Roadmap#

This page lists functionality solprop does not have yet, and known issues in what it has. It is not a schedule: the items are grouped by area, and none is promised. Some qualification sequences state the order in which their checks are needed. What exists is described in the user manual; how it is built, in the rest of these developer pages.

Every item is weighed against the project’s priorities (ADR-020 in decisions.md): accuracy and reliability first, then CPU performance, then GPU. Taking one up follows ADR-033: an issue or pull request, an analytic or independent reference with a stated tolerance, and the design pages updated in the same change. A major item — a new model, route or numerical scheme, a changed default or convention — starts as an enhancement proposal (SEP, ADR-035). Functionality that Luna.jl has and solprop lacks is added when a use case needs it; the deliberate differences are listed on Differences from Luna.

Accepted enhancement proposals#

SEPs that have been accepted and are not yet final, each waiting to be picked up or being implemented.

SEP 17 — Broadband two-photon continuum response beside strong-field ionisation: prescribed-field diagnostics first, then an Ar/He continuum driven by the squared VUV field, with polarisation and optical work from one interaction. The first model gives final yield and optical feedback; additional plasma needs a separate electron-timing qualification. Existing optmat OEP 6 quantities suffice for this approximation. Before production implementation, its qualification covers threshold data and completion, reactive matching, atomic validity and the complex convolution’s work, passivity and window convergence. A coupled optmat OEP is needed only if qualification identifies a new material contract. Ne/Xe, mixtures, mode-averaged and multimode routes, bound response and other geometries remain follow-up work; the first route is a single-mode capillary with node quadrature.

Draft enhancement proposals#

SEPs under discussion; their scope may change and implementation has not been accepted. The SEP index lists every proposal and its current status.

SEP 1 — Efficient causal Raman evaluation: exact sampled recurrences, including the current lag taper and silica endpoints, beside default exact FFT evaluation. A common execution contract retains analytic poles, permits exact FFT evaluation of density-shifted lines and supports SEP 10’s explicit approximation. Population dynamics are separated into SEP 32.
SEP 2 — Two-photon absorption: constant-coefficient, narrowband scalar absorption on field and envelope grids. Broadband and tensor absorption, and solid carrier dynamics, remain outside its scope.
SEP 4 — Bounded execution autotuning: explicit CPU policy search on short windows, with isolated processes, numerical validation, bounded total cost and reproducible execution plans. Native GPU controls, chunking, propagator selection and implicit tuning remain open below.
SEP 5 — Solid ionisation and bulk filamentation: scalar bulk silica/YAG photoionisation, explicit simple avalanche and carrier removal, collisional currents, optmat solid rate models and parameter bundles (which need upstream optmat work first), using SEP 34’s shared medium flux and radial adiabatic transport. Delayed-heating kinetics, subcycle rates and solid-guide ionisation remain follow-up work below; source-normalisation questions precede acceptance.
SEP 7 — Laser gain dynamics in solid-core fibres: Yb/silica spectroscopy through optmat, conventional steady and time-dependent populations, pump and ASE transport, and scalar spatial gain dynamics for nonlinear amplifiers. The broadband filtered-field closure has a separate qualification gate; general vector gain, other ions, stochastic field ASE and oscillators remain open.
SEP 8 — Second-order nonlinearity and birefringent crystal optics: vector forward propagation in homogeneous crystals, quadratic frequency mixing, BBO/LBO material curation, field/envelope grids, interfaces, physical flux and a separately qualified statistical OPG model. Guided overlaps, varying crystal profiles, resonant THz response and bidirectional propagation remain open.
SEP 9 — Broadband beam measurements and propagation statistics: frequency-dependent modal fluence and time-resolved beam widths, spectral filters, per-mode durations, transverse peak intensity and distinct quadrature/modal-span diagnostics. General-guide embedded quadrature and current-source span diagnostics remain open; SEP 34 covers the shared medium-flux correction.
SEP 10 — Low-rank Raman kernels along density paths: explicitly requested spectral compression of density-dependent Raman kernels, extending SEP 1’s contract and exact shifted-line FFT reference. Separate basis and coefficient-path budgets, independent-state validation and observable convergence bound the approximation; zero tolerance selects exact evaluation.
SEP 11 — Finite-density gas index and mixture optics: optmat’s Lorentz–Lorenz law with refractivity virials, applied once to a mixture, and a composition-preserving mixture EOS. One resolved state feeds linear optics, Raman amplitudes, widths and shifts. Current upstream model selections, estimates and validity records govern admission; nonlinear local-field corrections remain open.
SEP 21 — Temporal boundaries: wider windows, absorption and the escaped field: wider production windows without temporal absorption versus collar strategies at matched physical accuracy, including fixed asymmetric placement, prescribed frames, memory and guard-monitor costs. Its qualification decides between a monitored opt-in and a default, with the warning/error policy, before adoption. An implicit–explicit collar is pursued only where it can improve on the admissible retained-window alternatives. Absorbed-energy and occupancy diagnostics, keyword switches, and an opt-in escaped-field record after its own qualification remain in scope. Plasma comparisons use the held-current source, and the affected CPU free-space cost recommendations are remeasured on the current free-space evaluation. Spatial collar changes, adaptive recentering/expansion, a general offset-grid API, independently asymmetric absorber ramps and implicit Runge–Kutta remain open.
SEP 32 — Raman Maxwell–Bloch dynamics: a separate scalar two-level population/coherence response, with equilibrium subtraction, a weak-response limit and local temporal integration. Its ensemble, data and integration questions must be settled before acceptance; shared-level ladders, vector saturation and repeated-window histories remain future work.

The informal items below record work not fully covered by these proposals. When adding or revising a SEP, remove an informal item only if the proposal covers it in full; otherwise link the SEP, state what it covers and keep the remaining work explicit. Moving work into a proposal does not mean it is implemented.

Physics and models#

Spectral-domain responses. A response declares its domain, and spectral is one of the declared values (ADR-011). SEP 2 partly covers this item with a scalar, constant-coefficient, narrowband two-photon loss current. SEP 6 covers dispersive electronic third-order response in qualified transparent gases, with two FFT-based evaluators. SEP 8 covers second-order (\(\chi^{(2)}\)) mixing in homogeneous birefringent crystals, with vector linear/source operators, orientation, walk-off and explicit nonlinear dispersion. Guided quadratic overlaps, resonant THz/phonon response and general anisotropic third-order response remain open, as do broadband mixed-frequency and tensor absorption. SEP 17 proposes an effective continuum optical response for Ar/He with memory and a common polarisation/work derivation. Its pair-sum approximation does not determine general nondegenerate or channel-coherent amplitudes. Those, other species, solid responses and additional electron timing remain open. Ne/Xe, mixtures and mode-averaged or multimode gas configurations require later qualification beyond the first single-mode capillary implementation.

Beyond the scalar dispersive electronic response. SEP 6’s evaluators are scalar, off-resonant and for gases. Open beyond it:

tensor dispersion: the vector response holds the isotropic Kleinman ratio \(\chi_{xxyy}=\chi_{xxxx}/3\) at every frequency, where the tensor components of a gas disperse differently; independently supported components are needed first;
fused silica: its degenerate \(n_2\) curve does not determine its nondegenerate response, and no curated nondegenerate data exist to factorise, so silica keeps the carrier-evaluated cubic;
the GNLSE route, whose fitted \(\gamma\) is one constant and which refuses a dispersive policy by name;
the bounded region’s few xenon and CO quartets that mix legs on either side of a 400 nm carrier and are not shown better than it (SEP 6 §4.10): whether they need a narrower bound.

The constant cubic’s number-conserving rule beyond an octave. The field-grid thg=False rule \(3|E_a|^2E/4\) and the general modal route’s kerr_form="analytic" lack the projection onto positive frequencies, \(\Pi_+\), that the dispersive responses’ number-conserving sector applies. Below an octave the two agree exactly; beyond it the constant rule does a little net work on the field (SEP 6, §4.9). Adding \(\Pi_+\) to it changes results only for fields wider than an octave.

The strong-field approximation for harmonic generation. A single-atom dipole response computed in the strong-field approximation (the Lewenstein model), driven by the propagated field, would model high-harmonic generation and its propagation in gas-filled guides and in free space. It is a memory response of the fused time-domain column; its cost is set by the recollision integral, and a validation needs an independent single-atom reference such as a time-dependent Schrödinger solution. SEP 12 delivers a bounded TDSE integration for helium capillaries, with illustrative 120–200 nm reporting and explicitly unqualified physical-window and calibration errors; it does not implement the SFA response or complete high-harmonic propagation described here.

Gas charge-state ladders and avalanche ionisation. Each ionising species is one channel from the neutral (mixtures, ADR-027). Successive ionisation through coupled charge-state populations, and collisional (avalanche) ionisation, are not modelled; both matter at high intensity and pressure. Multiple ionisation and dissociation fit the kernel contract as further channels and further values of electrons_per_event, without a new seam. The solid avalanche closure proposed in SEP 5 does not cover gas avalanche or coupled gas charge states.

Ionisation in solids. SEP 5 covers scalar bulk free-space propagation with solid photoionisation, simple avalanche, carrier removal and collisional optical currents. Solid-core fibres and regional general modal guides remain open, as do delayed-heating avalanche (multiple/dynamical rate equations), genuinely subcycle solid rates, tensor/anisotropic carrier response, inter-pulse memory and GPU qualification. YAG Raman requires its own qualified data when needed. The proposed bulk model does not supply those extensions or material damage dynamics.

Phase-resolved ionisation rates. The Yudin–Ivanov rate is available for research comparisons of the envelope plasma model, not in production propagation. Adopting it needs a separate loss-current closure and a resolution of the rate/current questions set out on envelope plasma.

Raman population dynamics and data. The Raman responses are fixed kernels (Raman). This item is partly covered by SEP 32, which proposes scalar two-level population dynamics within one temporal window. SEP 3 delivered optmat’s per-line responses with density-dependent widths; SEP 1 extends exact evaluation to shifted lines along density gradients and SEP 10 adds optional compression. Shared-level rotational and vibrational ladders and accumulation between separate pulse windows remain open. SEP 3 retains the missing nitrous-oxide rotations and provisional vibrational data as limitations. It also refuses Raman in mixtures whose collision partners optmat has not curated, such as H2, D2, CH4, N2 or O2 in a noble-gas buffer other than argon. Lifting that refusal needs foreign-gas broadening coefficients curated upstream in optmat, which is an optmat roadmap item. Q-branch line mixing at high density, modern H2/D2 level energies and the H2/D2 gain excess are optmat’s open items too. No proposal supplies the glass overlap of a hollow-core fibre, so wall Raman remains open on the gas-capillary route; the general modal route’s material regions are the building blocks.

Broadband Raman coupling. Raman-shift dependence and carrier-selected strengths already exist; dependence on the participating optical frequencies across a broadband field does not. UV generation motivates a separate material/response proposal. An off-resonant separable coupling could retain FFT or recurrent oscillator histories, but its drive and emitted polarisation must be derived together, with reality, reciprocity and energy transfer checked. Curated frequency coverage and weak-line accuracy are open. This work is independent of Maxwell–Bloch populations and does not block SEPs 1 or 10. SEP 6’s electronic filters do not supply Raman dispersion; see SEP 1 §10.

Gain with saturation. A linear operator may carry prescribed gain; gain saturation, population inversion and an amplifier or laser population model do not yet exist. SEP 7 proposes Yb/silica amplifier dynamics with steady and transient rate equations, pump/ASE transport and scalar population resolution. Qualification of its broadband filtered-field closure is an explicit decision, not an assumed capability. General vector and moving-guide gain, additional dopants and their host-specific kinetics, stochastic field ASE, oscillator dynamics, thermal feedback and photodarkening remain outside that proposal’s implementation scope.

Resonant gas susceptibilities. The gas models are off-resonance fits. Near ultraviolet and vacuum-ultraviolet resonances a user-supplied dispersion does not make the nonlinear or ionisation response valid there; a resonant susceptibility model would. SEP 3 partly addresses the material data through curated dispersion and available absorption, but does not supply a complete resonant nonlinear or ionisation model. SEP 6 adds off-resonant electronic dispersion and identifies UV/VUV validity gaps; it does not supply the missing resonant response or its nonlinear absorption. Beyond a gas’s band edge its evaluators hold a constant labelled best effort, and the helium dispersive wave’s energy moves by 30 to 55 % between that and the carrier value, depending on the mode set (the worked example): a resonant electronic response of helium in the vacuum ultraviolet is the missing reference. SEP 12 implements a helium TDSE response (TDSE response); its ultraviolet validity and its behaviour near resonances remain open, and it does not supply a general resonant material model. SEP 17 proposes an Ar/He continuum optical response with its dispersive partner, subject to completion and background matching. It does not supply a validated general resonant susceptibility, bound-state dynamics or (2+1) ionisation; those material inputs and propagation models remain open. optmat OEP 8’s bound-state material work is separate from the continuum approximation.

Noise transport along a gradient. The modified noise is advanced by the unitary part of the linear flow (noise, ADR-008). With true-power amplitudes (ADR-030) that leaves the noise floor drifting along a pressure gradient, by at most 0.3 % of the noise amplitude over a 20 bar gradient. Transporting the draw with the power weight would remove the approximation. SEP 16 qualified modified-noise stepping while preserving the existing model; it did not remove this physical normalisation limitation.

Waveguides and geometries#

Material absorption in step-index fibres. The step-index eigenproblem takes the core and cladding indices as real, so absorption in the glass does not enter the modes’ loss. A confinement-factor correction, each region’s absorption weighted by the share of the mode’s power it carries, would add it without a complex eigensolve.

Integrated waveguides. Support chip-scale nonlinear optics in silicon nitride (SiN), silicon (Si) and lithium niobate waveguides, with geometry-dependent dispersion, loss and vector mode profiles. The general modal route provides a starting point; cross-section providers or imported mode-solver data, material-specific nonlinear responses and validated examples remain to be developed. The scope should account for second-order nonlinearity and crystal orientation, nonlinear absorption and carrier dynamics where relevant, alongside Kerr and Raman responses. SEP 8 proposes bulk crystal tensors and orientation handling; guided quadratic overlaps and integrated-waveguide providers remain outside its implementation scope.

Multi-pass cells (MPCs). Model pulse compression in gas-filled and solid-state multi-pass cells by composing repeated bulk propagation with focusing mirrors, dispersive optical elements and the chosen cell geometry. A cell-level workflow would need to carry the spatial and temporal field from pass to pass, account for mirror dispersion and loss, and expose the evolution of the beam, spectrum and compressible pulse duration. SEP 31 added field handoffs, focusing, reflection and spectral optics on top of the existing free-space routes (coupling simulations). Cell geometry, assembly and validation against independent propagation or experimental data remain open.

Coupling beyond the forward model. SEP 31 delivered forward overlaps, scalar homogeneous gaps and spectral compressors. A general junction calculation with reflected, radiated and evanescent fields, high-NA vector free-space transport, gradient transport outside SEP 34’s certificate, varying absorption, resolved prism/grating angular dispersion and spatial chirp, and feedback between stages remain separate extensions. SEP 34 covers the existing radial route’s certified forward amplitude correction, including admitted weak steps; it does not add general gradient scattering or a complex-index path certificate.

A graded-index facet envelope for the vector coupling model. vector_paraxial into a step-index guide counts a recorded envelope for the facet reflection and backward fields it omits (VECTOR_STEP_INDEX_ENVELOPE, against an independent full-vector mode-matching solution of the facet). A graded-index guide has no such envelope: it is admitted with an uncounted envelope item, its facet omission bounded only by the weak-guidance contrast screen and the H closure. A graded-index reference — the same mode-matching solver with a layered core — would give it a recorded bound (SEP 31, decision A1; coupling simulations).

The coupling envelopes beyond their measured range. Each recorded pair envelope holds only where it was measured (SEP 31, Part II, “Closing”):

the capillary envelope stops at the leaky parameter \(\delta = 10^{-2}\), so a narrow capillary in the infrared is refused; extending it needs a re-measured bound, and is worth doing only when such a pair is needed;
the beam-into-fibre envelope covers LP01 alone (V 1.5–2.4, contrast up to 0.01); near V 2.4 its in-cone figure is dominated by the cone’s truncation of the widening HE11 spectrum rather than by the vector correction, and could be sharpened;
the vector facet envelope rests on the analytic Fresnel ceiling in the narrow bands just above the HE12 and HE13 cutoffs, which no practical reference cylinder resolves;
the fibre-pair STEP_INDEX_ENVELOPE is a constant bound whose order in the index contrast was not measured; re-measuring it would let it scale as the others do;
every envelope is measured at the profile wavelength against the run’s frozen profile, so a fibre whose profile moves strongly across the band is represented only as the run represents it.

The general guide’s angular item at every occupied bin. couple into a general modal guide evaluates the target’s angular admission at no more than five bins of the band. Evaluating it at every occupied bin closes the gap, if its cost stays within a coupling’s; the cost is to be measured first (SEP 31, phase 3).

Other exit interfaces for Outcouple. Outcouple synthesises a guide’s exit field under the matched interface only. A declared planar-interface model at the exit (the end face of a fibre into air, for example) would make the facet’s reflection and refraction explicit, as PlanarInterface does between two media in transport (SEP 31, phase 3).

Vector coupling into a moving general guide. The input port of a general guide whose geometry moves along z (a sampled path) carries no E/H reference at the entrance, so it admits only same_basis; vector_paraxial needs the entrance section’s fields from the path (SEP 31, phase 3).

Elliptical hollow capillary modes. Extend hollow-capillary modelling to elliptical bores specified by their two semi-axes, with mode profiles, dispersion, wall loss and nonlinear overlaps that resolve the geometry and polarisation. This needs a mode model and transverse quadrature validated against an independent reference, including recovery of the circular-capillary limit as the semi-axes become equal. The choice between an analytic approximation and numerically supplied modes remains open.

Zeisberger antiresonant modes. The analytic antiresonant model of Zeisberger and Schmidt is not implemented; the antiresonant family is Vincetti’s tube-lattice model (modes). A second model gives an independent check on the first and covers geometries described by a single effective wall.

A coherent-wall (Robin) slab. The planar guides truncate the field at the plates, with one wall-loss pair per channel (spatial decompositions). The coherent wall that couples capillary modes through the condensed outgoing Robin condition has no planar counterpart; with one, slab modes would exchange power through the wall as HE_1m modes do.

Planar guides resolved in y, beyond one family. The hybrid planar route (slab modes in x, an FFT in y) carries one polarisation family with a constant gap and constant media. Both families together, a gap or medium varying along z, and a settled normalisation for a noise draw on the planar grid (refused at translation today) are open.

Single-azimuthal-order radial free space. A full radial and azimuthal expansion is not built: it needs one Hankel transform per order, on radial nodes the orders do not share, and beams that break cylindrical symmetry belong on the Cartesian grid. A single order-\(m\) transform is a small change to the order-zero quasi-discrete Hankel transform. It would be exact for cylindrical-vector (radially or azimuthally polarised) beams under isotropic responses; for a scalar vortex, exact only on the envelope grid without third-harmonic generation, and blind to azimuthal break-up.

Cartesian media varying along z. Cartesian free space requires a medium constant along z: its seam stores frequency-only material data and evaluates \(\Phi\) as \(D\,\Delta z\). Varying profiles and tabulation are refused by name (Cartesian). The radial grid already supports varying media through cumulative expansions. SEP 8 proposes homogeneous crystal slabs and resolved longitudinal nonlinear domain reversal; it does not cover continuously varying dielectric tensors, crystal orientation or temperature.

Taper coupling for tube-lattice fibres. A multimode capillary whose radius varies along z carries the taper coupling between its HE_1m modes by default. A Vincetti tube-lattice taper keeps the adiabatic route, although its fields are Marcatili’s and the connection would be the same; admitting it is a small change with its tests.

The envelope grid on the general modal route and the coupled flows. The general modal route, the coherent wall and the taper coupling run on the field grid; an envelope grid is refused by name (general modal). The coupled flows need an increasing frequency axis, and the envelope grid is in FFT order. prop_gnlse covers the scalar envelope case.

Hybrid graded-index sectors. Graded-index guides provide the LP sectors under weak guidance (index contrast up to 0.02) and the radial TE_0m/TM_0m sectors. Hybrid HE/EH sectors of azimuthal order one or more are refused; they are needed for strongly guiding profiles and for polarisation effects in graded-index fibres.

Numerics#

Multistep integration with general varying noncommuting coupled flows. SEP 16 covered diagonal and separable operators, adiabatic tapers and the constant-Schur and frozen-residual coupled sections, which already carry a constant block with an explicit residual, for its exponential formulation. CF4 sections and other general varying coupled generators without an explicit residual need a separate propagator/residual derivation, action-error control, history/dense-output construction and stability analysis. Exponentiating the integrated matrix is not generally its exact flow. The Runge–Kutta march keeps them; the Adams methods refuse such a problem by name rather than omit its coupling.

A coupled flow with no source under the Adams methods. Beside a frozen-residual coupled flow the explicit residual joins the compensation, which is sampled with the source (ADR-042, ADR-044). A run with no nonlinear source has no sample, so its steps are the frozen block’s alone and their error, second order in the step, is not seen by the controller: a linear graded-index taper at one bin is 6.4e-4 from the converged answer at ten steps whatever rtol (the Adams methods). Sampling the residual without a source, as a cheap evaluation of the linear part alone, would put it under control.

The exponential formulation beside the modified noise. The noise’s source is partly linear in each weak bin’s own field and rotates with its free phase, which the exponential formulation interpolates in the laboratory frame. Until a developed noisy window, such as the CtC comb, measures it, the manual prefers Multistep(formulation="interaction") beside that noise.

Qualifying the TDSE response. The helium TDSE response (SEP 12, SEP 20, SEP 26) is a research option, demonstrated over short capillary segments (accuracy and limitations). Open: a converged physical time window and atomic preparation, against which the default 60 fs window’s history sensitivity becomes a stated error rather than a recorded difference; a calibration of the single-active-electron current scale and of the ionisation yields that is independent of the linear, cubic and rate matching it rests on; numerical qualification over longer segments and other waveforms, then a full-length capillary demonstration with a worked example; and publication of the atomic provider, Pyroraptor (dependencies), so that the response can be installed from a released package. Other gases, double ionisation and solver comparisons on the actual atomic source lie beyond these.

A quadrature-error estimate on Cartesian waveguide runs. A radial run scores its transverse quadrature against an embedded coarse rule and reports the difference as the quadrature_error statistic; a Cartesian waveguide run has no embedded rule, so the statistic is unavailable there.

Light at a waveguide’s exact cutoff in free space. Grazing incidence on a spatial grid is handled arithmetically (the residuals are exact and a warning names significant content near cutoff), but there is no physical treatment of content sitting at the exact cutoff, where the longitudinal wavenumber vanishes; a run with significant power there is outside the model.

Open numerical questions in the stepper. Error norms for observables far below the pump, reject-free stepping at matched global error, schedule transfer beyond the measured ranges (including from a coarser transverse grid to the production one), round-off growth with varying operators and FSAL through a coupled flow are measurable questions rather than planned work; they are listed in open questions.

An embedded low-storage pair for adaptive runs. SEP 24’s march is replay-only; an optional extension would add an embedded pair with adaptive control on a fifth or sixth volume, qualified against the adaptive Runge–Kutta march as the Adams methods were. The SEP recommends starting from Balac and Mahé’s ERK4(3)-IP (doi:10.1016/j.cpc.2012.12.020), which keeps the march’s stencil, kernels and checks, once its coefficients, stages and FSAL structure are confirmed from the paper; a 2R or 3S* scheme only if one register decides, whose inner nodes need inverse flows and fresh qualification. No memory-bound adaptive run needs it yet, and the march with an external calibration (calibrate, replay) covers the fixed-step use.

An absolute tolerance on density paths. The adaptive Chebyshev builder accepts an interval by a tolerance relative to each element, with ADR-040’s sampling-noise floor where floating-point z is too coarse. Near a choked gas end the density, and with it the nonlinear response, is a fraction of a percent of its peak, so an absolute tolerance against the path’s peak would stop refinement there earlier. It would complement the floor, not replace it; worth adding if translation near near-vacuum ends shows up in a profile (SEP 14).

A narrow-support fallback for general guides. A general modal guide must be sampled over the run’s whole active band, which reaches to half the smallest declared frequency and is wider than the declared band. A guide that does not span it is refused, and the user narrows the grid. A fallback would propagate only on the guide’s support and warn when energy reaches a guard band inside its edge; it needs an optional band threaded through the shared linear-operator construction that every route runs, and a definition of the warning threshold.

Observable tolerances on the general modal route. observable_rtol and observable_atol accept only None, meaning the controller’s own tolerances. A number would need an error allocation over the table’s frequency interpolation, the transverse rule and the step, and named physical observables to allocate against; until then a number is refused rather than recorded and ignored.

Noise and batches beside a coupled flow. A noise model beside a coupled flow (a wall or the taper coupling) is refused, because a draw’s transport through the flow is not measured; so is a coupled flow in a lockstep batch. Both need a derivation and a test before they can be admitted.

Independent spatial references for the vector and graded-index routes. The scalar bore and taper are checked against an independent finite-element solve. The vector wall is qualified by the exact single-family Maxwell roots and the mixed-coupling eigenvalue evidence; a resolved-cladding vector solve in a mixed state would be the stronger check. Graded-index and vector spatial references, and a demonstration of a mode crossing its cutoff along a taper, are also open.

Performance and backends#

The CPU performance record lists what has been tried, adopted and rejected on the compiled CPU backend, with its measurements; read it before taking up any of the first ten items.

BLAS threads in the Windows and manylinux recipes: one OpenMP runtime for Kokkos, FFTW and OpenBLAS in the Windows wheel (set up in the wheels workflow, to be proved there and timed), the same check on the manylinux release recipe, and the Kokkos team-barrier backoff (sleeps that Windows rounds up to its scheduler tick) reported upstream.
The problem-size thread policy: the rows that measurement noise left unresolved are within about 1 % of the pre-policy build on the 36-core Xeon (2026-10-22). Open items:
- the scalar wall, 10 % slower there at the default width, against an A/A spread under 1 % (on Apple silicon it is faster); find the cause;
- the taper, short-line and planar controls;
- the stage and FFT starts on macOS, kept at 8,192 and 32,768 while Linux uses 4,096 and 16,384 (smooth grid sizes): on Apple silicon the serial stage is faster even at 8,196 elements, so sweep the crossover there on a quiet machine (native/benchmarks/size_cliffs.py and the SSFS and reduced comb-to-continuum cases) and set its own values.
The multimode step at full width on Linux: at 36 threads on the Xeon, general-modal RK is about 6 % slower than before the per-core changes (reduced comb-to-continuum is 4 % faster). Estimate planning is as slow, so FFT plan selection is ruled out; find the cause.
The TE/TM wall’s complex GEMM, slower in Kokkos Kernels and OpenBLAS than in a vendor BLAS for its shapes: improve the kernel without letting this rare configuration decide the wheels’ BLAS.
Windows re-measurement of the CPU fixes against numba_cpu on the single-runtime wheel, and a Windows CI lane that fits its memory (fewer xdist workers or a larger page file). Two numba_cpu cache tests exceed Windows’s 260-character path limit.
FFTW planning policy per transform shape: estimate, measure and saved wisdom, and wider codelets, compared per representative shape and processor and qualified on the minimum CPU model.
Compiler targets with identical libraries (x86-64-v2, v3, v4 and processor-specific targets), which matter most on the general modal route.
Explicit SIMD in at most one or two kernels (the general-modal response, synthesis and projection), after the compiler-target comparison.
Selective run-time kernel dispatch: AVX-512 variants of a hot kernel in the x86-64-v3 wheel behind a CPU-feature selector (SEP 29, §6.2).
Batched vector transcendentals for the plasma response, which need the dependent plasma recurrence split into a batched rate pass.

Further fusion in the stage passes. Candidates, each subject to the source-accumulation and stage-dependency checks of SEP 25 §6.7: stage.phase into the combination; the finish of one stage with the combination of the next; the zero-fill copies into the kernels that fill the rest of the padded buffer; and a 1×1 map’s transform scale, synthesis and projection into the response column (architecture.md §8).

Floating-point options and reduced precision. What the floating-point policy (SEP 25 §5) leaves untried, with the expected gains. Build options promise little: the transforms and the BLAS are in libraries these flags do not reach, and the policy already takes contraction and per-kernel reassociation. Only reduced precision promises a large factor, and mainly on GPUs with slow double precision.

Decided: contraction on, with the protected blocks strict; reassociation per kernel; no umbrella fast-math, nnan/ninf, nsz, or approximate functions and reciprocals. Measured and not adopted: libmvec exp and log under Clang (about 3 %, still behind GCC) and Accelerate’s vForce in the propagator (about 0.1 ms per window).
Value-safe options, not yet measured: -fno-math-errno and -fno-trapping-math change no rounded result. Under GCC (the Linux builds) the first lets sqrt compile to one instruction and vectorise; the second lets loops with conditional floating-point work be if-converted. A few per cent at most, in the custom kernels only.
Subnormals: no build sets flush-to-zero. On x86-64 an operation on a subnormal can cost of the order of a hundred cycles (NVIDIA GPUs and Apple silicon take them at full speed), and spectral wings held down by an absorber or a per-step exp(-rate dz) could reach that range on long runs. Count the subnormals in a long production run’s field first; flush-to-zero needs its own justification (SEP 25) only if they appear.
Double-precision transcendentals on GPUs with a reduced FP64 rate (NVIDIA’s GA102 and AD102 parts, such as the A40, run FP64 at 1/64 of FP32): the stage combination evaluates exp, cos and sin per bin and stage, which there may cost as much as the transforms. Profile first. Remedies that stay in double include deriving one stage’s propagator from another’s (squaring a half-step propagator) and reusing propagators while dz is unchanged.
A single-precision source evaluation inside a double-precision solver: the transforms, the responses and the modal GEMMs in FP32; the state, the linear phase increments, the Adams history and the running sums in FP64. Its rounding enters each step scaled by dz, about FP32’s epsilon times the step’s nonlinear phase, so the error estimate’s floor stays near \(10^{-8}\), against about \(10^{-5}\) for a solver entirely in FP32. Expected gain, unmeasured: up to the bandwidth factor of two on CPUs and data-centre GPUs, possibly several times on reduced-FP64 GPUs. A first test needs no new data types: round only the source evaluation to complex64 in numpy_ref, on coherent cases and one noise-seeded case. Noise-seeded runs may fail it, since FP32 rounding can sit near a one-photon-per-mode seed for pulses of a microjoule and above. Adoption would be a SEP.
Entirely single precision, and Metal: SEP 33 (final) built the precision setting, a numpy_ref oracle and an MLX backend for Apple GPUs on the fibre, radial and Cartesian routes with every response but the refused ones. Its open questions: whether a rule comparing binary32 Adams with binary64 Adams should admit Multistep() under single precision (binary64 Adams has the same errors at the tolerances the study used); mlx cases in the standard benchmark suite (SEP 30); and a fused stage kernel for the launch-bound small plasma capillaries. Open beyond it: single precision on native and numba_cpu (Kokkos, Kokkos Kernels and Kokkos-FFT support it on every backend in use), on the routes SEP 33 refuses, and noise models under single precision, which SEP 33 refuses as a known limitation: binary32 rounding can reach a one-photon-per-mode seed at high pulse energy. Kokkos has no Metal backend, and one would need a single-source C++ compiler for Metal (the runtime part is modest: 5 to 10 thousand lines for each existing GPU backend).

Benchmark platform qualification beyond the initial CPU suite. SEP 30 covers the standard suite, portable records and durable incremental exports, with CPU timing collections on macOS and Linux and an initial native coverage collection on macOS. Linux native coverage, Windows execution, cloud launch/collection recipes and GPU timing and coverage need separate qualification. Keep future platform results in distinct machine/build records; successful CPU coverage does not establish coverage of device code.

Remaining benchmark references and estimates. Complete independent reference checks for the 512-node radial filament, 128/256-square Cartesian filaments and 768-square Kerr beam when those workloads are needed for validated comparisons. Keep their incomplete qualification visible while permitting exploratory timing. Add runtime forecasts from compatible pilot receipts and automatic consumption of compatible measured memory receipts. Until then, planning reports unknown costs and admission uses conservative provisional budgets.

The low-storage march’s remaining volumes and names. Follow-ups of SEP 24:

The planar route’s stage register. Its finish could write the next pass’s input into the column volume, as the Cartesian finish writes it into the transverse scratch, saving one volume when no in-stage collar is present.
A varying modal prefactor on numba_cpu. On a modal route whose response prefactor varies along z, each pass combines the prefactor’s terms into a fresh state-sized array (both marches, NumbaCpuBackend._prefactor_for); a held buffer would take it into the ledger, as the march’s recomputed flows are.
Hook names. Rename _adams_lab_source and _adams_propagator, which the low-storage engines also call, with the next change to the kernel contract.
Device qualification. The march’s passes pass their device tests on a HIP build; the CUDA build is to be qualified on an NVIDIA GPU, and device timings against the IP-DOPRI5 replay are to be measured.

An opt-out for the held field-free plane. Under the low-storage march a field-free plane holds one extra field volume so that a run stopped in flight can resume from it (SEP 22). That volume is the difference between the march’s five volumes and a four-volume budget. Offer a setting that drops it for runs that never resume, such as many short runs in a scan, with the abandoned step then recorded without a checkpoint. The default stays as it is.

The dispersive polynomial on mixtures of distinct leg bounds. SEP 6’s polynomial evaluator shares its derivative fields only between species with equal leg bounds, so air (N2, O2 and Ar, three bounds) costs 4.3 to 7.6 times the constant cubic’s stage on numba_cpu, above the 4 times of its target. Reduce the per-species transforms without approximating the operator, or measure a bounded-support grouping first.

The on-axis statistics on device builds. The native backend computes SEP 19’s on_axis_* statistics from a host copy of the state, one per saved plane that requests them. A device scan of the axis row, as the SEP’s §6 describes, is not built.

The device gallery under the active-band norm. The stepper gallery’s AMD CDNA 3 and NVIDIA Hopper rows (choosing a stepper) were not measured under the error norm that scores the whole active band, which the CPU rows use. Regenerate the device rows under it and redraw the gallery.

A device-sized node chunk. The source’s node chunk is sized by SimulationSettings.node_budget (1 GiB by default on spatial grids, 64 MiB on modal routes) at translation, whatever the backend. On a device a chunk is a kernel launch: the 128 × 128 Cartesian filament evaluates its plasma in 161 chunks a call, and with every column in one chunk IP-DOPRI5 ran 4 times faster and the exponential Adams method 2.4 times. A chunk of any size gives the same fields to rounding, so the device backend can choose its own from the device’s memory when it is built, without translation importing the device runtime, which a scan must not initialise before it forks its workers.

Plasma statistics between planes from the peak-field node. A per-step statistics row’s plasma statistics scan every quadrature node’s time column. The Runge–Kutta march reads them off its last stage, which is the accepted state; an Adams step’s one evaluation is at its predicted state, so under the default a row scans afresh and, at a period of one, adds 80 % to the reduced M2 run (statistics). electron_density and peak_ionisation_rate describe one node, the one of largest ionised fraction: synthesising every node (a cheap matrix product) and scanning only the node of largest peak field, or the two or three largest, would cut a row’s scan by the node count (8 to 24 on a capillary, about 256 on a radial free-space grid). The peak need not be on the axis (a TE₀₁ or ring-shaped field), and the node of largest field is not always the node of largest fraction where pulse shapes differ across the beam, so the rule is to be checked against the full scan on the qualification cases and applied to plane statistics too, keeping a row the plane’s own to rounding. Statistics resolved over every node (species_ionisation) keep the full scan.

The split-step collar on the general modal route, native against the reference. With absorber_scheme="split_step" on a general modal guide, numpy_ref carries a launch’s content outside the active band (it decays over the steps) and native sets it to zero at once; inside the active band the two agree, and with the in-stage collar or none they agree everywhere. Not yet traced: one of the two applies the hard band limit where the other does not. SEP 21’s campaign needs it resolved before it compares the split scheme on a general modal guide.

Fixed-order Adams norms on devices. On a native host build the Adams engine’s magnitude and defect norms sum fixed chunks in index order and combine the partials serially, so an adaptive run is the same at every thread count, as the backends owe (backends). A device build still reduces them with Kokkos’s own parallel_reduce, whose association is the launch’s; the fixed-order device form (a fixed block count, each block reducing its own contiguous range) is to be built and measured.

Spatial convergence records on CUDA. The spatial routes’ convergence records (spatial_accuracy, and backend_equivalence for the polarised and free-space routes) have no native_cuda rows, so the native-parametrised tests that read them cannot pass on a CUDA build until they are recorded there.

Susceptibility sampling on tabulated and rectangular modes. Along a pressure gradient, these modes evaluate the gas susceptibility, and its dilute-scaling check, at every z sample instead of once per frequency array as the capillary modes do. It costs time, not accuracy; worth sharing the capillary’s sampling if a gradient run on these modes shows the cost.

Execution autotuning beyond existing CPU controls. SEP 4 covers explicit search over existing CPU pool, FFT-planning and OpenMP policies, including plan application and validation. Runtime-selectable Kokkos GEMM providers and GPU team/reduction geometries remain open: they need an enumerated, qualified native capability set before becoming candidates. SEP 25 profiled data-centre NVIDIA and AMD GPUs and tiled the device band column; runtime enumeration and automatic policy selection remain with this informal item, as do the coupled cases’ launch and copy counts that SEP 25 records. Chunk-size tuning needs route-specific memory admission; the general modal route does not use the ordinary transform’s node_chunk. Propagator-policy tuning must account for the requested mode in today’s resume hash and route-specific refusals. Implicit tuning on the convenience interfaces, broad reuse across scan points and native batch tuning also remain outside SEP 4; none may infer full-run savings from a short prefix.

CF4 on the compiled backends. The commutator-free fourth-order flow, the certified alternative to the frozen split for a varying coupled flow, runs on numpy_ref alone. The frozen split is the default and the more accurate on every case measured, so CF4 on numba_cpu and native is needed only if a use case favours it; its Gauss-node weights are evaluated inside the backend, not carried in the stage bundle.

The z table on GPU builds. Every CPU backend runs the general z table, the native CPU build in its own kernels (SEP 29 section 6.6), but the native GPU builds refuse it, naming its size: the table is the state’s size times the degree times the intervals, held resident. Two routes would admit it there. A memory check against available_device_memory before allocating would admit the tables that fit, at the cost of device qualification of the table kernel. For a SpecifiedMode whose dispersion is declared at several positions along the guide, a compiled representation would declare the interpolation coordinate a parameter and expand the dispersion in it, as the operator expansion already does for a fill’s density, so the GPU builds could run it without the table.

A compiled closure synthesis for numba_cpu. The local spatial statistics (fluence, intensities, the on-axis spectrum) need the closure G formed with the complex admittance; numba_cpu forms that volume with NumPy FFTs on the host, where native forms it in its own kernels (SEP 34, phase 2, Q2). A compiled pass would remove the host round trip.

The cost of a kept native save on Linux. On the snapshot benchmark case a kept native Adams save costs about 4.5 ms against numba_cpu’s 1.6 ms on Linux, unexplained; the run totals are level (SEP 29, phase 2).

The hybrid planar route on GPUs. native CPU executes the planar kind resolved in y; the CUDA and HIP builds have run it only in one short Kerr case of SEP 34’s device qualification. Its full device qualification follows the other routes’ parity bounds (backends).

The far-field pass on GPUs. The CUDA and HIP builds compile the native backend’s far-field pass (SEP 23) but have run it only on the short Cartesian and line cases of SEP 34’s device qualification: its band partials live in global memory, with a team barrier per column. Device qualification measures it against the CPU parity bounds of tests/backends/test_native_far_field.py, and its optimisation holds a thread’s slot accumulators in registers and runs a large aperture set in groups (SEP 23 §6).

Multi-channel columns on GPUs. On the CUDA and HIP builds a column with several ionisation or population channels (a gas mixture, for example) takes a serial loop over its nodes; only the single-channel column uses the team-parallel path. Device qualification covers the multi-channel results, not their speed; SEP 25’s device optimisation included these columns, and a channel-specific redesign remains open.

Device qualification of optmat material runs. No run with optmat materials, Raman responses or ionisation rates has been qualified on a CUDA or HIP build. The kernels see only material-agnostic arrays, so no change is expected; the next device qualification is to include one.

GPU continuous integration. The gate runs on CPUs. The native CUDA and HIP builds are qualified by hand on data-centre NVIDIA and AMD Instinct GPUs; a scheduled device lane would catch regressions in the gpu-marked tests between qualifications.

Native source builds without a checkout. The published wheels are built from the repository, but solprop-native cannot be built from a source archive or offline: its CMake configure records its own commit through git rev-parse and fails without one, and by default it fetches Kokkos, Kokkos Kernels and (for the Linux distribution build) OpenBLAS from GitHub (dependencies). A release archive that builds needs the commit to come from the package version and the sources to be vendored or fetchable by checksum.

Wheel qualification beyond CI (SEP 29, §7). The wheels’ test jobs run on GitHub’s runners, which carry compilers and current operating systems. Still to do: installs on clean machines at the declared minimum systems (macOS 14, the manylinux_2_28 generation, Windows 10); the slow lane on the native backend from the wheels, not only from a source build; and a lane that fails on unexpected native skips.

Per-platform performance of the wheels (SEP 29, §7). The standard suite’s core cases, adaptive Multistep and Runge–Kutta, native against numba_cpu with matched budgets, at one thread and the production width, on each wheel family, recorded in the benchmark ledger with any serious loss per case. To run once the first release is out.

macOS: pyFFTW imported before the native package. pyFFTW’s macOS wheel carries and initialises its own libomp, and the native package refuses to import after it (the rule is documented in Building the native backend). Lifting the rule needs a macOS native build without OpenMP (Kokkos Threads and threaded FFTW, measured first) or a pyFFTW wheel that does not initialise its own runtime.

The future of numba_cpu. With the native wheels published, a later SEP decides whether numba_cpu stays, on evidence from use: how often auto falls back, coverage across the installed platforms, default performance and coexistence reports (SEP 29, §6.3).

Raman convolution length. The Raman convolution is always built on a doubled grid of 2 n_to samples, including for short-lived responses such as silica’s. SEP 1 partly addresses this cost with exact sampled recurrences for supported kernels and the existing lag-taper correction. A convolution sized to the response’s useful support remains outside its implementation phases; it must stay causal, bound its truncation error and be benchmarked on short runs.

Batch fault isolation. In a lockstep scan batch, a fault at group level (a non-finite error norm, for example) fails every member of the group, so one diverging point can cost up to the group size less one points that Local or Pool would complete. Re-dispatching the survivors through the per-point path is the preferred fix; either it or a per-member abort changes the documented failure semantics of Batch (scans).

Runner reuse and field-free members in scans. A scan worker constructs one backend per point rather than keeping one SimulationRunner for its points, so a large spatial scan pays a backend’s allocation and transform planning per point; and a lockstep batch (observe_members) observes every member’s field at every plane, since Batch requires fibre geometry. Both were left out of SEP 22: the first needs the worker to check each point’s problem against its runner’s, the second a field-free member observation.

The compiled backends’ extra volumes in the memory estimate. estimate_memory counts the reference’s base buffers on every backend: on the Cartesian kind the Runge–Kutta march holds 13 state volumes on numpy_ref, 16 on numba_cpu and 15 on native CPU, and the difference, gigabytes on a large grid, is declared not counted (SEP 22 §6.1). Count them per backend so that admission is exact there too.

Cold-start compilation. numba_cpu’s parallel kernels are compiled in every new process rather than read from numba’s disk cache, because parallel closures compiled in different processes can collide in the cache. A collision-free cache would remove the first-run compile cost from every worker of a scan.

OpenMP core places and the Numba allocation cap. With OMP_PLACES=cores on a machine with two hardware threads per core, OpenMP can bind the main thread to two logical CPUs. solprop._process.allowed_cpus recognises a one-CPU runtime binding but treats this two-CPU mask as a deliberate allocation restriction, capping Numba, FFTW and BLAS at two despite a larger request. The SEP 27 campaign exposed this in execution metadata. Its corrected measurements use matching explicit places containing one logical CPU per physical core and assert actual pool sizes. A general repair must distinguish an OpenMP place from a deliberately narrowed scheduler grant and retain the existing affinity tests.

Held-current source cost and compiled attribution. The SEP 27 campaign measures the cost of replacing the gated equation on existing routes. On its 18-thread radial case, Numba’s source costs 1.67–1.72 times the gated source and its ten-step window 1.53–1.67 times the gated window, while the native backend’s corresponding ratios are 0.58–0.59 and 0.46–0.47. Other modal and spatial cases also have resolved increases; the per-case measurements and noise ranges belong to the benchmark ledger. The NumPy profiles identify extra transforms and chunk passes; compiled explanations still need direct profiling. Investigate the current products, projections, transforms, copies and chunk counts without changing the production equation or dropping accounted storage. The radial and modal routes are outside SEP 28’s free-space evaluation and need their own work.

The general route’s memory estimate. A general modal guide is sized before its route is built and admitted against a declared budget. The estimate does not count the transient knot table the reference is interpolated from, nor the base, delayed and flow groups, and a large guide is warned on twice (at construction and at run).

Large arrays in the run record. A spatial run’s metadata (metadata_for in solprop/driver/simulation.py) carries the geometry’s spatial_weights, spectral_weights, window and absorber_rate, four numbers per transverse node, and a mixture run’s processing record (processing_record in solprop/driver/material_output.py) carries the dense node synthesis matrix (\(n_r^2\) numbers on a radial grid) and the species absorption cross sections. All of them are JSON text inside the configuration that the resume hash covers: at 512 × 512 the four geometry vectors alone are about 20 MB, serialised and hashed on every resume check. The geometry vectors follow from the grid’s parameters, and the matrices could be stored as datasets outside the hashed text; either keeps the record small.

The radial statistics synthesis on the native backend. The radial spatial statistics (peak_intensity, the fluence widths) synthesise the intensity at every node and time sample by a direct \(O(n_r^2 n_t)\) sum rather than a matrix product. These are per-stage and per-plane times, not allocations.

A copy-free Adams sample. The Adams sample on numba_cpu makes 16 passes of which 4 are needed (identity fills, a stage clear, an unused accumulation, a returned copy). A numba_cpu Adams engine through the existing hook, or a destination argument for _adams_lab_source (a kernel-contract change), would remove them; either needs lifetime and rejection tests across startup, restarts, corrected-field evaluations and dense observation, for both formulations.

The march’s register traffic on devices. On a data-centre NVIDIA GPU the low-storage march takes 51.6 and 227 ms a step at \(384^2\) and \(768^2\) by 600 against 30.8 and 168.7 ms for a dedicated single-transform engine. A transform probe showed the transforms near their best (one rank-3 transform would save 4-12 %); the rest is the fold, already near the card’s bandwidth, and the volumes each pass reads and writes. A march that touches fewer volumes per pass is the remaining lever.

Population statistics without a current stash. On a field-grid run with plasma channels, each requested population statistic (the species_* statistics, electron_density, peak_ionisation_rate) rescans the populations at every node: on numpy_ref at every plane, and on numba_cpu when the accepted step’s stash is not current, as at the entrance. Seven requested names cost seven scans where one would do; the envelope-plasma path already shares one scan between them.

Tools and analysis#

Full dispersive-wave phase matching. The design tools phase-match resonant dispersive waves with the linear mismatch and an optional Kerr term (design tools). A plasma term, a Raman term and the full soliton wavenumber would predict emission where ionisation or the Raman shift moves the soliton.

Nonlinear diagnostics beyond the proposed statistics. SEP 9 covers the transverse-maximum modal intensity, modal beam widths, filtered quadrature norms where an embedded rule exists, and a polarisation-span residual. A general modal guide still needs an independently qualified embedded coarse rule before it can report quadrature error. Extending the span residual to plasma/current sources needs an explicit source-space convention; neither extension is included in SEP 9’s implementation scope.

A custom spectral rate apart from the collar’s. custom_absorber sums the user’s power rate into the collar’s spectral_rate, which is what the seam carries, so the spectral channel of the far-field statistic absorbed_power reports the collar and the user’s rate together. An omit-default user_rate on Absorber, summed at translation, would let the two be attributed without moving any record (SEP 22).

An absorber switch on the keyword interfaces. prop_capillary, prop_waveguide and their Simulation.from_* constructors build the default spectral and temporal collar when absorber=None, and there is no keyword to run without an absorber or with the split-step scheme: today that takes replacing the frozen SimulationSettings before the first run (boundaries). An explicit absorber=False and an absorber_scheme= keyword would make both one argument. Related: spatial runs carry no absorber unless one is given, although every spatial grid has a collar; whether they should default to one deserves a design check. SEP 21 proposes the two keywords and a qualified convenience path for monitored propagation without temporal absorption while retaining the spectral collar; the spatial default is not in its scope.

An offset or walk-off-centred time window. The temporal grid is symmetric about the frame, but walk-off is usually one-sided: the Raman self-frequency-shift example plots −200 fs to +1.7 ps inside a ±1.75 ps window, so most of its leading half carries nothing. An offset time axis, or a frame chosen to centre the expected walk-off, would roughly halve such grids and keep content away from the temporal collar. SEP 21 measures fixed asymmetric launch placement and prescribed frames, including fastest-light and midpoint frames, and delivers qualified setting support and guidance. The remaining work is a general offset time-axis API, independently asymmetric absorber ramps, and adaptive recentering or grid expansion, with consistent history, response memory, noise and output-coordinate transport. Those require a separate design and remain outside SEP 21.

A self-focusing collapse estimate. The critical power is provided (critical_power()), but not the Marburger estimate of the collapse distance, which the radial tutorial computes by hand.

A polarisation launch across mode rows. The keyword interfaces fill only the first row of a mode set, so a Jones vector spread over the two members of a degenerate pair needs a hand-built input; a vector capillary run through the keyword interfaces is also not yet validated end to end (polarisation). SEP 31 added explicit incident-field polarisation and coherent projection across target channels. The keyword-interface convenience and its complete vector-capillary qualification remain open.

Structured scan values. A scan axis takes one-dimensional scalar values, so a list of paired gradient pressures is expressed as an integer index with the profile built in the builder (scans). Structured axis values would need explicit serialisation and provenance in the scan file.

Further offline elements. SEP 31 added focusing, apertures and ordered field transport (coupling.transport) beside the existing material, mirror, grating and prism transfer functions (compression). Its aperture shares spatial.aperture_profile and its weighting with the in-run aperture statistics, so the two agree where their plane, flux and paraxial assumptions do (SEP 23). Offline propagation through a z-varying waveguide mode as an element remains open.

Field planes in the output schema. Two changes wait for the next output-schema change: field_plane on the general modal route’s accessor whitelist (the whitelist feeds the recorded schema), and the plane’s frame recorded in FieldPlane, so that transport no longer needs a frame= argument for a plane written in a frame other than its medium’s (SEP 31, phases 2 and 4).

Regrid’s band loss in physical energy. regrid reports the share of a plane lost outside the new temporal band in the record’s norm. Reporting it in physical energy as well, through the per-bin admittance the transport core already evaluates, would give the loss in joules beside the share (SEP 31, phase 2).

A public per-bin spatial energy spectrum. The transport elements use a private _energy_spectrum (coupling/elements.py); a public spatial.spatial_energy_spectrum per frequency bin would replace it. In the same change, spatial.aperture_profile’s docstring should say that its result is a field transmission under the default "amplitude" weighting and a power weight under "quadrature" (SEP 31, phase 4).

Field planes from modified-noise runs. processing.field_plane refuses every plane of a run with noise="modified", whose saved field leaves the one-photon background to the auxiliary noise_phase. Reconstructing that phase into the plane, following the documented saved-field meaning, would lift the refusal (SEP 31, decision A2; coupling simulations).

Far-field detection beyond SEP 23. The far-field statistics (statistics, Far-field apertures and angular regions) leave the following in-run extensions for later, each as new names or fields that leave the six statistics’ shapes unchanged (SEP 23). SEP 31 added coherent offline observation planes, apertures and polarisation optics; its partial coverage is noted below.

Polarisation-resolved detection: an analyser Jones vector on ApertureSet, or Stokes-resolved aperture spectra. Per-component spectra answer only an analyser aligned with the grid’s axes; any other angle needs \(\operatorname{Re}A_xA_y^*\) too.
Tabulated radial profiles: a user-supplied \(W(\rho)\) for a measured apodisation, with its own interpolation and parity rules.
The exact mapping: an option placing a component at \(z_m\boldsymbol\kappa/k_z^{\rm vac}\) with the vacuum obliquity, for wide-angle emission, where the paraxial map’s \(O(\theta^2)\) error matters.
Other shapes: rectangular and elliptical apertures, among them a two-dimensional slit (a line grid’s aperture already is one).
Radial and planar-hybrid geometry: refused today, having no Cartesian transverse wavevector; a centred aperture on a radial grid would weight the Hankel spectrum’s rings. SEP 31 added radial offline transport and apertures; extensions to these in-run statistics remain open.

Release preparation#

The first release. Publish solprop and solprop-native at one version from a tag, following the checklist in Releasing and the native wheels: the wheels workflow with publishing enabled, the archive DOI, and a fresh install of solprop[native] on each wheel platform. The native wheels and their recipe were qualified by SEP 29.

A resume-compatibility policy before the first public release. The output schema records every field of a recorded dataclass, at its default too (ADR-051 in decisions), so a release that adds a field moves every configuration hash, and a run written by the previous release can no longer be resumed. Before 1.0, decide between computing a resume hash with the stored file’s own schema rules and reintroducing omit-at-default as a documented rule from 1.0 on.

Known issues#

Large-V step-index modes. StepIndexMode’s characteristic equation and its derivatives use unscaled modified Bessel functions kv/kvp. At large cladding arguments both underflow to zero while their ratio is finite (at argument 1600, \(K_0\) and \(K_1\) are both zero; the scaled ratio is about 6.2e-4), and the division fails. A 300 µm, NA 0.22 fibre at 250 nm fails at construction with an invalid division and no roots assigned. The fix is exponentially scaled ratios throughout the characteristic equation, derivatives, normalisation and exterior fields, qualified at large V and against independent references at ordinary V; suppressing the warning would hide the failure.

The zero-dispersion search misses paired zeros. modes.base.zdw, and the zdw statistic that uses it, test only the signs of \(\beta_2\) at the ends of the window before a bracketed root search. An even number of interior crossings — two zero-dispersion wavelengths, or the sign changes of a gas resonance inside the window, as with krypton or xenon — reads as no zero, and the search raises. The design tools’ sampled search with its pole filter does not have the defect. Reconciling the two needs a policy for several roots (all roots, a seed, or the root nearest the pump) and regression tests for zero, one and several roots and for poles.

MemoryOutput.planes() with between-plane statistics. For a run that records statistics between save planes, planes() raises OutputError (“step statistics columns must have equal row counts”). A step-statistics column created when the first plane carrying rows arrives is short by the planes saved before it. SEP 9 requires this existing defect to be fixed before qualifying its additional periodic scalar statistics.

Modal accessors on spatial files. processing.energy, get_power and get_spectral_density return a modal reading of a radial, Cartesian or planar file instead of refusing it; the numbers they give there are not energies. The scan reductions built on them already refuse spatial runs by name.

Free-space flux beyond the scalar forward model. Spatial energy, intensity and fluence are the scalar axial flux of the medium at the plane (SEP 34), and a radial gas gradient is admitted only within its real-index adiabatic certificate. Gradient scattering outside that certificate, a radial path with varying (or any) loss, which needs a complex-index bound, varying media on the Cartesian and line grids, and the full vector TE/TM flux at high numerical aperture remain open. SEP 8 requires E/H flux for its crystal route.

A lossy multimode taper at the default step bound. The taper coupling carries each channel’s loss inside its block, and over the default dz_max of half the fibre the flow’s conditioning guard refuses a strongly lossy taper. The refusal names dz_max and the taper_coupling=False opt-out; a smaller default dz_max for multimode tapers would avoid it.

Scan queues and platform probes. A live but wedged queue owner on the same host holds its point indefinitely, because the lease age is not consulted while its process is alive. An interrupted Local or Pool scan skips its running points for stale_age seconds on restart. Container memory limits are read from the root cgroup rather than the process’s own cgroup path (CPU quotas follow the path, ADR-052), so a shared node under a cgroup manager without Slurm variables is mis-sized for memory, and a failing platform probe raises outside the SolpropError hierarchy. The scan file grows on every resume, because HDF5 does not reclaim the groups it recreates. SEP 4 partly covers the platform probes: it requires effective allocation detection shared with scans, including nested cgroups or an explicit conservative allocation when detection is unavailable. Queue leases, interrupted-point recovery and HDF5 growth remain separate work.

The dispersion-only Air entry in a hand-built mixture. gas="air" on the keyword interfaces becomes a nitrogen/oxygen mixture with every channel. The built-in entry Air itself carries only a refractive index: put into Mixture.fill by hand it has no Kerr, plasma or Raman data: with kerr=True it is refused (Air has no Kerr data), but with kerr=False a hand-built Mixture.fill({"Air": 1.0}, ...) run translates with no nonlinear response and no warning. It should refuse by name or steer to Mixture.air (mixtures). SEP 3 partly covers this area by migrating air data and named compositions; the handling of this hand-built mixture still needs an explicit refusal or constituent expansion and a regression test.

Provenance of the chirped-mirror curves. The measured mirror GDD and reflectance curves read by ChirpedMirror.lookup (compression) carry no vendor citation. They are device data, outside SEP 3’s material migration, and need their sources and measurement conditions recorded, or replacing with cited curves.

Further built-in materials. SEP 3 migrates the gases and solids solprop already ships, and adds none. optmat also recommends candidates:

CO2, with dispersion, Kerr, a Raman response with provisional Q-branch widths and a rate model; the first candidate;
CO, with a rate but no Raman data;
gold and copper, as tables;
further crystals.

Each needs separate admission and qualification as a built-in material. O3 waits for absorbing gas indices and non-ideal density to be handled properly.

Memory keywords on the keyword interfaces. memory_budget= is a keyword of Simulation.run and Simulation.from_waveguide; the prop_* functions read only SOLPROP_MEMORY_BUDGET (choosing a backend). Likewise the field retention of SEP 22 (fields=, field_planes=) is set on SimulationSettings or through run_scan, not as a keyword of the prop_* functions. Offer both there together.

Two slow coupling-envelope tests in the medium tier. Two of the recorded-envelope oracles in tests/coupling/test_envelopes.py take about 30–50 s each in the medium tier. If CI time becomes a constraint, they can move to the slow tier, with a cheaper check of the same envelope left in the medium tier (SEP 31).