Industry · MEMS & sensors
MEMS & sensor technologies
Inertial MEMS lives or dies on a matched pair of resonant modes. When drive and sense frequencies drift apart across temperature, sensitivity and bias stability follow, and the dead-reckoning that ADAS leans on when GPS drops starts to wander.
01The problem
Consider an automotive-grade inertial unit like Bosch's SMU300, built to hold a vehicle's orientation and trajectory when GPS, radar, or vision fall away. Its gyroscopes depend on drive and sense modes staying matched across the full automotive temperature range, yet thermal expansion and quality-factor degradation shift each eigenfrequency at its own rate. Any residual frequency split between the two modes degrades Coriolis energy conversion, sensitivity, and bias stability, so that margin has to be engineered into the geometry long before a wafer is cut.
02How we solve it
Geometry that holds mode-match across temperature.
Agentriq treats mode-matching as a variation-aware geometry search, pairing high-fidelity multiphysics with surrogate models so robustness is designed in, not trimmed out later.
FEM eigenfrequency and modal analysis
Solve the drive and sense eigenmodes directly to read frequency split, mode shapes, and spurious modes without slow frequency sweeps.
Shape and parametric geometry optimization
Parametrize spring and proof-mass geometry, then tune the design variables to close the frequency split and push spurious modes clear of the operating band under manufacturability constraints.
Electromechanical multiphysics co-simulation
Couple structural mechanics with electrostatic actuation and anchor or squeeze-film damping to capture quality factor and scale factor, not just bare frequencies.
Surrogate modeling and variation-aware optimization
Train Gaussian-process surrogates with uncertainty and polynomial chaos, then run Bayesian, multi-objective searches that stay robust to fabrication spread.
03What it produces
Fewer solver calls, robust margins.
In published MEMS optimization work, an adaptive Gaussian-process surrogate reached comparable multi-objective designs using about 2.8% of the finite-element evaluations, roughly 70 instead of 2,500, turning weeks of solver time into hours.
04The agentic loop
From geometry sweep to verified decision.
Agentriq runs the whole search as a closed agent loop: it proposes geometries, evaluates them in the multiphysics stack, verifies mode-match and robustness against targets, and only promotes a design once the evidence holds.
- 01Propose candidate geometries and simulation cases from the current surrogate and the design targets.
- 02Run FEM eigenfrequency and electromechanical co-simulation, then retrain the surrogate on the fresh results.
- 03Verify frequency split, quality factor, and temperature-swept mode-match against spec, flagging any run that misses.
- 04On a miss, trace root cause through parameter sensitivities and an Agentic RAG pass over prior designs, papers, and process notes.
- 05An LLM judge weighs the evidence, accepts or rejects the candidate, and writes the decision, metrics, and rationale back to the design record.
05Tooling
Simulation, solvers & frameworks.
Get started