Platform · Optimization
Optimization agents
Every strong MEMS design lives on a trade-off: push sensitivity and you give up robustness to fabrication scatter. Optimization agents map that whole Pareto front while spending only a fraction of the expensive finite-element solves a direct search would burn.
01The problem
Bosch builds MEMS sensors at wafer scale, where a pressure membrane can deflect a few picometers under a small pressure change and still has to survive process scatter across thousands of dies. Its engineers frame the work plainly: multiple requirements that are hard to meet at once, so they run multi-objective optimization. The catch is that each candidate needs a full multi-physics solve, so a naive search spends thousands of runs just to draw one clean front.
02How we solve it
Surrogates that learn where to solve.
Agentriq trains a probabilistic surrogate on a handful of FEA runs, then lets acquisition functions, learned policies and evolutionary search decide which candidates are worth a real solve.
Multi-objective Bayesian optimization
A Gaussian-process surrogate models each objective with calibrated uncertainty, and expected-hypervolume-improvement picks the next design that most expands the Pareto front.
Reinforcement-learning policy search
When a design is built as a sequence of decisions, a policy learns which move to make next from the reward each candidate earns, the same family of methods now driving chip floorplanning and inverse device design.
Surrogate-assisted evolutionary search
NSGA-II and NSGA-III evolve a population against the surrogate, so the front is refined in software before any expensive solver is called.
Active learning and Latin-hypercube design
Latin-hypercube sampling seeds the surrogate with broad coverage, then only high-uncertainty candidates reach full FEA and the surrogate is retrained each wave so accuracy concentrates where the decisions are close.
03What it produces
A Pareto front in fewer solves.
A converged sensitivity versus robustness front reached on a small fraction of the solver budget: recent MEMS optimization work hit high-quality solutions using roughly 2.8% of the FEM calls, about 70 solves instead of 2,500.
04The agentic loop
Closing the optimization loop.
Agentriq wraps the whole search in a closed agent loop that proposes candidates, verifies the promising ones against the real solver, and only writes back a front it can defend.
- 01Seed a space-filling DOE, fit the Gaussian-process surrogate, and let NSGA-II with expected-hypervolume-improvement propose the next candidate batch.
- 02Verify the most promising candidates with a full multi-physics FEA solve and fold the true results back into the surrogate.
- 03When a candidate misses target or the surrogate and solver disagree, trace the root cause and pull method and prior-design context through Agentic RAG.
- 04A judge scores the refined Pareto front against the sensitivity and robustness targets and decides accept, iterate, or escalate.
- 05Write the accepted front, the winning designs, and their full provenance back to the record.
05Tooling
Simulation, solvers & frameworks.
Get started