3D EM

What circuitRF's 3D solvers are for, how to get them, how to set one up and read the answer, and a worked example you can open.

circuitRF can hand a layout to a 3D solver: Palace or EMerge, two finite-element (FEM) solvers, or openEMS, a finite-difference time-domain (FDTD) solver. It builds the 3D model itself, from the layout, its technology and its bond wires. It meshes the model with Gmsh, runs the solver, and reads the answer back as an ordinary result. The same .cem setup document drives it. You choose a 3D solver in that setup's Solver group, and nothing else about the workflow changes.

This page is the guide. The setup panel's 3D controls, and every .cem field they write, are described control by control in EM Setup ▸ Solver — planar or 3D. This page links into that one rather than repeating it.

When to use 3D, and when not

The planar solvers stay the default wherever they fit. circuitRF's planar method-of-moments solver (see the MoM engine) and the bond-wire kernel in wBond put their unknowns on the metal only. A 3D solver meshes the volume, including the air around the structure. A 3D run therefore costs minutes and gigabytes where a planar run costs seconds. A 3D solver is also a separate program you install. Use 3D for what the planar solvers cannot represent, and as an independent check on them. Do not use it as a replacement.

Which 3D solver. FEM (Palace, EMerge) and FDTD (openEMS) suit different geometry. The setup panel says which fits your model when you choose one. This is advice, not a restriction. The full comparison of FEM, FDTD and the planar MoM solver — cost, memory, and which problems suit each — is EM Solvers: MoM, FEM and FDTD.

• FEM (Palace) fits curved and diagonal metal: bond wires, round vias. Its tetrahedra follow a surface, and its adaptive refinement puts small elements only where the error is. It is the only one of the two that computes capacitance and inductance matrices, and it finds a cavity's resonances directly. FDTD would have to ring a high-Q cavity down for a long time.
• FEM (EMerge) fits the same geometry as Palace and finds resonances too. It installs in about two minutes with nothing to build and solves a small driven model in seconds, with lumped or wave ports. It is a second FEM answer that shares no code with Palace's.
• FDTD (openEMS) fits Manhattan geometry on a stackup, because metal aligned with its grid costs nothing. One pulse covers a whole band. Its absorbing boundaries let the air box around a radiator stay small. One fine feature sets its cell size, and through that its time step, so a single thin wire in a large model is expensive.

Running two or more on one setup (Compare…) gives answers from solvers that share no code, and FEM and FDTD share nothing numerically either; their agreement is strong evidence that each has converged. It is a cross-check, not a reference: every solver reads the same generated model, so a mistake in building that model would appear in all of them.

Getting the solvers

circuitRF ships no solver and no mesher. Opening a workspace, editing a 3D setup and looking at its model need none of them. Simulate is the first step that does. On a machine without them it is refused, naming what is missing, and the message offers Install Palace … (or Gmsh, openEMS or EMerge). Settings ▸ 3D EM has the same Install … buttons, and circuitrf solver install palace --yes does the same from a terminal. EM Setup ▸ Letting circuitRF install the 3D solvers is the full description: what is shown before anything downloads, where each program goes, and what happens when an install fails.

What it costs, as measured. Each figure is one machine's measurement, and yours will differ:

Palace is always built from source, because no ready-built Palace exists for macOS or for the Linux subsystem. The build is long but runs in the background, with its progress in the Messages panel, and Cancel stops it at any point.

On Windows, Palace runs inside your own Windows Subsystem for Linux distribution, and circuitRF treats it like a Palace on this computer. Gmsh, openEMS and EMerge run natively. Installing Palace on Windows walks through it step by step, and EM Setup ▸ Palace on Windows has what each missing piece's message says.

By hand. If you prefer to build the solvers yourself, circuitRF finds them on PATH, in a Spack install tree, in a conda environment, or where you name them in Settings ▸ Solvers. An EMerge of your own is never taken from PATH: name its Python, or its virtual environment's folder, in that row or in CIRCUITRF_EMERGE. It runs only the Palace versions it has validated; any Gmsh, openEMS or EMerge runs, with a warning when it is not a version circuitRF has tested. The exact versions and the three build problems met on macOS are in EM Setup ▸ Installing the 3D solvers by hand.

Removing them. Uninstall … in Settings ▸ Solvers, or circuitrf solver remove palace --yes. circuitRF removes only what it installed, and never touches your documents or results. See EM Setup ▸ Removing the 3D solvers.

Setting one up

A 3D setup is an ordinary EM setup whose Solver is not Planar. Create one as any other (New ▸ EM Setup on a layout, see EM Setup), then choose the solver. Every control writes a field of the .cem, so a setup can equally be written as text. The fields below link to their full description.

• Solver (Solver3D): Palace, OpenEms, or Both to run the two and compare them. More.

• Problem (Problem3D, Palace only): Driven for S-parameters (the default), Electrostatic for a capacitance matrix, Magnetostatic for an inductance matrix, Eigenmode for resonant frequencies and

  17. Package RLC, Eigenmodes. When an electrostatic solve is the right way to get a capacitance, and how to put the answer in a schematic, is AN-03 Extracting capacitance.

• Ports. A layout's port labels become ports. In 3D a port is lumped by default: a sheet from the line's end to its return, with the port's Z₀ across it. The return is found in the layout: the nearer of the first plane with metal under the end below the line and the first above it, a ground-reference layer preferred, and a port can state another in the port table's Return column (Return). A wave port (Ports3D) is fed and measured in the line's own field at the air box's face instead, and has no sheet parasitic. Both solvers build one, except that openEMS cannot feed a hollow waveguide and Palace cannot give a port with several terminals. Wave ports.

  How wide a lumped port's sheet is. As wide as the port's bar in the layout, which the port's Use edge width and Width set (Ports): the whole conductor edge with the box ticked; otherwise the Width you state, or 2 × H (H, the dielectric height from the port's conductor to its return) centred on the port, cut short where the edge ends. The run's notes say which, and why. A sheet across a whole edge can reach metal the port was never near, such as a via on a pour's edge; when that refuses the run, the refusal names Use edge width as the way to narrow it. On a stepped end, the edge is the straight run the port stands on, and the copper beside it stays. Two ports on one sheet are refused, naming both.

  A port can drive a line on a layer the stackup marks as ground, such as a trace in a clearance of a bottom pour, when the port is on that layer. A port on the layer's largest piece, the plane itself, is refused.

  A port on a curved edge. A port on a round pad's arc has a flat sheet against a curved outline, which leaves slivers between the two that the mesher has to resolve. The run cuts the pad flat across the sheet's width, at the sheet, and its notes say how much metal that removed.

  An internal port, placed inside a pad rather than at a line's end, is a pad-to-plane port: a vertical lumped sheet from the pad to the nearest ground-reference plane, above or below, that has metal across the whole sheet. A plane with an opening under the pad is passed through, so a cut-out under a test pad moves the port's return to the next plane down, which is what the cut-out does to the real pad. The sheet is centred on the label and is as wide as the pad is across it. When a via stands at the label (a test point's through via), the sheet stands beside the barrel instead, on the side away from the trace that feeds the pad, and moves out to where the plane's clearance round the via ends. A port whose Return states a layer lands on that layer only. Metal of another layer across the sheet's path is refused, naming that layer. A delta-gap port (a cut across a line) is planar only.

  Ports a component carries. A port label inside a placed part is that part's terminal and never a port of the board — except a port the component declares as an EM port. A TESTPAD with Centre = Port is one: its pad centre becomes an internal port numbered PortNum, wherever the part is placed, turned or mirrored, and in 3D it is the pad-to-plane port above. A planar setup takes it as an internal port to ground. Its number is checked against the board's own port labels and every other component's: a number used twice is refused, naming both.

  A port across each part's own pads (ComponentPorts). To co-simulate a board with the parts placed on it, each part needs a port of the board's N-port at its own pads. The setup names them by reference designator — ["C1", "C2", "L2"], or "*" for every placed two-terminal part — in its Component ports list (tick the parts; Check shows what would be built) or with em <view.c3d> --component-ports "C1,C2" for one run. That setup is a 3D setup of a 3D view that places the board's layout, or a .cem on the layout, Planar or 3D: the list is stored in the setup that names it, never in the layout (Where a port is stored), and on a 3D .cem it is built in the 3D view described here, made in memory. Component ports has the Planar construction and the limits. In 3D each is a lumped gap port between the part's two pads: its ends at the closest copper facing the gap across the port's width — the pad, a trace's round cap, a zone's arc — and where that copper is not one straight face, a landing rectangle extends it to one, reaching back into solid copper across the whole width. So a port between pads that end in rounded or scalloped copper, as imported artwork often does, meets that copper along both of its edges rather than at a few points. Landings are part of the solved problem and are drawn in the 3D view as part of the copper they extend, so the view shows what is solved; the layout is not changed. Before anything is meshed, in check as in the run, each port is verified: no landing touches other copper (it would join two nets), each end of the sheet touches its own pad's copper and nothing else, its sides touch nothing, no copper lies inside it, and its two pads are on different nets. A part that fails is left out of the run with a warning naming it and what was found, and the other parts' ports are built; the run refuses only when no port is left. Port numbers follow the 3D view's own ports, in list order, without a gap; an entry may state its own Z0, its Reference (one of the part's pins) to say which pad is the negative end, and its WidthUm, the port's width across the gap in micrometres, centred where the pads put it (omitted: the width the two pads share). An entry beside "*" is that part's override. The .sNp a run writes starts its header with one line per port saying which number is which port, a component port by its part (! Port[3] = C1, ! Port[4] = U1.IN). The ports are drawn in the 3D view and listed under Component ports in the Object Tree as soon as the setup names them, a part left out with why; selecting one shows its Z0 and width in the Properties Inspector, which writes them into the part's entry. Its place and orientation come from the pads, so they follow the layout. A part with more pads, such as an amplifier, gets one gap port per signal pad against its own ground pad (a pad on the ground net, or named GND); a part whose ground pads on top and bottom copper lie over each other is a coaxial launch, and its shell is modelled by a copper block through the substrate joining the two. explain <view.c3d> --component-ports lists each port's pads, conductors, gap, width, landings and verdict without meshing. A finished driven run also writes <setup>_cosim.cnl: the board's block, a Port for each of the 3D view's own ports and each connector, and each part's model from the board's schematic from its port's node to ground — a part with no model there is a placeholder value, named in a comment and by check. With its radiation pattern on, part values then move the pattern with no new run: Co-simulation. A 3D view of a cell's layout is made by Design ▸ New 3D View from Layout, or new view3d <cell> --from-layout. New 3D View brings in the component ports of each .cem on that layout, a Planar .cem's as a 3D setup of its own named after it. After that, the two are separate: changing the parts in one leaves the other alone. Each part is checked again in 3D and may be left out there.

  Why a gap port, and not a pad-to-plane port at each pad. A part between two pad-to-plane ports carries its current, as far as the EM is concerned, down one port's sheet, along the plane and up the other's: a loop that is not on the board. A gap port puts the part where it is, across its own pads, and the co-simulation connects it from that port's node to ground — never between two port nodes. A hand-drawn pair of pad-to-plane ports is a different, approximate model; its error grows with the pads' spacing against the substrate's height.

  Port numbers run 1 to N. A Touchstone file numbers its ports by position, not by name, so a layout whose ports are 1, 3 and 4 would write P3 as the file's port 2, and a schematic that connects the file connects it by position. A run, planar or 3D, is therefore refused while the numbers skip one, and the refusal names what is missing (Port numbers skip 2: the 3 ports are 1, 3, 4 …); check and explain say the same. Renumber the ports 1 to N. A port outside the setup's solve region is refused too, by name.

• Terminals (Terminals3D, static solves): the matrix's rows and columns, each named by a net, an object's path or the end of one. An electrostatic terminal takes the whole piece of metal it names; a magnetostatic terminal also names the port its current enters by. Metal in no terminal floats, is ground or refuses the run, as Other metal (UnassignedConductors3D) says. Static problems.

• The air box (AirBox): how far each face of the solved region lies beyond the geometry, and what the face does to the field. It can be Absorbing (open space), PEC (a metal wall), PMC or Symmetry. A face's padding is a share of the geometry's extent along that face's own axis: 0 % touches the geometry, and 100 % on +Z puts one board-thickness of air above it. By default, on a driven setup, the top is padded by 100 % of the geometry's height, the four sides by 20 H, where H is the dielectric height under the top metal, and the floor by 10 H, where H is the dielectric height over the bottom metal. Each is rounded to the nearest 5 % of its axis. No default is more than an eighth of the longest wavelength or the structure's largest extent. On a 1.6 mm board whose top dielectric is 0.2 mm that is about 4 mm of air on the sides and 1.6 mm on top; state a larger +Z padding when the field above the board needs more room. The floor sits on the lowest ground plane, as a PEC face, when that plane is undrawn, or when it is drawn under all the other metal. A drawn bottom conductor that covers 95 % of the footprint is that floor even when the technology does not mark it as ground, and the notes say so. A hole in that plane is allowed when nothing of another net lies in or over it, so the drill holes of ground vias do not keep the floor off. A hole with a signal via or other metal of another net in it does, because a floor there would short that metal to ground. When a hole keeps the floor off, the notes say where it is. A setup that asks for a radiation pattern keeps every face absorbing, padded by a quarter wavelength at its pattern frequency (PatternFrequencyGHz, the sweep's lowest frequency when none is set). That is never more than the structure's largest extent, and never less than an eighth of a wavelength at the sweep's lowest frequency. A static or eigenmode setup keeps the wavelength rule. A closed box of PEC faces is a shielded enclosure, and it is how the example models a package lid. Each padding box in a 3D view's Properties panel shows the padding as a length, and each face row gives that length with the rule that set it (30.6 mm — λ/4 at 2.45 GHz, 0.69 mm — 10 × H). Type a length, or a number ending in % for a share of the extent. The tooltip and circuitrf explain give the rule in full (default: 4.572 mm, 20 × H, H = 203.2 µm of dielectric under the top metal, to the nearest 5 %).

• Robust meshing (RobustMeshing): removes outline points that sit within a fraction of a micrometre of their neighbours' line, which imported artwork often has and which can stop the mesher. On unless the setup says false; ports are inferred on the outline it leaves, by check and by the run alike. See EM Setup ▸ Robust meshing.

• Thin Metal Correction (ThinMetalCorrection): on a driven Palace or EMerge run, gives a conductor under three skin depths thick the impedance of its thickness, so its resistance levels off at the DC value. On unless the setup says false. See EM Setup ▸ Thin Metal Correction.

• Vias are solid copper barrels in every EM problem, whatever their stackup entry's Fill says: a plated barrel's air core changes nothing an RF answer can see and makes the mesher build inverted elements. A thermal solve keeps the wall, which is where the heat runs. circuitrf explain names each entry that was built solid.

• Mesh order. A layout's geometry, whose only curved surfaces are its drilled barrels, is meshed with straight-sided second-order elements; on an imported board that took Gmsh 10 to 30 s where curving them ran for more than a quarter of an hour. Genuinely curved geometry keeps curved elements, for at most MeshCurvingMinutes (5 by default) before the mesh is made straight-sided, and the run says when it was.

• Quality (Quality in the Palace section): Draft, Standard or Accurate. A new setup starts at Draft; one that names none runs at Standard. These set the element order, the refinement passes and the sweep tolerance together, and any one of those can be overridden on its own. The measured cost of each preset is in EM Setup ▸ Solver. Which one is enough depends on the structure, not on a rule: the example below uses both Draft and Standard, and says why for each.

• Fields (SaveFieldsGHz): which frequencies' fields are saved for the 3D view. By default only the sweep's centre is saved, because fields are large. Fields for the 3D view.

Leaving something out. In a 3D view, untick an object's, a placed cell's or a port's Model to solve without it while keeping it drawn. A run leaves it out entirely — the air box too is sized without it — and its notes name what was left out. A port that is off is absent, not terminated: no excitation, no sheet, no lumped element, so the gap it spanned is open. The result's ports are the modelled ones renumbered 1…N in their order (P2 off: P1, P3, P4 become 1, 2, 3), and the .sNp header carries one circuitRF-EM 3D port map: line per port, and the .npy a DocumentPort array, so the file says which is which. A reference to something that is off — a modelled port's conductor, a wire's pad, a face boundary — is refused naming both, never dropped; every port off is refused (there is nothing to excite).

Look before you solve. Show 3D on the setup's panel opens the 3D view on the model circuitRF built. It works before any solver is installed, and a port that cannot be built does not empty it: the board is drawn without that port, and the reason the setup would not run is shown across the top of the view. circuitrf explain on the .cem prints the same model as text: every solid, its material and its extent, the ports, the air box and its faces, and an estimate of the problem's size. A .c3d that embeds several setups takes --setup <name> to say which one. circuitrf render x.cem --section xz@y=0um -o cut.svg draws a section through it.

Watching it run, and reading the answer

Before anything runs, circuitRF checks whether the run fits in this machine's memory. The first check uses the model's volumes and the refined shell round each conductor and port. A second check, once Gmsh has reported how many tetrahedra it made, uses the real count. On a bond wire or another small curved conductor the second is the one that counts, because the refinement its curvature asks for is nearly the whole mesh. Past 75 % of memory the run carries a warning naming the estimate and what would shrink it. Past 100 %, Simulate asks before going on.

While it runs, the progress row names the solver's own stages, read from its own log: Meshing (Gmsh) with Gmsh's own stage (curves, surfaces, 3D, the second-order elements, and high-order optimisation with the node count it works on), and gmsh.log in the run folder fills as Gmsh runs; Solving: refinement pass k of N with the unknown count; Adaptive sweep, with the model change — how much the frequency model still changes when a new frequency is solved — and the tolerance it is done below; Reading results. The memory the solver is using is beside it. A 3D run cannot stop early and keep a partial result, so the button reads Cancel.

When it finishes, one line states what the run cost: the wall time, the solver's own peak memory, the tetrahedra before and after refinement, the unknowns, and the preset. The same line is recorded in the result's header.

A driven Palace run also states the power each port loses: driving that port with the others matched, 1 − Σᵢ|Sᵢⱼ|², at the sweep's lowest and highest frequency. This is power that comes out of no port: it went into the dielectric and the metal, or out through an absorbing face. A lossless circuit model cannot show it, and |S21| hides it behind mismatch. A port that loses more than three times the median port, and more than 1 %, is named. If the loss is nearly the same at both ends of the band, it is unlikely to be dielectric loss or radiation, which both grow with frequency. Look at what that port's field reaches.

Where the answer lands. A result is named after its solver, so a 3D run never replaces a planar one or the other solver's:

For a setup embedded in a .c3d, <setup> is the file's stem and the setup's name with a space between: the setup Driven in board.c3d writes results/board Driven.palace.s15p for fifteen ports. -o moves the Touchstone only. run analysis=em lists every file it wrote in its outputs, the Touchstone first, so the path never has to be put together by hand.

results/<setup>.palace/ keeps everything the run made: the Gmsh script, the mesh, the Palace configuration, both programs' logs, and the saved fields. The run can be repeated by hand from it. An unchanged model reuses its mesh. results/<setup>.emerge/ is EMerge's: the model as data (problem.json), the driver script that reads it, EMerge's log, the arrays it wrote and the saved fields, in the layout a Palace run's fields use, so every field plot draws either. A static or eigenmode setup keeps its own folder beside the driven one: results/<setup>.palace_es/, _ms/ or _eig/.

A port with two terminals on Palace (On Palace) also keeps what its terminal S was converted from, per frequency, in the palace group of <setup>.palace_em.npy: modal_S and modal_V (Palace's modal S and the voltage along each terminal's path, port by excitation), modal_Zpv and modal_kn (each mode's impedance and wavenumber), and modal_G and modal_K (the fitted overlap of a face's modes and its Robin correction). A disputed result can be re-derived from them without running Palace again.

Is this solved?

A 3D run takes minutes to hours, so a 3D view (.c3d) says, for each of its setups, whether a run's result still matches the model. You can tell without opening the fields, and the answer is still there after you close and reopen the document or the workspace.

What "solved" means. A result exists, and nothing the run read has changed since. That covers the 3D view itself (unsaved edits included) and every file the run was solved from: placed layouts and their sub-cells, nested 3D views, the technology, and its material libraries. A circuit-driven thermal run also counts the schematic it ran. A run keeps a copy of the model it solved beside its result, and the check compares contents, not an edit count.

What makes a result out of date. An edit to anything that setup's solver reads: geometry, materials, variables, an object's Model switch (Model: kept in the drawing, left out of the solve in Drawing in 3D), and the setup itself. Each setup is judged on its own:

• Another setup never counts. Adding, editing or removing one setup leaves every other setup's result as it was. The exception is a thermal submodel, whose result is cut from its From setup's, so editing that setup counts for it.
• An EM result also counts ports, face boundaries and the air box's material. Heat sources, probes, contact resistances, effective blocks, symmetry planes and the wire ground plane are thermal only and never count. Mesh regions count for Palace but not for openEMS, which builds its own grid.
• A thermal result also counts heat sources, probes and the rest of the thermal places, and only the ports its currents are driven through (every port, when a current comes from a circuit). Adding an EM port, or editing one no thermal current uses, never counts. Nor do face boundaries or the air box's material.

What never counts, for any setup: hiding or showing an object, its transparency, the field plots, choosing the active setup, the display unit and the snap grid. Those are how the model is drawn, not what is solved. Undoing back to the solved model makes the result current again.

The marks. Each method has its own shape: a diamond for FEM (Palace or EMerge), a triangle for openEMS (FDTD) and a circle for a thermal result; the mark's tooltip names the program. Each method shows at most one mark:

• Filled, solid: the active setup has a result of this solver, and it matches the model.
• Hollow: the active setup has a result of this solver, and the model has changed since. This wins even when another setup's result is current.
• Faded, filled: the active setup has no result of this solver, but another setup does, and it matches the model.
• No mark: nothing of this solver matches the model. An out-of-date result of a setup you are not using shows no mark. The setup list says it in words.
• A star after a mark (*): that result is partial. The run was cancelled, interrupted (circuitRF closed while it ran), or did not converge, such as openEMS stopping at its step limit. A partial result can still be current.

A Compare setup counts for each solver it runs. Shape and fill carry the meaning, so the marks read the same without colour. Hover over a mark for its words: the setup, when it was solved and how long it took, or what has changed since.

Where it appears.

• The setup list (Simulate ▸ Setup Analyses…, and the Analyses panel while a 3D view is active). Each setup's card has a status line: Solved 14:32 (18 min), Solved 14:32 (18 min), not converged, Out of date: 'Board.clay' has changed, Cancelled 14:32: no complete result, Not run. A Compare setup has a line per solver. This is where to decide whether to run.
• The document tab, after the name. The unsaved-changes dot stays in front of the name.
• The workspace tree, after a .c3d's name. An open document shows what is in the editor, unsaved edits included. A closed one shows what its file says.
• A torn-off window's title, when its active document is a 3D view, as words: Connector.c3d — solved (FEM, thermal). Only filled, solid marks are named, with partial where it applies. The main window's title names the workspace and does not change.

The checks run in the background and never delay opening a document or a workspace. A mark appears once its check finishes, a moment after the document is on screen. After an edit, the check waits for half a second of quiet. It runs again when any run finishes and when a file a 3D view is solved from is saved.

Before a re-run. When the setup you run already has a complete, current result, Simulate asks first: Palace result for 'EM1' is current (solved 14:32, took 18 min). Run again? A Both setup names both results. A partial or out-of-date result runs with no question, because there is no complete result to lose.

From the command line. circuitrf explain <view.c3d> has a Solved section: one line per setup and solver with its state, whether it is partial, when it was solved, how long it took and what has changed. --json carries it too. circuitrf find lists each 3D view's setups with their state. circuitrf em never asks before a re-run; it notes on stderr that the result was already current and runs. See The command line.

The 3D view

Show 3D on a setup's panel opens the 3D view beside it, drawn by the graphics card. It shows the model circuitRF built, and after a run it also shows what the solver made. It is the 3D editor's own view with nothing to edit — the same toolbar, tree, keys and menus, without the drawing tools, the setups, Simulate, the unit and the snap step — and its status lines under the view (the setup's refusals and notes, the mesh, the field) are always shown.

• Model. Every solid, coloured by its layer, with the dielectrics, the air and the air box's faces each switchable. The faces are coloured by boundary kind: grey metal, blue absorbing, orange PMC, violet symmetry. The object tree lists every solid, and a click in the view picks one, names it and brings its row into sight; a row's tick hides that solid in this view only. The tree is headed Objects and lists the objects by material (the default — objects with no material come first, in their own group) or by type (boxes, sheets, cylinders, …), chosen in its header. By material a row does not repeat the material its group is headed by; an object's construction order is its name's tooltip. The air box appears — drawn and listed — once there is a solid or a sheet to size it: with no setup yet it is the box a new setup would solve in. By material it is listed under the material that fills it, first under Boxes by type, and says which setup it belongs to. It is filled with Air unless you choose otherwise: select it and pick its Material in the Properties panel (Vacuum for free space, or any material of the technology). Its tick hides it — faces and outline — and it stays hidden until you show it again. The filter at the header's right hides tree rows by type or by material, and its icon changes while it hides anything; it never hides an object in the view.

• Camera. Middle-drag or Alt-drag to orbit; Shift-drag or Ctrl+middle-drag to pan; right-drag, Shift+middle-drag, scroll or pinch to zoom; a left drag selects with a rectangle. With Settings ▸ General ▸ Use Hertzon interactivity turned off, a left drag orbits and right-, middle-, Alt- or Shift-drag pans. With no wheel and no middle button, the keyboard does all three: Ctrl/Cmd+= and Ctrl/Cmd+− zoom, the arrow keys pan and Shift+arrow orbits 15°. Home, Ctrl/Cmd+0 and View ▸ Zoom to Fit fit what is shown, keeping the direction you are looking from; Ctrl/Cmd+E fits the selection, or everything shown when nothing is selected. 1 is isometric, 2–7 are the six orthographic views, and P switches between perspective and orthographic. Double-click the axis indicator for the same views: an axis looks straight down it — Z the top view, Y the front, X the right — and again turns to the opposite side; anywhere else inside its ring is isometric. While you orbit or pan nothing is highlighted and nothing snaps. These are the 3D editor's keys too: every 3D pane uses the same ones.

• Selecting. O, F, E and V choose what a click selects: a whole object, one face, one edge, or one vertex (the toolbar has a button for each). What is under the cursor is highlighted. A click selects it; Shift-click adds to or removes from the selection; Esc clears it. A selected face's area and normal, an edge's name, kind and length, or a vertex's coordinates, are shown under the view, in the layout's unit. B steps to the next thing behind the selection along the line of sight through the cursor, and Shift+B steps back toward you — the status line says where you are (Face zmin · "trace" · 2 of 5). A face selected behind others is drawn through what is in front of it. Right-click opens the menu for what is selected: Hide, Isolate, Show All, and Select Owning Object. Nothing here changes the setup.

• The 3D editor. A 3D view (.c3d) opens in this same pane with drawing added: boxes, cylinders, sheets, faces pushed and pulled, placed cells, bond wires, ports, setups and the air box, all with these keys. It is circuitRF's own and needs no solver. The 3D Editor is its chapter, and describes every control.

• Measure (M, the ruler button, 3D ▸ Measure; in the editor and in this view). Click two points, snapped as drawing is; after the first a line follows the cursor. A card in the corner of the view gives both points' x, y and z, their differences (the second minus the first) and the distance, in the document's unit — changing the unit re-spells them. Every number can be selected and copied; hover one for its copy button, or Copy All for the whole card as a table that pastes into a spreadsheet. A copied value carries its unit (12.5mil) and pastes back into any dimension box as exactly the same length. ≈ marks a point that is not exactly on the database-unit grid. A third click starts a new measurement; Esc ends it. Measuring changes nothing and adds no undo step.

• Clip plane (C). A plane along an axis or the view direction, dragged through the model. It shows inside a package, under a lid, or through a via's clearance.

• Mesh. The mesh Gmsh made, once one is on disk for the setup, with or without fields: the boundary triangles, or the tetrahedra on the mesh's own plane. Looking at a 3D mesh before solving makes one with no solve.

• Grid. For an openEMS setup, the FDTD grid on the clip plane and where it meets the metal. A cylindrical grid is drawn as it is: on a clip plane across its axis, a circle at each radius and a spoke at each angle; on a plane along it, the radii as lines along the axis; on the metal, where its circles, spokes and planes cross it.

• Field plots. After a run that saved them, a field is drawn by a field plot: a row of the Object Tree's Field Plots group, saved with the 3D view like any object, so a plot is still there when the file is closed and opened again. Add one with the + on the group's header (or right-click the group, New Field Plot…); it is drawn at once and its fields open in the Properties Inspector:

  • Setup — the setup whose run it reads. A new plot keeps the setup that was active when it was made, so switching setups never changes what an existing plot shows; Active follows whichever is active.
  • Solution — what the run saved: a frequency (driven), a mode (eigenmode), a terminal (static), a sweep point (thermal). The plot keeps the value, not its place in the list, so a later run that saves one more frequency never moves it to another. On Palace, a port with two terminals is solved one mode of its face at a time, not one terminal, so its choices read 5 GHz, Left mode 2 driven: the field of the face's second mode launched, which is not terminal 2 driven alone. An openEMS run drives each terminal, and its choices name the port. The list ends with Other frequency… (needs a re-run), which adds a frequency to the setup's Save fields at and moves the plot to it. Until the setup has run, the list holds one greyed row saying what will appear, such as Terminals of the electrostatic setup, once it has run.
  • Quantity — only those in the solver's files are offered: the electric field |E|, the surface current J_s on the conductors, and for a static solve the potential. Until the setup has run, the quantities its run will offer are listed greyed.
  • On — a clip plane (the plot's own axis and position, so two plots can cut in two places), surfaces (the solid selected in the tree, or the conductors for J_s), or faces (below). Beside a clip plane's position is a slider across the geometry: the cut is drawn as you drag it and kept, as one undo step, when you let go. The field is saved for the whole volume, so the plane can be moved at any time, before or after Simulate, without running again. While the plot is selected the plane is drawn in the 3D view as a translucent rectangle across the air box, labelled with its axis and position, so it can be placed before anything is solved; Always show the plane keeps it drawn when the plot is not selected. Through… puts the plane through an object: press it, then click the object (a pad, a via) in the 3D view, and the plane moves to the object's centre on the plot's axis. Esc cancels. The plane is the plot's own. The toolbar's section plane neither moves with it nor hides it, so cut the geometry wherever helps you see. Section Through… (the Field Plots menu, or the 3D view's menu once a run has fields) makes a plot on a plane no axis gives: click two points (an object's centre, a vertex and an edge snap, so a pad's centre and a via's axis are one click each) and the plot is cut on the vertical plane through both, across the whole height. The toolbar's clip plane is set to the same plane (its axis reads Section), keeping the side away from you, and the view turns to face it square on. One undo takes all three back. In Properties the plot shows its two points; Use an Axis returns it to an axis plane.
  • dB and a range percentile, which keep one singular edge from washing out the picture. A field strength in dB reads in dBµV/m (H and surface current in dBµA/m), the EMC convention; any other quantity reads in dB re 1 of its unit. A power or energy density (the Poynting vector, the energy densities) is 10·log₁₀ of the value, and a field strength is 20·log₁₀.
  • Drive — for a driven solution, two more rows set the drive the field is shown at. Drive power is the power, in W, mW or dBm, set per plot and saved with it. Referred to chooses Incident, the power a source matched to the port's Z₀ makes available, or Accepted, the power that enters the port after reflection, with every other port terminated in its own Z₀. Neither row is shown for an eigenmode, static or thermal solution.

  Up to four plots are drawn at once: tick them in the tree, so two slices of one quantity at two positions, or two quantities on crossing planes, are on screen together. Ticking a fifth is refused, and the status line names the four drawn so you can untick one. A new plot made while four are drawn is added hidden; Show all ticks the first four in the list and says which stay hidden. Plots of the same quantity at the same solution, in the same dB and range percentile, share one colour range, taken over all of their triangles together, so their colours compare. Each such group has one legend, titled with every plot in it. The legends stack down the right of the view, all one width, and a legend that would run past the bottom is left off with "+N more" under the last one. Where two plots cover the same place, the one selected in the tree is drawn on top, and the value under the cursor is the front plot's, named. Each change is one undo step. Adding, editing or hiding a plot never marks a result stale. A plot whose data is missing — no run of its setup yet, a run that did not save its frequency, a quantity the run no longer offers — stays in the tree with a warning mark and draws nothing; its tooltip, and the Inspector, say which, and name what the run did save. It is never moved to the nearest frequency for you. In a setup's own 3D view (Show 3D View from a .cem) plots last for the session: that view has no document to keep them in.

  What drive a field is shown at. Each driven solution is the field of one port driven, with every other port terminated in its own Z₀. The solve is linear, so the drive only rescales what is shown: field strengths by the square root of the power ratio, power and energy densities by the ratio. Changing it never needs a re-run and never marks a result stale. The default is the power the solver's own field is at, 0.5 W incident: Palace drives each port at unit power on peak phasors, which is 0.5 W time-averaged. Magnitudes are peak; the RMS value is 1/√2 of the peak. Under an automatic range the colours do not change with the drive, but the legend's numbers and the value under the cursor do, and the legend states the drive, the port and its Z₀. Accepted needs the run's record of the port's reflection at the plot's frequency; where the run kept none, the plot is refused and the Inspector says why, and Incident still works. openEMS fields are shown at the same drive as Palace's, so the two solvers' plots of one structure compare directly. An openEMS run that kept no record of its port draws in relative values, without a unit, and its legend says so. Driving several ports at once is not offered.

  Without a window, a .c3d's clip-plane plot is drawn as the section it cuts: circuitrf render cavity.c3d -o cut.png --field Field1 (add --phase 90 for an instantaneous quantity), and circuitrf render cavity.c3d --list-fields lists the plots and whether each one's data is there. A surfaces or faces plot is drawn from a direction you name, since the window's camera is not saved with it: --iso, or --view-dir top (front, right … or x,y,z), as a .png. It reads the same run, solution and range the 3D view does, and refuses a plot whose data is missing with the same sentence. See render.

• One face. In the 3D editor, right-click a face, Plot Field, to add that face to the plot being drawn when it is a faces plot — the top of one trace, one face of a substrate — or to start a new faces plot with it; again to take it off, and several faces accumulate. A conductor has no inside to read, so its face shows the field in the material next to it (or J_s, when that is the quantity chosen). A sheet asks Top side or Bottom side: the field's normal part jumps across a sheet that carries charge, so the two sides are two different pictures. The quantity is never changed for you — a face that cannot show the chosen one says why under the view. On an openEMS run the face is read half a grid cell off the metal, where openEMS records the field.

• Animation. With the plot drawn selected, the play button at the foot of the Inspector sweeps the phase through one cycle and draws the instantaneous field Re{E·e^jφ}. That is the standing wave in a cavity, or the current running along a wire. The phase and its speed belong to the view, not the plot: they are not saved. There is one phase for every plot drawn, so two animated plots move in step, and moving φ under one plot moves the others. For two quantities at one frequency that shows their true relative phase. Two plots at different frequencies each run one cycle of their own per loop, which compares them phase for phase rather than in real time; their legends say so.

• Pictures. Export picture … saves a PNG of the view at a multiple of the window's size, with or without the legends (the same stack as the view's) and caption. Right-click in the view — on anything or nothing — for Copy, which puts the view on the clipboard at four times the window's size (the 3D editor's canvas offers it too).

What the mesh tells you about trusting the answer. Look at where the small elements are. After refinement they gather where Palace's own error estimate said the answer needed them: along the edges of strips, around a wire, in a port sheet. If the refined region is on the metal you care about, the refinement did its work. If a feature you care about sits among large elements, the answer at that feature is the coarse mesh's, and a Standard run (with refinement) is the one to trust. Palace also saves its error indicator per element, which the view offers as a field: it shows the error directly, where the mesh shows it only by implication. On an openEMS setup, the grid is the whole accuracy story. A wire much thinner than the cells around it has too much inductance, and the run's notes say so.

How far the dielectric reaches

A stackup gives each dielectric a thickness and a material but no outline, so circuitRF decides how far it reaches sideways. The answer depends on the solver, and for the 3D solvers it is part of what gets solved.

• Planar (MoM) setups. The planar solver treats every dielectric as laterally infinite: a board that goes on forever under the metal. Show 3D on a planar setup has to draw the substrate some size, so it draws each dielectric to the board outline, or, if the layout has none, to the box round what the layout draws. That shape is for viewing only. The planar result is the same whatever the picture shows.
• 3D setups (a .cem run in Palace or openEMS). The dielectric reaches the air box, so a board with no drawn edge reaches the absorbing boundary as though it went on forever. A layout that draws a board outline is cut to that outline instead.
• A layout placed in a 3D view (a .c3d). There is no air box around the layout, so every dielectric is finite, and its sideways shape is part of the problem Palace and openEMS solve. Every slab is cut to the layout's board outline (closed shapes on the technology's outline layer: the layer exported as Edge.Cuts, or a Gerber profile), or, with no outline, to one rectangle round everything the layout draws, the same for every slab. Nothing reaches past the artwork, and a placed layout brings no metal it does not draw: a ground plane that a planar setup takes from the stackup must be drawn. The Messages note for the run says which bound the slabs took.

Draw a board outline whenever the board is bigger than its copper, for example when a trace's fringing field runs through substrate beside it, for radiation from the board edge, or for coupling across it. How a placed layout comes into a 3D view is in Drawing in 3D ▸ A placed layout's dielectric.

Walking through the example

File ▸ New ▸ Examples ▸ 3D EM copies a workspace with three cells to a folder you choose and opens it. Each cell is something a planar solver cannot do, or can do only as a comparison. Every setup runs in minutes on a 16 GB laptop. Its README.md opens first and carries the same numbers as this section. The numbers also live in expected-numbers.json beside it, which circuitRF's own tests re-run against, so the two cannot drift apart.

Every time and memory figure below was measured once, on an Apple M4 with 10 cores and 16 GB, running macOS 27.0, Palace 0.18.1 on 10 processes, and Gmsh 4.15.2. Times are wall clock for the whole run. Memory is Palace's own peak over all its processes.

Without Palace, start anyway. Open any setup and press Show 3D to see the model. Press Simulate, and the refusal offering Install Palace … is the expected first step.

Bond wire

A 1 mil (25.4 µm) gold wire with a hexagonal section and a wedge foot at each end, between two 100 µm pads on 100 µm of alumina, inside a closed metal box 3 × 2 × 1.1 mm. Its loop is not symmetric: it rises steeply from the left pad to 290 µm above the pads about a third of the way across, then falls in a long slope to the right pad. A port sheet runs from each pad's outer edge down to the ground.

Bond wire 3D.cem runs Palace at Standard. (Measured on the wire's earlier, symmetric 150 µm loop; the Palace figures in this section have not been re-run for the loop above. Kernel W's have.) It takes 179 s, with a peak of 3.1 GB, and gives |S21| = −0.765 dB at 10 GHz and a series inductance of 906 pH at 1 GHz. (The inductance is the π-model's series term, −Im(1/Y21)/ω, from the S-parameters.)

• Draft is not enough here. It takes 5 s, but reads 813 pH and −0.514 dB. The starting mesh around a 25 µm wire is exactly what refinement exists to fix. A wire is small metal in a big box, so it is the case where the preset matters most.
• Accurate (three refinement passes and tighter tolerances) took 8.5 min at 7.0 GB and reads 908 pH and −0.772 dB: within 2 pH and 0.01 dB of Standard, for nearly three times the time.
• Kernel W, the wire kernel in wBond, reads 768 pH and −0.800 dB from the same .wBond (open it and Export Touchstone … with the Distributed model). The difference between the two answers is where the terminals are, not a disagreement between the solvers. Kernel W's terminals are the wire's own ends. Palace's ports are at the pads' outer edges, so its answer also contains the pad between each port and the wire, and the 100 µm drop from the pad to ground.

Via through a plane

A 400 µm microstrip on top of a four-layer board, a via through a 900 µm clearance in the middle ground plane, and an inverted microstrip underneath leaving the other way. The laminate has εr 3.66 and tanδ 0.004, and the metal is copper. Both lines return through the same plane, one from above and one from below.

First, the planar solver. Via planar.cem is the same structure for the planar solver, and it is refused:

This EM setup names 'Ground Plane' as its return plane, but its top surface is at 270 µm, which is NOT below the lowest analysis level 'Bottom Copper' at 0 µm. A return plane must lie BENEATH the conductor it feeds, or there is no dielectric slab between them to solve on. Name a conductor below 0 µm, or restrict this setup's analysis levels to conductors above 'Ground Plane'.

circuitrf check reports the same sentence as an error on that setup. It is the only error in the workspace, and it is there on purpose.

Then Palace. Via 3D.cem runs at Draft over 0.1–20 GHz, with 1.5 mm of air on every side of the board and absorbing walls. It takes 62 s, with a peak of 4.8 GB, and gives |S21| = −0.228 dB at 10 GHz and −0.430 dB at 20 GHz.

• What Draft trades away is phase. Against an independent hand-built Palace model of the same via (element order 2, 155,883 tetrahedra), Draft's |S21| is within 0.1 dB, but its phase is 12° off at 10 GHz and 23° off at 20 GHz.
• Element order 2, with Draft's other values ("ElementOrder": 2 in the Palace section), takes 6.5 min at 9.2 GB and lands within 0.02 dB and 5° of that model.
• Standard took 35 min at 9.3 GB on this structure, when the presets were measured. Accurate has never finished on it: it was stopped after 23 minutes in its third solve.

Package

A ceramic package with its lid on. The 8 × 8 mm cavity has a 254 µm alumina base on a metal floor. Two 200 µm gold leads stop 100 µm short of the side walls, and a 1 mil bond wire runs from each lead to a die pad inside. The lid and seal ring are the closed metal box around it: side walls 100 µm beyond the leads, and a lid 700 µm above the wires, 1.13 mm above the floor. The die itself is not in the model.

• Capacitance (Package C.cem, electrostatic, Standard): 8 s, 1.1 GB. Each lead with its wire and die pad has 586 fF to ground, and the two leads couple through 0.74 fF.
• Inductance (Package L.cem, magnetostatic, Standard): 22 s, 1.8 GB. Each path has 1.98 nH, and the mutual inductance is 0.032 nH. It runs on Package shorted.clay, a copy of the layout with a via from each die pad to the floor. Shorting the far end gives each terminal's current a loop, which is the ordinary way to read a series inductance. This is external inductance, the RF value: conductors carry current on their surfaces only.
• Draft is not enough for either. It runs in 2–3 s but reads 831 fF (42 % high) and 1.66 nH (16 % low). A static solve has no frequency to size its elements by, so its first mesh is coarse, and refinement is what finds the field around the metal.
• The lid's modes (Package lid modes.cem, eigenmode, the three modes above 5 GHz, at Draft with element order 2): 166 s, 1.7 GB. Mode 3, at 23.65 GHz, is the lid's first cavity mode, with 94 % of its energy in the air under the lid and a Q of 1,662. That Q counts the gold's loss and the two ports' 50 Ω as loads. The estimate for a cavity much thinner than a wavelength, with the base and the air in series under the lid, puts the empty cavity at 23.7 GHz. Modes 1 and 2, at 14.3 GHz, are not the lid: they are the leads ringing as lines into their ports, with a Q below 1. For a band that reaches 20 GHz, the lid is clear, and a cavity 20 % larger each way would bring its mode to about 19.7 GHz, inside the band.
• The eigenmode preset. Draft takes 13 s and finds the lid mode at 23.39 GHz, but reads its Q as 176, ten times too low: order 1 cannot resolve the loss. Element order 2 fixes that without a refinement pass. Standard adds refinement passes to Palace's nonlinear eigenvalue solve, and was stopped after 28 minutes in its last pass.

What the example does not show. Every setup is Palace's. The via and the bond wire can also be run with FEM & FDTD - Compare, to see how far openEMS lands from Palace on the same model. That is a cross-check, not a reference.