A multi-physics simulator for post-silicon and 3D-stacked transistors. It ranks 2D channel materials, device geometries and stacking choices on speed, heat and manufacturability together, before anything goes near a fab.
Silicon transistors are running out of room to shrink. The leading replacements are atomically thin "2D" semiconductors such as MoS₂ and WSe₂, stacked in new geometries: gate-all-around nanosheets, CFETs that stack the n- and p-type transistors on top of each other, and fully monolithic 3D chips.
In June 2026 imec, ASML and TSMC demonstrated 2D-material CMOS on 300 mm wafers. It confirmed an awkward fact: MoS₂ makes the best n-type channel and WSe₂/WS₂ the best p-type, and no single material does both well. So the useful question isn't "which material is fastest?" It's "which combination of materials, dielectrics and stacking is fast enough, stays cool enough, and can actually be manufactured?"
Answering that means joining physics that usually lives in separate, expensive tools. Quantum Foundry joins it in one open pipeline.
Five stages, each behind a common interface with two interchangeable backends: a fast "stub" for exploring, and a real physics solver for answers. Switching between them is a config change, not a code change.
Kwant for atomistic quantum transport, DEVSIM for classic drift-diffusion, HotSpot as a fast thermal pre-filter before full 3D FEA. All open source, so every result can be reproduced.
No mature open-source yield tool exists, so yield is modelled directly with Monte Carlo and Poisson/Murphy models. Every candidate is checked against back-end-of-line thermal budgets.
CFET, GAA and monolithic-3D devices run per-tier, with multi-tier heat flow including the thermal resistance at each bonded interface.
That rule caught real mistakes. Three of them:
Power density was normalised by the wrong device dimension, which scaled total power by L/W. For the default 20 × 100 nm device, every junction temperature, including real HotSpot runs, was fed a fifth of the real power. Caught by checking the formula's units, then fixed in both thermal backends.
A drain-induced barrier lowering value looked like real short-channel physics. Reading the model showed its potential has no drain-voltage dependence, so the number was a by-product of the method. It is now flagged in every output instead of being shown as physics.
MoS₂'s thermal conductivity was seeded as 34.5 W/m·K, "from Yan et al. 2014". The paper's own abstract says 24 ± 4 W/m·K. The higher number circulates in later papers that cite it. The value was corrected to the source.
The n/p channel search runs every seeded 2D material in both roles instead of assuming MoS₂ is n-type and WSe₂ is p-type. On the current physics, MoS₂, WSe₂, WS₂, MoTe₂ and WTe₂ all land on the Pareto front for both roles. That's a genuine trade-off, so the tool declines to name a single winner.
Graphene and black phosphorus are excluded, with the reason printed: their effective masses aren't in the cited data yet. Leaving them out openly is the correct behaviour.
These rankings come from stub-level transport and thermal models, labelled as such. They show that the search machinery works, not what the final answer is.
$ qf heterostructure-search
NOTE: Tj values below come from thermal/stub.py's
lumped-RC model, not the real HotSpot backend …
=== n-type role ===
MoS2_monolayer: Ion=240 uA/um … [pareto-front]
WSe2_monolayer: Ion=437 uA/um … [pareto-front]
WS2_monolayer: Ion=471 uA/um … [pareto-front]
MoTe2_monolayer: Ion=259 uA/um … [pareto-front]
WTe2_monolayer: Ion=272 uA/um … [pareto-front]
black_phosphorus_monolayer: EXCLUDED — …
refusing to fabricate.
graphene_monolayer: EXCLUDED — …
-> no single recommendation: 5 candidate(s)
on the pareto front, a real trade-off
rather than a clear winner
| Layer | State | Detail |
|---|---|---|
| Materials database | real | 10 cited 2D materials and gate dielectrics |
| Validation | real | 6 literature benchmarks, each tied to a specific figure or table |
| Quantum transport | real | Kwant NEGF |
| Drift-diffusion | next | DEVSIM backend scaffolded; needs a solver environment |
| Thermal | real | HotSpot compact model incl. multi-tier stacks; 3D meshing via gmsh |
| 3D FEA | next | FEniCSx/Elmer on the generated mesh |
| Yield | real | Monte Carlo + Poisson/Murphy/negative-binomial |
| Ab-initio | next | Quantum ESPRESSO + Wannier90, designed |
The remaining solvers need real compute, and the full parameter sweeps approach HPC scale. That's the next stage of the project, and a good place to contribute.
pip install -e ".[dev]" qf run MoS2_monolayer --backend stub qf heterostructure-search --format markdown
The stub pipeline installs in a minute with no heavy solvers. Real backends install through one conda setup script, and qf preflight reports exactly what's available before any real run.
Contributions are welcome, especially cited materials data, validation benchmarks and new solver backends. See CONTRIBUTING and the roadmap.