Platform · Simulation at scale

Simulation at scale

A real design study is slow in more than one place: building and meshing the model, running each high-fidelity solve, and sweeping thousands of variants all take time, and no team can run them one at a time. We train a surrogate on the solver, search the whole space on it in seconds, and validate only the best, so a sweep that would take months resolves in days or hours, and one search returns a whole portfolio, not a single design.

01What it produces

Not one design, a whole portfolio.

From millions of candidates to a validated portfolio. About twenty five algorithms sweep millions of candidate designs on the surrogate in seconds, against roughly five a day by hand; only the best are validated in the full solver, and the search returns a portfolio tuned for different objectives.

The gains were dramatic: sensitivity climbed to several times the baseline, bandwidth on some designs rose more than twenty times over, and the strongest objectives improved by orders of magnitude, with hundreds of designs landing in performance territory the baseline never came near. Because we could run so many simulations, tens of thousands in the full solver and millions more on the surrogate, we did not tune a single device: opening the design space that far let us build a whole portfolio of devices with completely different performance objectives, from automotive crash sensing to inertial navigation and seismic monitoring. Those solver runs alone would have taken about fifteen years at five simulations a day by hand; here they took days.

02The problem

Take a MEMS resonant accelerometer. Its sensitivity, bandwidth and measurement range pull against each other across an eighteen-parameter geometry, and every honest evaluation means setting up, meshing and running an expensive finite-element solve. By hand, an engineer runs on the order of five simulations a day, so exploring that space properly, which takes tens of thousands of runs, is out of reach. It is the same wall the largest engineering groups hit: NASA's own CFD ecosystem answers it by pairing job orchestration with trained surrogates, so large batches run at once and the model predicts the rest. The bottleneck is no longer the physics, it is how much of the design space you can afford to search.

Design-space exploration over a surrogate field. Two geometry parameters on the axes, an objective such as sensitivity or bandwidth as a surrogate-predicted contour, with the solved designs and the next adaptive batch on top. The surrogate turns the whole space into something you can sweep.

03How we solve it

Sweep millions on the surrogate, validate the best.

We train a surrogate on a batch of full-fidelity solves, then set a portfolio of about twenty five optimization algorithms loose on it. Together they generate and score millions of candidate geometries in seconds, we keep the best and most interesting, and we validate the top ones back in the full solver.

Orchestrate, learn, schedule the next batch. The parallel-plus-surrogate pipeline: solves fan across the machine, the surrogate learns the field, and active learning decides which designs to run next.

A surrogate you can sweep millions of times

A neural-network surrogate trained on the solver predicts a design's performance in microseconds, so the search scores millions of candidates for the cost of a handful of real solves.

A portfolio of about twenty five algorithms

Bayesian optimization, NSGA-III and other evolutionary search, CMA-ES, reinforcement learning, and local and gap-filling searches all run together, each probing the space differently so no promising region is missed.

Multi-objective, not a single answer

The search maps the whole trade-off surface between competing targets instead of chasing one number, returning a Pareto front of designs rather than a single point.

Validate the best in the full solver

The most promising and most interesting candidates are re-simulated at full fidelity and checked against every physics constraint, so each recommended design is real, not a surrogate's guess.

04The agentic loop

From sweep to a validated portfolio.

Agentriq runs the whole search as a closed agent loop: propose with the algorithm portfolio, screen on the surrogate, validate the best in the full solver, check the physics, and keep only the designs that survive.

From work to a verified decision. The same closed agent loop that turns a raw sweep into a validated, traceable portfolio of designs.

01Train a surrogate on a batch of full-fidelity solves spread across the design space.
02Turn about twenty five optimization algorithms loose on the surrogate; together they generate and score millions of candidate geometries in seconds.
03Shortlist the best and the most interesting, then re-simulate them in the full solver and check every physics constraint.
04On disagreement, retrain the surrogate on the fresh results and search again, sharpening it where the decisions are close.
05Keep the validated portfolio, a Pareto front of designs tuned for different objectives, with full provenance.

05Tooling

Simulation, solvers & frameworks.

Get started

From engineering problems to autonomous solutions.

Talk to our engineers john.ferderigos@agentriq-sa.com