Simulation economics
Cut the simulation cost
of superconducting design.
Most of a superconducting design budget is spent re-solving the same physics at thousands of points — every pad geometry in a cQED layout, every bias point in an SFQ cell. A learned surrogate solves that physics once, at training time, then answers the whole sweep at millisecond latency, reserving your solver for the candidates that deserve it.
Why simulation gets expensive
A single high-fidelity solve is affordable. The cost explodes because design questions are sweeps, not solves: every pad geometry that lands a qubit on its target frequency, every bias point that keeps a logic cell inside its margins, every junction spread a fab run might ship with.
At minutes per solve and thousands of points per question, teams end up rationing their own engineering curiosity — exploring less of the design space than they should because compute is metered in cluster-hours.
A worked example · SFQ margin sweep, measured
Live models, published validation
Cheap only counts if the answers hold up.
Every model ships with a validity report: held-out validation against its reference solver, the classical baseline it had to beat, a conformal band where one is fitted — and abstention in the regions the reference physics itself cannot label.
RSFQ screening — Fluxus
Margin, timing, yield & design-rule screening for superconducting logic (MIT-LL SFQ5ee+)
When a surrogate is the wrong tool
If you need one certification-grade answer, run the full simulator — that is what it is for. Surrogates also stay inside the input ranges they were validated on; they will not extrapolate to physics they were never trained against.
The economic case is exploration: screen the whole design space for the cost of a lunch, then spend your solver budget on the handful of candidates that survive.
Where the savings come from
Training happens once, on physics generated by the reference solver. Serving a trained model costs milliseconds of CPU time per design — so the screening loop that used to be metered in cluster-hours becomes effectively free, and your solver budget concentrates on the candidates that matter.
FAQ
Simulation cost, answered plainly.
Why are physics simulations so expensive?
A single high-fidelity solve (electromagnetic eigenmode, transient circuit, FEM) costs minutes to hours of CPU time, and real design questions are never a single solve — they are sweeps: every pad geometry, every bias point, every junction-spread variation a fab run might ship with. Multiply solver time by thousands of design points and you get cluster bills and week-long queues.
What is a simulation surrogate?
A surrogate is a model fitted to many runs of a reference simulator — an electromagnetic solver for qubit chips, or JoSIM for SFQ logic — until it reproduces that solver's outputs across a validated range of inputs. The model family is whatever measures best: on RSFQ margins the served model is gradient-boosted trees; the resonator and TransmonCross engines are log–log ridge regressions on HFSS data. Inference takes milliseconds instead of seconds to hours, which is what makes large sweeps cheap. Our own models are trained on simulator output, not on measured hardware; a per-process model is trained on your reference simulator's output, under your agreement.
How accurate are surrogate models?
Accuracy is per-model and must be measured on held-out cases, not claimed. Cogitan publishes a validity report with every model, including the classical baseline it was measured against — and we publish that comparison whether or not it flatters us. On RSFQ margins a gradient-boosted-tree baseline beat our neural surrogate on every schema-derivable head, so on 2026-08-29 the trees became the served model and the neural surrogate was deleted. Inputs outside a model's validated envelope are refused rather than answered; served SFQ margins carry a split-conformal band, and ambiguous cases are deferred to JoSIM rather than answered.
How do engagements work?
A fixed-scope feasibility study first: we take your simulator, build the labeled dataset, benchmark a surrogate against a strong classical baseline, and deliver the validity report — you see the evidence before committing. If the numbers justify it, the surrogate ships as a managed endpoint your team calls at millisecond latency, with retraining as your process or design library evolves.
When is a surrogate the wrong tool?
When you need one high-stakes, certification-grade solve, run the full simulator. Surrogates are also restricted to the input ranges they were validated on — they will not extrapolate to novel physics. The economic case is exploration: screening a large design space cheaply, then spending your solver budget on the handful of candidates that matter. Our feasibility study will tell you honestly if a classical model or your existing solver is the better answer.
Do I need my own GPUs or solver licenses?
No. The delivered surrogate runs as a managed endpoint on Cogitan infrastructure (or on-prem where the data can't leave), and screening results come back at interactive latency. Your solver seats get reserved for the shortlisted candidates that deserve a full-fidelity run.
See it run, on captured output from the real models.
demo.cogitan.ai