},{"comment":"The table uses \"×\" entries without a caption explanation. Please state explicitly that \"×\" means the simulation was not run or could not be run on the available hardware/resource limits.","section":"Table 2"},{"comment":"There is a typo: \"the algorithms returns an error message\" should be \"the algorithm returns an error message.\"","section":"Section 2.3"},{"comment":"The Meep validation is presented only as field-distribution images; no quantitative error metric (e.g., maximum or mean normalized error across the planes) is reported. Adding a single error number would make the validation more convincing.","section":"Appendix A"},{"comment":"Reference [13] is cited as \"Quantized inverse design for photonic integrated circuits\" (2024) without a venue, arXiv identifier, or DOI. Please provide a complete citation so readers can locate the prior implementation of the inverse time-stepping method.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a software-presentation-style manuscript whose central contribution is a memory-efficient AD-FDTD implementation. The main risk is overclaiming: the optimized device results rest on an unvalidated backward pass, and the speed comparison is not hardware-controlled. Both issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The open-source release is a genuine strength and should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Useful open-source FDTD inverse-design package, but the paper currently asks reviewers to trust the time-reversal gradient without a single numerical check, and the speed comparisons are on non-comparable hardware. Both are fixable.\n\nWhat is actually new: a JAX-native FDTD with multi-GPU, time-reversal AD storing six 2D PML slices, and an object API with relative positioning and sizing constraints. The relational constraint API is the most novel piece; I have not seen that in other EM simulators. The forward solver is checked against Meep in Appendix A and the field distributions match well, which gives some grounding. The optimization examples (waveguide bend at 92% efficiency, stitching robustness) are plausible demonstrations of the framework.\n\nSoft spots, in order. First and most important: there is no gradient-accuracy check. Section 2.2 describes the inverse update and the six-slice boundary saving, but nothing verifies that the computed gradient matches a finite-difference or adjoint reference. The Yee update with sources, material averaging, and interpolation may not be exactly reversible; the PML boundary reconstruction could lose information. Without this check, the 92% bend and stitching results are unsupported as statements about the true objective. A simple finite-difference comparison on a small problem would settle it. Second, the speed comparison mixes hardware: Meep and Ceviche on CPU, OmniSim on a laptop, Tidy3D on undisclosed cloud hardware, FDTDX on H100. The paper admits this, but the abstract-level claim of outperforming all open-source software is too strong as stated. Third, one sentence in Section 2.2 says time-reversal AD \"has never been implemented in a powerful automatic differentiation framework\" and then says \"we follow previous work of Ref. 13\" to implement it. If Ref. 13 did it, the sentence is wrong; if not, the citation is doing too much work. This needs clarification. Also, devices are simulated only, not fabricated; that is fine for a methods paper, but it should be labeled as such.\n\nNone of these are fatal to the central value. The package is open source, the forward physics checks out against an external tool, and the API is a real usability improvement. The gradient issue is the load-bearing one; I would want it addressed before trusting the optimized devices, but it is easy to address.\n\nThis paper is for people who want a free GPU FDTD inverse-design tool; it deserves a serious referee. I would accept it for review with the expectation of major revision, mostly adding a gradient check and rewriting the performance and novelty claims.","headline":"A genuinely useful open-source GPU FDTD inverse-design tool whose time-reversal gradient is unverified and whose speed claims compare unlike hardware; fixable, worth reviewing.","tokens_in":12197,"tokens_out":2130,"would_cite":true,"duration_ms":19421,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An open-source FDTD package makes gradient-based inverse design of 3D nanostructures practical by reconstructing earlier fields through time-reversed Maxwell updates, needing only six 2D boundary slices per step for memory.","keywords":["FDTD","inverse design","automatic differentiation","time reversal","nanophotonics","JAX","GPU","photonic integrated circuits"],"falsifier":"Run a small 3D FDTD device (e.g., the 3-million-cell bend) and compare the reverse-mode gradient from FDTDX against a central finite-difference gradient or an independent adjoint implementation; if the normalized gradient error exceeds a few percent or the reverse pass is unstable, the time-reversal memory saving is not accurate enough to rely on.","tokens_in":11154,"feed_emoji":"💡","tokens_out":4850,"duration_ms":38102,"temperature":0.7,"pith_summary":"The paper introduces FDTDX, an open-source Python package that performs inverse design of three-dimensional photonic nanostructures by differentiating through Finite-Difference Time-Domain (FDTD) simulations. Its central claim is that Maxwell's equations' time reversibility makes gradient computation memory-efficient: rather than storing the full 3D electromagnetic field at every time step, the backward pass reconstructs earlier fields with an inverse update, needing only six two-dimensional slices at the PML boundaries. This reduces memory enough that devices with millions of design parameters can be optimized on a single consumer GPU, and the package scales from one to multiple GPUs for simulations with billions of grid cells. The authors demonstrate the approach by optimizing a silicon waveguide bend to 92% efficiency (1.6 µm² footprint) and a two-photon-polymerization stitching element robust to random offsets up to 2 µm. If correct, this makes large-scale photonic inverse design accessible to researchers without commercial software licenses.","feed_headline":"Open-source FDTD solver automates 3D nanostructure inverse design","feed_subtitle":"Time-reversal gradients cut memory so millions of parameters fit on one GPU; silicon bend hits 92%.","key_machinery":"The central object is the inverse FDTD update step: because the discretized Maxwell curl updates are invertible in the interior of the simulation domain, the automatic differentiation backward pass can reconstruct earlier field states instead of storing them. Only six two-dimensional slices of field data at the convolutional PML boundaries are checkpointed per time step, cutting memory from a full 3D volume per step to a small boundary set; this is implemented via reverse-mode automatic differentiation in JAX, with the Yee grid providing the staggered discretization that makes the curl operator local and reversible.","core_discovery":"FDTDX claims that gradient-based inverse design for 3D FDTD can be made practical by replacing checkpointed field storage with an inverse time-stepping pass. The electric and magnetic field updates on the Yee grid are reversible for linear materials, so the backward pass recomputes the fields from time step t+1 to t, and only the six 2D slices at the PML boundary need to be saved because the PML itself is not invertible. The authors show this memory-efficient differentiation works within the JAX framework, supports multi-GPU execution, and produces optimized devices: a silicon corner bend with -0.36 dB attenuation (92% efficiency) over 3 million grid cells, and a 3D polymer stitching device that keeps attenuation between -1 dB and -5 dB under random translations up to 2 µm, versus -35 dB for a standard waveguide. They also report speed comparisons where FDTDX outperforms other open-source FDTD solvers and approaches the speed of a commercial cloud solver at moderate resolutions.","pith_inferences":["A direct numerical gradient check (e.g., against finite differences or an adjoint solver) would confirm whether the inverse time-reversal pass is exact; the paper reports no such check, so the 92% bend and stitching results rest on the assumption that the saved PML slices fully capture the reverse dynamics.","The same time-reversal trick should extend to other reversible update schemes beyond FDTD, such as pseudospectral time-domain methods, whenever the interior update is invertible and the boundary condition is the only lossy element.","The stitching-device demonstration suggests a design principle for 2PP fabrication: deliberately insert a collimation and gathering stage so that translational stitching errors are converted into small coupling losses rather than mode mismatch; optimizing for robustness in z-translation is the natural next step the authors name.","Because the framework is open source, the community can implement custom CUDA kernels or alternative boundary conditions without needing a proprietary solver, which may shift the standard practice for large-scale photonic inverse design toward in-house GPU clusters."],"forward_implications":["A single consumer GPU with roughly 20 GB of memory can run forward and backward passes for a 3-million-cell inverse design, making large-scale optimization accessible without specialized hardware.","The same solver scales to 2.3 billion grid cells by distributing across multiple GPUs, which the authors demonstrate up to four H100 cards.","The time-reversal gradient approach incurs only a small runtime overhead compared to checkpointing, because fields are recomputed in the backward pass rather than over many forward sweeps.","The relational object-constraint API lets users specify simulation scenes in relative coordinates, which the authors claim is new among FDTD tools and reduces setup errors.","Devices optimized with this framework achieve concrete performance targets: a 92% efficient silicon bend and a 2PP stitching element that outperforms a plain waveguide for misalignments above 0.5 µm in the y-direction."],"supporting_citations":[{"why":"Introduces time-reversal differentiation of FDTD, the method FDTDX implements for memory-efficient gradients.","marker":"[12]"},{"why":"Presents the prior JAX-based implementation and quantization constraints that FDTDX builds on.","marker":"[13]"},{"why":"Provides the silicon bend design problem and the strict foundry-constraint optimization approach used for the 92% device.","marker":"[16]"},{"why":"Meep is the open-source baseline for speed comparison and the validation target for field-scattering simulations.","marker":"[1]"},{"why":"Defines the Yee grid discretization that the reversible FDTD updates rely on.","marker":"[6]"},{"why":"Supplies the convolutional PML boundary conditions whose non-invertibility forces the six-slice checkpointing.","marker":"[7]"},{"why":"Contributes the random-scattering field dataset used in the Appendix A validation against Meep.","marker":"[29]"}],"fun_headline_variants":["FDTDX: Reversible time steps for 3D photonic design","Time-reversal gradients cut memory for 3D FDTD design","Open-source FDTD inverse design handles millions of parameters","Memory-lean FDTD gradients shape 3D nanostructures","3D photonic inverse design with reversible Maxwell steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that saving only six two-dimensional field slices at the PML boundaries each time step is enough for the inverse time-stepping pass to reconstruct all earlier fields exactly, so the computed gradients match the true objective; the paper does not include a numerical gradient-accuracy check.","fun_headline_variants_meta":{"raw":{"variants":["FDTDX: Reversible time steps for 3D photonic design","Time-reversal gradients cut memory for 3D FDTD design","Open-source FDTD inverse design handles millions of parameters","Memory-lean FDTD gradients shape 3D nanostructures","3D photonic inverse design with reversible Maxwell steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1390,"prompt_tokens":1004,"completion_tokens":386,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":297}},"tokens_in":620,"tokens_out":386,"duration_ms":4232,"temperature":1.0,"reasoning_tokens":297,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:09:30.902912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a small 3D FDTD device (e.g., the 3-million-cell bend) and compare the reverse-mode gradient from FDTDX against a central finite-difference gradient or an independent adjoint implementation; if the normalized gradient error exceeds a few percent or the reverse pass is unstable, the time-reversal memory saving is not accurate enough to rely on.","supporting_citations":[{"cited_title":"Time reversal differentiation of fdtd for photonic inverse design,","cited_arxiv_id":null,"evidence_quote":"Introduces time-reversal differentiation of FDTD, the method FDTDX implements for memory-efficient gradients."},{"cited_title":"Quantized inverse design for photonic integrated circuits,","cited_arxiv_id":null,"evidence_quote":"Presents the prior JAX-based implementation and quantization constraints that FDTDX builds on."},{"cited_title":"Inverse design of photonic devices with strict foundry fabrication constraints,","cited_arxiv_id":null,"evidence_quote":"Provides the silicon bend design problem and the strict foundry-constraint optimization approach used for the 92% device."},{"cited_title":"Meep: A flexible free-software package for electromagnetic simulations by the fdtd method,","cited_arxiv_id":null,"evidence_quote":"Meep is the open-source baseline for speed comparison and the validation target for field-scattering simulations."},{"cited_title":"Numerical solution of initial boundary value problems involving maxwell’s equations in isotropic media,","cited_arxiv_id":null,"evidence_quote":"Defines the Yee grid discretization that the reversible FDTD updates rely on."},{"cited_title":"Convolution pml (cpml): An efficient fdtd implementation of the cfs–pml for arbitrary media,","cited_arxiv_id":null,"evidence_quote":"Supplies the convolutional PML boundary conditions whose non-invertibility forces the six-slice checkpointing."},{"cited_title":"Neural operator-based surrogate solver for free-form elec- tromagnetic inverse design,","cited_arxiv_id":null,"evidence_quote":"Contributes the random-scattering field dataset used in the Appendix A validation against Meep."}],"review_version":1}