← All impact studiesClient · RF Simulation Automation
Millibeam

Product ID in. Optimized design and datasheet out.

An engineer-day of expert Ansys HFSS work per customer configuration became 1.5 hours of unattended agentic work — optimized, verified, documented, and replayable after the fact, with nobody at the keyboard.

CLIENT
Millibeam
SECTOR
RF Simulation Automation
INDUSTRY
Semiconductor Packaging & RF Hardware

Millibeam sells configurable antenna-in-package modules. The Heaviside line spans WiFi 6, WiFi 7, and 5G AiP parts, and every customer order arrives with its own band, gain, and matching requirements — so every configuration has to be verified, optimized, and documented in Ansys HFSS before it ships. That was an engineer-day of expert solver work per part: loading the catalog model, setting targets from the customer spec, optimizing within manufacturer-approved bounds, verifying the winner, and writing the datasheet by hand.

Published with Millibeam's permission. The workflow, scope, figures, and recorded run in this study are verified by Divergent Physics, 2026.

The Millibeam Heaviside antenna-in-package product catalog — WiFi 6, WiFi 7, and 5G module variants

One catalog, thousands of customer-specific variants: the Heaviside line spans WiFi 6, WiFi 7, and 5G AiP modules, and each order arrives with its own band, gain, and matching targets.

Outcomes
1.5 h
of unattended agentic work, end to end
replacing a full engineer-day per configuration — load → baseline → optimize → datasheet
20+
full-wave candidates explored per run
overnight-scale exploration, on demand
0
human touches in the loop
the solver scores every candidate against the spec

The same skill runs any product in the catalog: new part, same process.

Watch the Run

Watch the run, start to finish

The complete pilot run in one recording — catalog load, baseline solve, the optimization loop, verification, and the generated datasheet. Unattended from the first frame to the last.

The full Millibeam AiP workflow, one unattended run. Published with Millibeam's permission.

The Problem

Configuration work is the kind of expert labor that does not compound. Each part needed the same full-wave cycle as the last one, done by someone senior enough to judge whether the result was defensible, and the datasheet at the end was transcribed by hand from the solved model. The cost was an engineer-day per configuration, and the constraint was that it scaled only by adding engineers who could do it — which is exactly the constraint a growing catalog runs into first.

What We Built

We encoded the team's own process once, as a reviewable skill document the agent follows — data, not code, so their engineers can read, audit, and edit every step of it. One unattended run does the baseline solve, checks KPIs against the customer spec, runs Bayesian optimization over manufacturer-approved parameters only, re-solves the winning candidate to verify it, and generates the customer-facing datasheet — patterns, gain, matching, efficiency — with no hand-editing. Every solve, variable change, and decision is logged, and each run keeps its request, decision log, and working solver project on disk, replayable after the fact.

Sequence diagram: Bayesian optimizer, agent, and HFSS working the optimization loop

Inside the loop: the optimizer proposes parameters, the agent updates the model and solves in HFSS, scores against spec, repeats — and re-solves the winner before reporting it.

Workflow workspace file tree: catalog, requirements, skill, and one folder per run

The workflow workspace: product catalog, requirements, the skill document, and one folder per run.

Single run file tree: request, decision log, argument files, and the working HFSS project

Inside a single run: request, decision log, argument files, and the working HFSS project — replayable after the fact.

Stack
Ansys HFSS integrationWorkflow-as-skill documentBayesian optimization loopAutomated datasheet generationPer-run audit trail
Time to Production

Under 5 weeks, start to a working end-to-end run

Questions an Engineering Leader Will Ask
Can an AI agent run Ansys HFSS optimization unattended?

Yes. In this pilot a single run is 1.5 hours of unattended agentic work covering the full cycle — load the catalog model, baseline solve, KPI check against the customer spec, Bayesian optimization over approved parameters, re-solve of the winning candidate, and a generated datasheet. It explores 20+ full-wave candidates per run with zero human touches in the loop, replacing an engineer-day of expert HFSS work per configuration.

How do you validate a simulation an engineer did not watch run?

Every solve, variable change, and decision is logged, and each run keeps its request, decision log, and working solver project on disk. The run is replayable and reviewable after the fact, and the winning candidate is re-solved before it is reported — so the number in the datasheet comes from a verified solve, not from an interpolated surrogate.

What happens to the automation when Ansys ships a new release?

We carry it. The workflow is encoded as a skill document rather than a scripting project, which is what makes it cheap to keep current, and maintenance against solver-version drift is part of how we deliver rather than a separate project each time.

Bring us one workflow

A 30-minute screen share of the work your engineers repeat most — free, under NDA. We tell you on the spot whether it is automatable, what a pilot looks like, and what it would save.