v2.13.0.dev0

What's Changed

Added

• Added tidy3d.plugins.autograd.SafeUpdate and tidy3d.plugins.autograd.BacktrackingSafeUpdate, integrated through tidy3d.plugins.autograd.optimize(..., safe_update=...). The latter checks global backtracking candidates first, then optionally recovers the remaining update one parameter at a time. bounds and safe_update apply to the initial parameters before optimizer initialization and the first objective evaluation.
• downsample on SpatialDataArray, TriangularGridDataset and TetrahedralGridDataset, resampling a field onto a coarser uniform Cartesian grid.
• Added tidy3d.plugins.klayout.BatchedDRCChecker for hierarchical batched KLayout DRC and max_results_per_cell support in tidy3d.plugins.klayout.DRCRunner.run(), tidy3d.plugins.klayout.run_drc_on_gds(), and tidy3d.plugins.klayout.DRCResults.load().
• Optional downsample_dl argument to Scene.perturbed_mediums_copy and Simulation.perturbed_mediums_copy, to help reduce custom medium size in large perturbed simulations.
• Added AutoUnstructuredGrid for TCAD meshing, including ratio-based baseline and targeted interface refinement, reference-relative manual region and line refinements, and helpers for mapping DistanceUnstructuredGrid sizing controls. The mapping preserves sizing parameters, but AutoUnstructuredGrid may refine explicit interface-boundary-condition surfaces that legacy DistanceUnstructuredGrid skipped.
• Field projection monitors, including DirectivityMonitor, accept use_colocated_integration=False to compute the far-field surface currents from fields at their native in-plane Yee-grid positions, including with interval_space spatial downsampling.
• Multi-mode ModeABCBoundary: pass a tuple of mode indices to mode_index to absorb several waveguide modes simultaneously at an InternalAbsorber. primary_mode_index selects the expected dominant mode and tolerance groups modes whose effective index is near-equal to the primary mode's at the central absorption frequency. The multi-mode form is only valid as InternalAbsorber.boundary_spec; the single-integer form is unchanged and remains valid as a domain boundary. A multi-mode absorber must also keep at least two grid cells between itself and the domain boundary it samples: below itself when absorbing in +, above when absorbing in -.
• FieldTimeMonitor and FluxTimeMonitor can now specify their temporal recording cadence by physical time (sampling_dt, seconds) or by a fixed number of uniformly spaced samples (num_samples), as alternatives to the integer-step interval. The three fields are mutually exclusive; an unset cadence still defaults to interval=1, so switching an existing monitor to a physical cadence requires clearing the interval in the same update, e.g. monitor.updated_copy(interval=None, sampling_dt=...). Cadences finer than the simulation time step are clamped to one sample per step, and the field-decay shutoff can end the recording before the requested schedule completes.
• A FieldStructureMonitor and its FieldStructureData record matched electric field, diagonal complex permittivity, and per-Yee-component structure-ownership maps on their native Yee grids, exposing the derived displacement field Dx/Dy/Dz (in E-field units, with colocate_displacement) and absorbed_power_density. SimulationData.plot_field(...) also accepts these displacement components and 'D' (their vector magnitude). SimulationData.optical_generation(...) turns that data into an OpticalGenerationData with absorbed power, band-gap-gated generation_rate, and thermalization_heat_rate, ready to feed the Heat/Charge solvers via DistributedGeneration/HeatSource for coupled optical→charge/heat workflows. SimulationData.per_component_optical_generation(...) returns the same quantities per electric-field component on each native Yee grid (no colocation), each tagged with its component, for interface-accurate per-component downstream use.
• Added experimental, opt-in PCA compression for compatible point, line, and planar
  FieldData-derived adjoint current sources with exact matching frequency and component sets.
  Enable it with config.adjoint.field_source_reduction_mode = "pca"; the default 0.999
  energy coverage permits approximate compression, while 1.0 retains the full numerical rank.
• Added the material-local quasi-Fermi-gradient formulation to the Selberherr impact-ionization model for accelerated isothermal charge simulations in two and three dimensions.
• SelfHeatingMonitor and SelfHeatingData record the volumetric heat a Conduction or isothermal Charge simulation generates. to_spatial_data_array(...), available on any heat-charge monitor data, converts a recorded field into a HeatSource rate for a separate Heat simulation, and SelfHeatingData.total_power() integrates the heat rate so the cost of that resampling can be checked. HeatFromElectricSource is no longer accepted in an isothermal charge simulation, which solves at several bias points; use the monitor instead. On a non-isothermal charge simulation it is ignored with a warning, since that already couples heat self-consistently.
• SteadyChargeDCAnalysis.temperature accepts a SpatialDataArray, so a Charge simulation can run on a prescribed, spatially varying lattice temperature: the temperature-dependent physics is evaluated on the supplied field, with no thermal solve and no feedback from the device. Leaving it None keeps the existing self-consistent coupling. Together with SelfHeatingMonitor this enables a manual charge → heat → charge iteration. Requires the accelerated charge solver, and warns if the field does not cover the semiconductor region.
• Added SteadyGenerationRecombinationMonitor, reporting the per-node net generation-recombination rate U = R - G entering the carrier continuity equations (recombination positive, generation negative), including every active mechanism, together with a per-mechanism breakdown, of which only impact-ionization generation is currently reported. Available only on the accelerated charge solver.
• Added "temporal" and "auto" field-source adjoint reduction modes to config.adjoint.field_source_reduction_mode: "temporal" combines the retained spatial modes into one adjoint simulation using synthesized multi-tone source waveforms, and "auto" picks between that and the separate per-mode simulations.
• Added the TIDY3D_EXECUTION_FLEXCREDIT_LIMIT environment variable for setting a process-wide cumulative maximum FlexCredit estimate for Tidy3D cloud task execution. Blocked submissions raise FlexCreditLimitExceededError before any proposed task starts, and web.execution_flexcredit_status() reports the current maximum-cost reservations.
• Added CanaliFieldDependence, an extended Canali high-field velocity-saturation model that can be attached to ConstantMobilityModel, CaugheyThomasMobility, and MasettiMobility through their new field_dependence parameter (accelerated charge solver only). Velocity saturation must be specified for both carriers of a medium or for neither. Every mobility model now serializes field_dependence.
• Added structural preprocessing cache reuse for compatible ordinary FDTD simulations. Set
  store_preprocess_cache=True on a producer job, then pass its task ID through parent_tasks to
  compatible child runs. See How do I reuse structural preprocessing in a source
  sweep?.
• Added MethodZip for deterministic paired parameter sweeps, pairing values by index across named parameters and producing N cases from N values per parameter; unequal lengths are rejected.
• Added HurkxTrapAssistedTunneling, a field enhancement of Shockley-Read-Hall recombination by elastic band-to-trap tunneling, set through the new field_enhancement parameter of ShockleyReedHallRecombination and SurfaceShockleyReedHallRecombination: it shortens the bulk carrier lifetimes and raises the surface recombination velocities according to the local electric field and the tunneling effective masses m_t_n, m_t_p. Available only on the accelerated charge solver.
• 1D (line) and point monitors now work in Charge simulations. Such a monitor has no unstructured representation, so its recorded field comes back as a bias-resolved Cartesian SpatialVoltageDataArray; select a bias with to_spatial_data_array(voltage=...). A line monitor's field can be plotted with plot_field(...) or SteadyEnergyBandData.plot(...); a point monitor holds one value per bias, which is read from the data array directly.
• UnsteadyHeatAnalysis.initial_temperature accepts a spatially varying SpatialDataArray in addition to a scalar, e.g. the temperature profile of a previous heat simulation, for cooling or heater shut-off analyses. Mesh nodes outside the field's bounding box start at the new background_temperature (default 300 K).
• Added HeatChargeSimulation.linear_solver, selecting between the default iterative linear solver and a sparse direct one that is more robust where the iterative solver stalls. On large meshes the direct solver's larger memory requirement can increase queue time, and its runtime can be longer or shorter than the iterative solver's depending on the case. It applies to charge, heat and conduction simulations alike.
• SteadyConvergenceData.cfl_history (new CFLHistoryDataArray) records the CFL number the charge solver held at each Newton iteration of every sweep bias, dimensioned (v, pseudo_step) to share the pseudo_step axis with residual_history. Reading the two together shows, for a bias that did not converge, whether the CFL had reached its cfl_min bound, its cfl_number bound, or was still adapting between them. When the trace for a non-converged bias conclusively ends on one bound, the non-convergence warning issued when the simulation finishes now names it.
• Added the Pc low-doping cutoff and mu_min2 transition-term floor to MasettiMobility, completing the Masetti model for boron-doped silicon; the defaults (Pc=0, mu_min2 equal to mu_0) keep existing results unchanged.
• Added direct autograd support for the parameters of uniform Lorentz, Drude, Debye, and Sellmeier media (eps_inf and coeffs); gradients for those media and for the poles and residues of PoleResidue and CustomPoleResidue now agree with finite differences.
• Remote adjoint gradients now reject surface sample sets larger than 5 GB before upload, with an error that names the largest structures and the settings to change to reduce the size.
• Added optional structural preprocessing-cache reuse to dipole emission studies.
• Added config.validation.mode = "fast" and from_file(..., validate="fast"), which skip tidy3d's check validators while keeping pydantic type checking, every validator that derives a value, and serialization, so already validated data loads several times faster; the per-call override is context-local and the option is never saved to disk.
• Added DipoleEmissionStudyData.angular_radiation_transfer() and angular_radiation_transfer_at_positions() for bulk-normalized dipole radiation into each far-field direction.

Changed

• config.adjoint.monitor_interval_poly and config.adjoint.monitor_interval_custom no longer affect standard shape (geometry) gradients, whose surface point-cloud sampling density is controlled by config.adjoint.default_wavelength_fraction; the interval settings still apply to the volumetric adjoint monitors used for medium, numerical-structure, and custom-vjp derivative paths.

• SurfaceRecombinationBC is now rejected at setup time when its placement names both sides of an interface that has no semiconductor side, or that is a heterojunction.

• A VoltageBC with model="ohmic" is now rejected at setup time when its placement names both sides of an interface with an insulator side and no semiconductor side, unless the other side is a ChargeConductorMedium carrying work_function.

• Shape gradients at boundary points with PEC on both sides are now exactly zero (previously a half-weighted PEC/dielectric blend), so gradients of overlapping PEC structures may change.

• SurfaceRecombinationBC no longer adds recombination where its surface meets an ohmic contact, which already fixes the carrier concentrations there, so results shift slightly for setups with that arrangement.

• electron_affinity is required on every SemiconductorMedium, including the background medium, when a charge simulation includes an insulator-only gate terminal. Different semiconductor electron affinities establish heterojunction band offsets and can change results.

• Charge terminals are now identified by the ChargeConductorMedium structure a contact is made on. One metal declared through several boundary_spec entries, or spanning several material zones, is swept, driven and current-summed as a single electrode; two different metals no longer merge into one contact even when they share a bias. A metal named by two VoltageBC entries whose sources prescribe different drives is rejected at validation time, so a saved simulation with such a setup no longer loads: a metal is equipotential. Sources are compared by what they prescribe, so a different name or attrs is not a disagreement and GroundVoltage() counts as DCVoltageSource(voltage=0).

• A VoltageBC whose placement puts a contact face on an insulator requires the accelerated charge solver; use_accelerated_solver=False is now rejected for such a simulation. This covers both a gate over an oxide and an ohmic or Schottky contact whose metal is clad by an insulator, because only the accelerated solver references such a face to the metal work function.

• A charge VoltageBC that matches no meshed face is now reported even when another VoltageBC with identical settings does match, and two VoltageBC entries with different contact models that resolve to the same face are rejected.

• The accelerated charge solver rejects an insulator contact face whose terminal reaches the semiconductor but shares no mesh node with it, since such a face cannot take its potential reference from the semiconductor. It also rejects touching contacts of different terminals that prescribe incompatible potentials or small-signal drives at a shared point, and warns once when the built-in potential along a contact tied to an insulator face varies by more than kT/q.

• Reduced peak memory when computing TriangleMesh geometry gradients by chunking surface-gradient interpolation and accumulation.

• colocate is locked to True for flux and field-projection monitors; the surface-integration scheme is controlled solely by use_colocated_integration.

• An InternalAbsorber placed against a domain boundary is now rejected: its PEC backing plate sits on the absorber's non-absorbing side and would fall outside the simulation. Move the absorber one cell further in, or absorb at the edge with a domain boundary condition.

  When a ModeABCBoundary absorbs more than one mode as InternalAbsorber.boundary_spec, its plane no longer has any effect: the absorber's own box and position determine where the modes are solved.

• Changed adaptive shape-gradient sampling to resolve per structure from the simulation definition
  and all adjoint frequencies; config.adjoint.minimum_spacing_fraction remains the active lower
  bound as a fraction of the shortest free-space adjoint wavelength.

  Geometry sampling metadata on DerivativeInfo (adaptive_vjp_spacing(),
  discretization_wavelength(), and the clip-context material length scale) is now resolved per
  structure from the simulation definition rather than from chunked permittivity data, and only for
  structures with a traced geometry path. A custom_vjp callback on a structure whose traced fields
  are all under medium now raises AdjointError from those accessors instead of returning a
  value derived from the adjoint monitor data; all other derivative inputs are unchanged.

• Changed cylinder geometry adjoint calculations to use symmetric angular sampling of the cylinder surface.

• The "Selberherr" and "PQ" impact-ionization formulations now take their driving force from the conduction-band-edge gradient rather than the electric field, and their generation rate vanishes at zero bias. Ionization is unchanged in a uniform material under bias; a band offset across a heterojunction no longer acts as a driving force.

• TriangleMesh.inside() now warns once when it uses trimesh's slower pure-Python ray_triangle backend instead of Embree.

• Adjoint simulations now warn when they end before their fields decay below the shutoff threshold that the forward simulation reached, since gradients from a truncated adjoint are biased without any visible failure.

• Charge simulations with a non-uniform lattice temperature (self-heating or a prescribed temperature field) now evaluate carrier flux, terminal currents and small-signal response with one temperature per mesh edge, so terminal currents balance under a temperature gradient. With Fermi-Dirac statistics the degeneracy-factor terms of the Jacobian and small-signal response use that same edge temperature, and the self-heating source is assembled with that edge temperature and with the carrier mobility the transport uses, so its power balance closes under a temperature gradient and with velocity saturation. Results for such simulations may shift slightly; isothermal simulations are unaffected.

• EME now reports setup errors earlier for unsupported mode-solver settings, including group-index calculation and non-increasing custom interpolation frequencies. If an older EME data file reports inconsistent mode frequencies during extraction or basis conversion, re-run the simulation to regenerate it.

• Reduced the size of gzip-packaged SimulationData results containing large monitor arrays.

• Changed tidy3d mcp to delegate to the independently released tidy3d-mcp runtime while preserving the integrated Tidy3D command.

• Documented the Tidy3D units of SolidMedium.capacity and FluidMedium.specific_heat, and recommend their from_si_units() constructors as the entry point.

• Shape (geometry) gradients in autograd are now substantially cheaper to compute — adjoint fields are recorded only at structure surfaces (pre-sampled point clouds) instead of over entire structure volumes, which shrinks recorded monitor data, downloads, and gradient postprocessing time and memory.

• The default ChargeToleranceSpec.cfl_min is now 1. Charge simulations that do not set cfl_min may behave slightly differently.

• Users can now request any positive whole-number vGPU allocation supported by their license instead of being limited to 1, 2, 4, or 8. Requests above the license limit are rejected with a clear message before submission.

• Improved PolySlab arc-discretization performance without changing the generated coordinates.

• TriangleMesh shape gradients use fewer surface sample points on long, thin triangles.

• Adjoint shape-gradient surface-sampling spacing is now bounded below by the smallest grid step near each traced structure as well as by config.adjoint.minimum_spacing_fraction of the shortest adjoint wavelength, whichever is larger, so surfaces are no longer sampled more finely than the simulation grid; this reduces the number of sample points and memory for PEC and metal structures.

• Changed the dipole emission plugin to deprecate DipoleEmissionStudyData.radiation_intensity_transfer() and its position-resolved counterpart in favor of the angular radiation transfer methods.

Fixed

• Fixed mode solver material classification to inspect the optical component of a MultiPhysicsMedium, so a wrapped FullyAnisotropicMedium selects the tensorial solver and a wrapped LossyMetalMedium is seen by the surface-impedance conductor check behind precision="auto".
• Fixed PEC shape gradients so ordinary surface samples are no longer handled as one-dimensional
  edges.
• NumPy arrays and other array-likes now work as selectors along non-spatial dimensions of unstructured datasets, in sel(), isel(), and interp().
• Fixed SurfaceRecombinationBC being rejected on an interface between two materials that both take part in the charge simulation, such as a semiconductor/insulator interface or a junction between two differently doped regions of the same material.
• Fixed longitudinal mode-field grid corrections and near-zero clipping at simulation boundaries.
• Fixed VoltageBC contacts placed on an insulator (a gate over an oxide) in charge simulations: the gate potential is now referenced to the metal's work_function, giving the correct flat-band voltage instead of one shifted by the work function minus the semiconductor electron affinity. Two existing setups shift with it: a gate whose metal already carried a work_function for an unrelated Schottky contact, and the oxide-clad face of a Schottky contact that reaches no semiconductor node of its own.
• AutoGrid no longer over-refines the regions on either side of a narrow feature when the maximum step size in effect there is much coarser than the spacing between nearby structure boundaries. Steps now coarsen away from the feature by up to the max_scale ratio instead of one side being tiled uniformly at the feature's own step size, which also made two regions equally distant from the feature grid differently. QuasiUniformGrid is affected in the same way: when an interval between structure boundaries or snapping points is shorter than the requested dl, the rest of the axis is no longer refined below dl on account of it, and the cells stay within dl and the max_scale ratio while generally being fewer and coarser.
• Reduced mode-solver setup time when precision="auto" or a tensorial solver has to be selected, with classification results unchanged. Most noticeable for EME, where every cell plane paid this cost.
• TriangleMesh shape gradients now require a closed, outward-oriented volume mesh.
• perturbed_mediums_copy() accepts a plain spatial xarray.DataArray on every path, and rejects one still carrying a non-spatial coordinate with a message naming the field and how to drop it.
• Fixed fully anisotropic mode operators for magnetic coupling, invariant dimensions, PML, and backward propagation.
• On a periodic axis whose neighbouring regions have different step sizes, AutoGrid now sizes the cells near the boundary from the region actually adjacent across it; they could previously come out coarser or finer than intended, while still respecting the local maximum step size and the max_scale ratio.
• The error from Job.task_id on a multi-step job, which every HeatSimulation and HeatChargeSimulation job is, now names job.estimate_cost() and job.real_cost() alongside job.task_ids.
• Fixed mode-field rotation across the angular branch cut and bend preprocessing of off-plane media.
• Fixed subpixel continuity for derived custom media whose parent is the first material entry.
• Fixed slanted PolySlab.inside() and Transformed.inside() for same-shape coordinate inputs supplied as scalars, lists, or arrays, and fixed Transformed.inside_meshgrid() for transformed slanted PolySlab geometries with 1D coordinate vectors.
• Fixed the local field projection (FieldProjector) handling of integration bounds, spatially downsampled near fields, and projection windowing, so that locally projected fields agree much more closely with server-side projections of the same near-field data.
• Fixed slightly mistuned absorption in ABCBoundary and ModeABCBoundary: the propagation index, including an explicitly set permittivity, now accounts for the grid's numerical dispersion, which improves absorption most on coarse meshes.
• Surface monitors now reject an empty fields tuple up front, rather than being accepted and then failing when their data is loaded.
• Fixed inaccurate adjoint shape gradients for field-monitor objectives with staggered field samples near absorbing simulation boundaries.
• Fixed a small permittivity error at dielectric-vacuum interfaces when edge_singularity_correction is enabled, which slightly shifted results for geometries such as metal traces on a substrate.
• Fixed excess reflection from InternalAbsorber / ModeABCBoundary absorbers when a conductor passes through the absorber plane with edge_singularity_correction enabled.
• Fixed the small-signal AC conductance of charge simulations with Fermi-Dirac statistics, which omitted the carrier-density dependence of the degeneracy factor and was overestimated at high doping.
• A HeatSource with a spatially varying rate (a SpatialDataArray) is now applied instead of being silently ignored, so affected heat simulations no longer solve as if no heat were generated.
• Fixed vacuum fallback when serialized CustomMedium data arrays are unavailable.
• Fixed CustomMedium adjoint validation to reject gradients for inactive data representations.
• CustomAnisotropicMedium.is_spatially_uniform now returns True only when all tensor components are spatially uniform, and False when any component varies in space.
• A TemperatureMonitor in an isothermal Charge simulation now reports the uniform lattice temperature the run was solved at instead of 0 K.
• Fixed instability of an InternalAbsorber whose plane crosses a conductor, which could make the simulation diverge while the equivalent ModeABCBoundary placed at a domain edge stayed stable.
• A MeshOverrideStructure with shadow=False whose boundary falls very close to a grid snapping point now reuses that snapping point as its boundary, instead of adding a separate grid line beside it and refining the resulting sliver. Meshes in affected setups may have slightly fewer or shifted grid lines than before.
• Fixed batches restored with Batch.from_file() failing to serialize again or load in a fresh Python process.
• The fill_fraction_box metric now uses the tangential projection of the bounding box, so direct queries, filtering, and sorting ignore its position and extent along the mode-plane normal axis. Mode sources and monitors in Simulation, user-supplied mode monitors and cell modes in EMESimulation, ModeSimulation, and ModeSolver now validate that the box intersects the effective mode-solving domain along both tangential axes before solving. Direct ModeSolverData.sort_modes() calls remain permissive: a tangentially disjoint box produces an all-zero metric.
• Fixed ModeSolver accepting custom microwave impedance paths outside the mode plane.
• Fixed web.load_simulation() and web.download_json() failing for mode solver and component modeler tasks, and web.download_simulation() failing for component modeler tasks (modeler batch ids remain unsupported by these functions).
• Lossy conductive media are no longer treated as perfect conductors by the mode solver at low frequency.
• Fixed validation so semiconductor simulations with electrical boundary conditions require an
  explicit charge analysis_spec.
• Fixed ModeSimulation.reduced_simulation_copy to preserve permittivity and medium monitors and their complete sampled regions.
• Fixed EME mode-solver monitor frequency handling for interpolated cell modes, including reduced-data storage, basis conversion, and cases where some requested modes are unavailable.
• Fixed mode solves failing with an IndexError when the mode plane coincided with, or lay beyond, the outermost grid point along its normal direction. Where the correction could still be computed it was in some cases taken from the wrong side of the plane. Finite-grid corrections are now computed against the original simulation grid, and are skipped with a warning when the plane is not bracketed by that grid. A mode plane placed exactly on the simulation boundary is now rejected with an explanatory error instead of failing inside the grid machinery.
• Fixed Simulation.to_gds_file() retaining lower-priority geometry on the same GDS layer beneath thresholded custom-medium structures, including custom optical media wrapped in MultiPhysicsMedium.
• Fixed EME validation and classification to inspect the electromagnetic material stored in MultiPhysicsMedium.optical. A wrapped AnisotropicMedium, FullyAnisotropicMedium or custom medium is now seen by the bent-anisotropy checks, the global-frame custom-medium guard, and the reciprocity check that rejects non-reciprocal fully anisotropic media, all of which previously treated a wrapped medium as absent.
• Fixed 2D unstructured charge monitor data to preserve separate material-side values at coincident interface nodes.
• Fixed frequency tracking rotating mode phases instead of aligning them, and mode profiles not being normalized in the dot-product convention their amplitudes are decomposed in. Recorded amplitudes and phases can change for lossy modes, so stored results may need re-baselining.
• Adjoint runs now immediately report failed simulation tasks, including server-provided failure details, instead of processing unavailable gradient data.
• Fixed a duplicated density unit conversion that made the transient heat characteristic-time estimate 1e18 too small, so the excessive-simulation-time warning fired for every physically meaningful UnsteadyHeatAnalysis.
• In a 2D Charge simulation, a monitor that is also flat along one of the two resolved axes now records a line instead of failing. SteadyCapacitanceMonitor, SteadyElectricFieldMonitor and SteadyCurrentDensityMonitor still need two non-zero dimensions, and now say so before the simulation runs rather than failing during postprocessing.
• HeatChargeSimulation now reports a heat boundary condition placed where no solid material is present: an error when nothing else anchors the temperature -- previously the condition was silently dropped, and a steady solve left unanchored that way returned non-physical values under a successful task status -- and a warning when another condition still anchors it.
• A slice taken across a material interface no longer reports one material on both sides of it. Where the grid holds a node per side — how a heterojunction carries its band offset, or a contact resistance its temperature step — a line profile now keeps both values at distinct coordinates, separated by a small fraction of the local sample spacing, so the step is no longer merged away when the profile is selected from, interpolated, or reordered. Because both samples now sit just off the interface, its own coordinate no longer names one: select a side, or pass method="nearest". A 2D plane slice of a 3D simulation is fixed the same way. A slice lying exactly along an interface can only report one side and now warns, instead of silently choosing one by its position in the grid. Slicing a tetrahedral dataset works again where the plane cuts a cell at more than two of its vertices.
• UnsteadyHeatAnalysis.initial_temperature and background_temperature now reject infinite and NaN scalars, and a SpatialDataArray initial temperature must be real, non-empty and on finite coordinates. These previously passed validation and reached the solver as an unusable initial condition.
• Custom doping is no longer dropped at a doping box face whose coordinate is not representable in single precision, such as the ±0.11 µm bounds of a 220 nm layer. A mesh node sitting exactly on such a face was treated as outside the box and doped to zero; it now takes the box's concentration. Charge results for devices with a junction on a layer face change accordingly.
• Fixed two-element real arrays being incorrectly parsed as complex numbers during deserialization.
• Fixed symmetric Simulation.to_gds and Simulation.to_gds_file exports for nonzero simulation centers to reflect around the simulation center.
• Changing config.adjoint.default_wavelength_fraction or config.adjoint.minimum_spacing_fraction no longer warns that remote gradients ignore them; both now take effect for remote gradients.
• Fixed 3D PEC shape gradients so that surface points within half a grid cell of an edge are integrated as PEC instead of dielectric.
• Reduced validation time when bent anisotropic EME simulations use length sweeps or are reused during mode-solver setup.
• HeatChargeSimulation now detects invalid volume-mesh connectivity early and retries automatically instead of failing later with an unclear conversion error.
• HeatChargeSimulation now warns when relative_min_dl overrides a mesh size requested through dl_bulk or through a grid refinement region or line.
• Fixed the dipole emission plugin's radiation intensity for non-unit collection index and source amplitude, and matched DipoleEmissionStudyData transfer normalization to the simulated source spectrum.
• Setting config.microwave.suppress_rf_license_warning to True now suppresses the RF license warning emitted during simulation validation.

Removed

• Removed tidy3d.plugins.microwave and the terminal half of tidy3d.plugins.smatrix, including TerminalComponentModeler, its lumped and wave ports and the RF material library, after their move to flexcompute-rf; the old import paths now report the flexcompute.rf.tidy3d replacement before they fail, and the modal component modeler is unaffected.

Breaking Changes

• Removed sweep_index from sweep-invariant EME mode, flux, correction, and aggregate-overlap data; singleton axes in released legacy result datasets are normalized, while multi-entry legacy axes must be selected or migrated before loading.
• Removed EMESweepSpec.sweep_modes and EMESimulationData.port_modes_list_sweep; EMESimulationData.port_modes_tuple now exposes the single shared port basis used by all supported sweeps.
• EME overlap FlexCredit estimates now include both per-cell self-overlap and per-interface cross-overlap tasks, including the union of interfaces required by periodicity sweeps, so estimated costs increase for existing jobs.
• Removed the deprecated EMEFreqSweep; list physical frequencies directly in EMESimulation.freqs and use EMEModeSpec.interp_spec, or migrate saved frequency-sweep results with an earlier Tidy3D version by multiplying their base frequencies by freq_scale_factors.
• Selecting a voltage when plotting charge monitor data now requires a bias the simulation solved at; a value between two recorded bias points raises DataError instead of quietly snapping to the nearest one. This applies to unstructured (2D and 3D) monitor data as well as the Cartesian data recorded by a line or point monitor.

Planned Deprecation

• Deprecated the FieldProjector.trapezoid helper, which is no longer used by the local field projection and will be removed in Tidy3D 2.13; use numpy.trapezoid instead.
• tidy3d.rf now carries only the RF classes defined under tidy3d.components and will be removed in 3.0; RF development continues in flexcompute.rf.tidy3d.
• Explicitly setting precision="single" or precision="auto" on optical, EME, or microwave/RF mode specs is deprecated; use "double" instead, which is already the default for all of them. "auto" resolves to single precision unless the mode solve contains a good conductor, so it carries the same accuracy risk.
• unstructured=False is deprecated on TemperatureMonitor and SteadyPotentialMonitor, the only heat-charge monitors that still accept it, and will be removed in Tidy3D 3.0; use to_spatial_data_array() on the monitor data where a Cartesian SpatialDataArray is needed.

Full Changelog: v2.12.0...v2.13.0.dev0