[IHP] Multiple bugs in mesh and simulation pipeline

[delficomerso]

Description

While generating meshes and running simulations to obtain S-parameters for different IHP components from gdsfactory/IHP (specifically IHP/ihp/cells), we encountered several bugs.

Reproducible notebooks containing the meshes, errors, and preliminary S-parameter calculations can be found in the s_parameters branch of EpsilonForge/gsim under nbs/IHP_components/.

---

Bug 1: Mesh fails for cmim and rfcmim with planar_conductors=False

Calling:

sim.mesh(preset="graded", margin=1.0, refined_mesh_size=0.2, planar_conductors=False)

fails for the cmim and rfcmim components. The workaround was to use planar_conductors=True, which allows the mesh to complete, but at the cost of losing information about the real complex layered structure of these components.

References:

• nbs/IHP_components/mesh_cmim.ipynb
• nbs/IHP_components/mesh_rfcmim.ipynb

---

Bug 2: Unexpected dimensions in mesh output (Z or XY axes)

In some components, the reported mesh dimensions are larger than expected — either along the Z axis or in XY. Examples:

• cmim: Dimensions: 9.9 x 9.9 x 23.3 µm — see mesh_cmim.ipynb
• cmom: Dimensions: 2.7 x 8.1 x 17.8 µm — see mesh_cmom.ipynb
• inductor2 and inductor3: XY dimension issue, partially mitigated by setting margin=1.0 — see mesh_inductor2.ipynb and mesh_inductor3.ipynb

The margin=1.0 workaround does not fully resolve the root cause for all components.

---

Bug 3: Simulation fails without margin=1.0 on inductor2 and inductor3

For inductor2 and inductor3, the simulation fails to run unless margin=1.0 is set. We suspect this is related to port positioning — without enough margin, the ports may not have sufficient space to be correctly placed.

A potential fix could be adding an offset parameter to add_port, similar to the existing behavior in add_cpw_ports.

---

Bug 4: Disconnected mesh on inductor3

The mesh generated for inductor3 appears visually disconnected — the conductor segments look separated rather than forming a continuous structure, as seen in the image below. This suggests the meshing is not correctly joining the geometry across the component.

For the placement — I'd edit the original issue and add it as Bug 5 there, since it's a new bug report rather than an update on an existing one. New bugs belong in the issue body alongside the others for visibility, while comments are better for progress updates on existing bugs.

Here's the text to add as Bug 5:

---

Bug 5: Mesh fails for rhigh, rppd, and rsil

Same class of issue as Bug 1 — marker and process layers being extruded as 3D volumes cause Gmsh to fail during boolean fragmentation.

rhigh

Removing HeatResdrawing, GatPolydrawing, and Contdrawing is necessary, otherwise mesh generation produces:

Conductor volume 1 on poly invalidated by dedup
Conductor volume 2 on poly invalidated by dedup
Conductor volume 4 on heatres invalidated by dedup
Error   : Invalid boundary mesh (overlapping facets) on surface 29 surface 93

rsil

Same error with different surface indices:

Conductor volume 1 on poly invalidated by dedup
Conductor volume 2 on poly invalidated by dedup
Conductor volume 4 on heatres invalidated by dedup
Error   : Invalid boundary mesh (overlapping facets) on surface 36 surface 93

As with Bug 1, the manual layer filtering workaround allows the mesh to complete, but produces lacking meshes with inflated XY dimensions (related to Bug 2), for example for rhigh:

References:

• nbs/IHP_components/mesh_rhigh.ipynb
• nbs/IHP_components/mesh_rppd.ipynb
• nbs/IHP_components/mesh_rsil.ipynb

[vvahidd]

Volker has documented some simulations here, you can compare s-params against
https://github.com/VolkerMuehlhaus/gds2palace_ihp_sg13g2/blob/main/doc/IHP_Palace_Workflow_evaluation_October%202025.pdf

[delficomerso]

---

Update on Bug 1: Root cause identified and partial workaround found

We investigated further and found that meshing with planar_conductors=False fails for both cmim and cmom caused by marker and nofill layers being extruded as 3D volumes, creating degenerate shared faces with the real conductor layers in Gmsh.

cmim

• Vmimdrawing (129/0) shares z-faces with adjacent conductor layers, causing Gmsh's boolean fragmentation to fail
• Nofill layers (Metal4nofill, Metal5nofill, TopMetal1nofill, TopMetal2nofill) and MemCapdrawing with overlapping footprints trigger the same collision

Error:

Exception: Invalid boundary mesh (overlapping facets) on surface 48 surface 18

Workaround:

keep_layers = [
    (67, 0),   # Metal5drawing
    (67, 2),   # Metal5pin
    (36, 0),   # MIMdrawing
    # (129, 0),  # Vmimdrawing — excluded, causes overlapping facets
    (69, 0),   # MemCapdrawing — harmless, no stack entry
    (126, 0),  # TopMetal1drawing
    (126, 2),  # TopMetal1pin
]
cc = c.extract(layers=keep_layers)
cc.add_ports(c.ports)

cmom

Via layers cause a different Gmsh error — same root cause, via z-extents share faces with adjacent metal layers.

Error:

Exception: PLC Error:  A segment and a facet intersect at point

Workaround:

keep_layers = [
    (8, 0),    # Metal1drawing
    (10, 0),   # Metal2drawing
    (30, 0),   # Metal3drawing
    (30, 2),   # Metal3pin
    # (19, 0),  # Via1drawing — excluded, causes PLC intersection
    # (29, 0),  # Via2drawing — excluded, causes PLC intersection
    (69, 0),   # MemCapdrawing — harmless, no stack entry
]
cc = c.extract(layers=keep_layers)
cc.add_ports(c.ports)

With these extractions, planar_conductors=False meshes successfully for both components. However this remains a manual workaround — ideally users should not need to know which layers conflict in order to generate a valid mesh.

Remaining issue (related to Bug 2)

Even after the layer extraction workaround, the mesh dimensions are still inflated:

• cmim: Dimensions: 107.9 x 107.9 x 20.3 µm for a nominally 9.9 x 9.9 µm component
• cmom: Dimensions: 101.7 x 107.1 x 17.8 µm

Preliminary S-parameter results

We ran some preliminary S-parameter simulations using the workarounds above. However we are not yet satisfied with either the mesh quality or the S-parameter results — the inflated domain dimensions and the manual layer filtering make it hard to trust the outputs at this stage. Below are some representative plots:

[delficomerso]

Update on Bugs 3 & 4

Bug 3: Port geometry for inductor2 and inductor3

We found meshing no longer requires margin≠0:

sim.mesh(preset="graded", refined_mesh_size=1.0)

However the mesh dimensions remain inflated (related to Bug 2):

Dimensions: 129.4 x 159.4 x 30.3 µm

The more fundamental open question is which port geometry is correct for this component. The current options we explored are:

# Option A: inplane (current approach)
sim.add_port("P1", layer="topmetal2", geometry="inplane", excited=True)
sim.add_port("P2", layer="topmetal2", geometry="inplane", excited=False)

# Option B: inplane with offset
sim.add_port("P1", layer="topmetal2", geometry="inplane", excited=True, offset=2.0)
sim.add_port("P2", layer="topmetal2", geometry="inplane", excited=False, offset=2.0)

# Option C: via (following the pattern of the non-IHP inductor notebook)
sim.add_port("P1", from_layer="INDdrawing", to_layer="topmetal2", geometry="via", excited=True)
sim.add_port("P2", from_layer="INDdrawing", to_layer="topmetal2", geometry="via", excited=False)

Looking at the non-IHP inductor component in mesh_inductor.ipynb, the ports are via — but that works because layer_inductor is a physical metal layer in that PDK. In the IHP inductor2, INDdrawing is a process marker with no physical conductor in the stack, so Option C is not viable without adding INDdrawing as a conductor layer, which would be physically incorrect.

We ran preliminary S-parameters using Option A (inplane) but are not satisfied with the results:

Bug 4: Disconnected mesh on inductor3

We can now see the vias geometry appearing in the mesh for inductor3