Dynamic Multi Node Thermal State Estimation in Anisotropically Etched MEMS Resonators
Dynamic multi-node thermal observers track non-uniform crystal-plane heat gradients in MEMS resonators to cancel frequency drift under fast thermal transients.

Gradient
Heat dissipation through microscale silicon structures depends on crystalline orientation and etch geometry. Single-crystal silicon has a nominal room-temperature thermal conductivity of 148 Watts per meter-Kelvin along bulk directions, but etching trenches into a wafer alters this transport profile. Wet anisotropic chemical etching using potassium hydroxide or tetramethylammonium hydroxide exposes {111} crystallographic planes at an angle of 54.74 degrees relative to the (100) wafer surface.
By contrast, deep reactive ion etching produces vertical sidewalls with microscopic periodic scallops from alternating etch and passivation cycles. These geometric boundaries constrict the mean free path of acoustic phonons, which dominate heat conduction in non-metallic crystalline lattices.
Phonon boundary scattering becomes pronounced when characteristic structural dimensions approach the phonon mean free path, approximately 200 to 300 nanometers at 300 Kelvin. In suspended suspension tethers with widths below two micrometers, effective thermal conductivity drops by 20 to 60 percent, causing localized thermal choking. Energy deposited by drive electrodes or piezoresistive readout tracks cannot equalize instantaneously throughout the proof mass.
Temperature fields across the vibrating element turn non-uniform, establishing spatial variations that shift the mechanical resonance frequency through the temperature coefficient of elasticity.

Crystal Planes and Heat Conduction
Anisotropic chemical dissolution rates differ by orders of magnitude across silicon crystallographic families. Under standard bath concentrations of 30 percent at 80 degrees Celsius, the etch rate of potassium hydroxide on {100} planes exceeds that on {111} planes by more than one hundred to one. The resulting slanted sidewalls introduce a trapezoidal cross-sectional taper along suspension beams, so heat flux traveling from the central vibrating body toward the substrate anchor encounters a cross-section that expands or contracts nonlinearly along the beam axis.
Silicon conducts heat along crystal axes.
Anchor thermal resistance reaches 42000 Kelvin per watt under high vacuum at 300 Kelvin.
Heat transfer modeling treats conduction along the tether as a one-dimensional flow path governed by Fourier law modified for boundary scattering. The thermal resistance of a trapezoidal tether etched along (110) directions on a (100) substrate incorporates this variable width profile. Conduction along {110} directions shows slight bulk anisotropy compared to {100} directions, with conductivity variations between 5 and 15 percent depending on doping concentration and ambient temperature.
Boron doping above 1.0e19 atoms per cubic centimeter, commonly introduced for ohmic contacts or etch-stop definitions, reduces lattice thermal conductivity through impurity scattering, worsening thermal isolation between the resonant body and the package heat sink.

Thermoelastic Dissipation across Asymmetric Boundaries
Mechanical vibration induces cyclic localized strain fields within the resonator body. Compressed regions undergo momentary heating while expanding regions cool, generating an irreversible heat flux across the structural strain gradient that causes thermoelastic energy loss and limits the mechanical quality factor. In symmetric beams, the thermoelastic relaxation peak frequency aligns with beam thickness and thermal diffusivity, but angled sidewalls from wet anisotropic etching break spatial reflection symmetry across the neutral bending axis.
This asymmetric cross-section shifts the neutral axis and broadens the spatial strain distribution. Heat flows both across the beam thickness and along angled facet diagonals, coupling out-of-plane and in-plane thermal gradients. Because of this, standard one-dimensional Zener approximations underestimate thermoelastic damping in trapezoidal cross-sections by 18 to 35 percent.
Thermal equilibrium relaxes across multiple distinct timescales ranging from hundreds of nanoseconds within the flexure neck to tens of milliseconds across the entire proof mass structure.
Whether structural facet asymmetry induces non-monotonic spatial temperature distributions under rapid ambient thermal transients remains unresolved in literature dealing with sub-micron anchor interfaces.

Mesh
Discretizing the suspended micromechanical resonator into distinct lumped thermal regions enables mathematical tracking of transient internal temperatures. A single lumped thermal capacity model fails during fast thermal slew events exceeding one Kelvin per second, where heat input from electrostatic actuation, piezoresistive sensing bias, and external package conduction distributes unevenly. A multi-node network partitions the device into physical sub-domains: the main proof mass plate, the flexural tether constrictions, the anchor mounting pads, and the die attach interface.
Each node represents a discrete thermal capacitance linked to neighboring nodes through thermal resistances, with physical dimensions and material properties establishing the network parameters. Node temperatures serve as internal state variables whose values govern the frequency drift of the mechanical modes while anchor pads sink heat into substrates.
| Sub-Domain Node | Etch Profile | Thermal Resistance (K/W) | Thermal Capacity (nJ/K) | Thermal Time Constant (ms) |
|---|---|---|---|---|
| Vibrating Plate Central Node | Vertical DRIE Sidewall | 14200 | 1.85 | 0.026 |
| Vibrating Plate Central Node | Wet Etched 54.74 Degree Facet | 16800 | 1.62 | 0.027 |
| Suspension Tether Flexure | Vertical DRIE Sidewall | 52000 | 0.14 | 0.007 |
| Suspension Tether Flexure | Wet Etched Trapezoidal | 68500 | 0.11 | 0.008 |
| Anchor Interface Pad | Oxide Boundary Layer | 31000 | 12.40 | 0.384 |
| Package Substrate Cavity | Ceramic Carrier Base | 1850 | 450.00 | 0.832 |

Discretized Thermal Resistance and Capacitance Networks
Spatial discretization maps the heat diffusion partial differential equation into a set of coupled ordinary differential equations. The thermal capacitance of node index i equals the product of mass density, specific heat capacity, and node volume. The thermal resistance connecting node i and node j accounts for conductive path length, cross-sectional area, and effective thermal conductivity.
In anisotropic wet-etched geometries, the cross-sectional area changes continuously along the coordinate axis, requiring integral evaluation of the conductive resistance between node centers.
Convective heat dissipation to surrounding gas molecules vanishes inside hermetic packages maintaining vacuum levels below 0.01 millibar. Heat rejection from the central resonator plate occurs almost exclusively through conductive transport along the suspension tethers into the anchor, as radiative exchange between the resonator surfaces and the silicon cap wafer accounts for less than 0.5 percent of total heat transfer below 350 Kelvin. The high thermal resistance of narrow tethers isolates the proof mass, creating thermal time constants three orders of magnitude larger than the mechanical oscillation period.
Thicker anchor tethers shift heat dissipation from radiation to bulk conduction while lowering mechanical quality factor.

Can Dual Mode Drive Isolate Tether Conduction?
Interrogating two distinct mechanical resonance modes simultaneously yields independent observable metrics for tracking internal thermal states. A square Lamé-mode resonator exhibits a primary extensional vibration mode and an auxiliary out-of-plane flexural mode, each possessing a unique spatial strain distribution across the plate and anchor joints. The temperature coefficient of frequency differs between these modes because acoustic wave velocities depend on different combinations of single-crystal silicon elastic stiffness coefficients.
The frequency tracking circuit monitors both modes in real time, using the difference in instantaneous frequency shifts to separate uniform bulk die temperature changes from localized tether heating. Localized dissipation in the flexure anchors exerts a pronounced effect on high-strain flexural modes while leaving the volume-dilatational Lamé mode comparatively undisturbed. Tracking both outputs creates two measurement equations that map directly into the two dominant thermal state nodes of the discretized physical network.

Observer
Mathematical observers reconstruct inaccessible internal temperature states using available frequency outputs and known excitation power inputs. These physical states represent the temperatures at the proof mass, tether, and anchor nodes relative to ambient reference. In state-space notation, the continuous-time thermal diffusion dynamics follow a linear differential matrix system.
The state matrix contains the conductive couplings divided by nodal heat capacities, while the input matrix maps electrical heating from bias currents and actuation signals into specific thermal nodes where electrostatic drive voltages induce ohmic losses.
Discrete-time implementation on digital signal processing hardware converts the continuous thermal model into a recursive filter. The measurement equation links the observed resonance frequency shifts to the state vector through modal temperature coefficients. A discrete Kalman filter updates temperature estimates at each measurement epoch, weighting model predictions against noisy frequency counter readings via the Kalman gain matrix.
| Estimator Architecture | Node Count | Tracking Bandwidth (Hz) | Peak Estimation Error (mK) | Execution Cycles (Clock Cycles) |
|---|---|---|---|---|
| Static Polynomial LUT | 1 | 1.2 | 840 | 12 |
| Luenberger Observer | 3 | 45.0 | 115 | 180 |
| Linear Kalman Filter | 3 | 85.0 | 28 | 640 |
| Extended Kalman Filter | 5 | 120.0 | 14 | 2150 |
| Execution cycles benchmarked on a 32-bit ARM Cortex-M4 running at 120 MHz with single-precision floating point operations. | ||||

State Space Observer Formulations
The state-space model defines the state vector x as the temperature offsets of the n nodes. System dynamics matrix A contains negative diagonal elements representing total thermal conductance out of each node, while off-diagonal elements represent conductance between adjacent nodes. Input vector u contains power dissipation terms from drive electrodes, sense resistors, and die-level circuitry, and output scalar y denotes normalized resonance frequency deviation in parts per million.
The filter execution cycle runs two alternating steps: time propagation and measurement correction. During time propagation, the state estimate projects forward using the discrete state transition matrix derived from matrix exponential calculation, and the error covariance matrix expands to account for unmodeled thermal noise and ambient temperature fluctuations. During measurement correction, the discrepancy between actual measured resonance frequency and predicted frequency updates the nodal temperature estimates proportionally to the calculated gain.
Silicon lattice orientation governs thermal expansion mismatch against borosilicate glass substrates.

Which State Matrices Preserve Numerical Stability?
Thermal time constants in micromechanical structures span multiple orders of magnitude, causing the system dynamics matrix to exhibit stiff mathematical characteristics. The ratio of the largest to smallest eigenvalue can exceed 1.0e4 when modeling tiny anchor necks alongside bulky package substrates. Standard explicit Euler discretization causes numerical instability unless the sampling interval remains smaller than twice the smallest thermal time constant, which often demands sampling frequencies exceeding 50 kilohertz.
Implicit backward differentiation or bilinear Tustin transformations preserve numerical stability across wide sampling intervals. The Tustin transformation maps the left half of the complex s-plane into the interior of the unit circle in the z-plane, guaranteeing bounded-input bounded-output stability for all non-negative sample periods. Matrix inversion operations required during implicit updates can be precomputed offline when thermal conductivities remain linear over the operating range from minus 40 to plus 85 degrees Celsius, though dedicated floating-point units raise silicon area costs.
Neglecting node-to-node thermal gradients during fast temperature slews causes the frequency compensation loop to apply incorrect bias voltages, inducing frequency overshoot exceeding five parts per million and triggering synchronization loss in communication timing receivers.

Screen
Empirical validation of multi-node thermal models demands precision characterization benches capable of imposing controlled heat flux transients while monitoring resonance frequencies. Environmental chambers cycling ambient air temperature introduce slow thermal slews capped near 0.2 Kelvin per second due to thermal mass in chamber walls. Specialized validation fixtures integrate focused infrared laser pulses or localized resistive micro-heaters to deliver step changes in heat flux with rise times below five milliseconds.
The vacuum environment inside the testing fixture replicates operational hermetic packaging conditions. Pressure levels inside the chamber must stay below 1.0e-4 millibar to eliminate air damping and parasitic gas conduction across sub-micron capacitive electrode gaps. A turbomolecular pumping station paired with an oil-free scroll backing pump maintains baseline vacuum levels without introducing hydrocarbon contamination onto exposed silicon proof masses.
Thermal tracking validation identifies several critical structural disturbance vectors across the operating envelope:
- Radiant Heat Absorption from neighboring power management integrated circuits causes asymmetric surface heating across unshielded silicon die.
- Conductive Anchor Shocks transmitted through printed circuit board solder joints create localized mechanical strain across the package die attach boundary.
- Piezoresistive Bias Dissipation elevates flexure tether temperatures by ten to forty millikelvins above the central resonator body.
- Gas Conduction Leakage occurring during package seal degradation alters the thermal resistance between the proof mass and cap wafer.

Radiant Pulse Bench Instrumentation
Characterizing the transient step response of individual internal nodes involves delivering calibrated optical energy to the resonator die. A 980-nanometer diode laser focused through an optical microscope objective illuminates a ten-micrometer spot on the proof mass plate. A high-bandwidth arbitrary waveform generator modulates the laser diode drive current to produce square-wave heat pulses, and optical power absorbed by the silicon surface is quantified by subtracting reflected beam power measured via an auxiliary calibrated photodiode.
The resonance frequency shift is recorded using a high-resolution frequency counter synchronized to the laser pulse modulation trigger. Operating at a 10-kilohertz sampling rate with single-millihertz frequency resolution, the counter captures a transient response profile that reveals the multi-exponential decay characteristics of the thermal network. Fitting this measured decay curve against the multi-node state model allows direct extraction of thermal resistances and nodal time constants.

Laser Doppler Vibrometry Thermal Mapping
Spatial temperature distributions induce thermal expansion fields that alter mode shapes and velocity distributions across the vibrating micro-structure. A differential laser Doppler vibrometer scanning across the resonator surface measures out-of-plane velocity profiles with picometer-level displacement resolution. Localized thermal stress modifies resonant mechanical amplitude distributions along flexural suspension beams, where wet etching leaves angled crystal planes and crystal defects pin mobile dislocations.
Executing an automated parameter extraction sequence confirms state-space observer tracking performance across controlled operating conditions:
- Evacuate the vacuum test chamber to an operating pressure below 1.0e-4 millibar while maintaining baseplate temperature at 25 degrees Celsius.
- Drive the primary mechanical resonance mode into stable self-oscillation using an external phase-locked loop tracking circuit.
- Apply a calibrated five-milliwatt optical heat pulse of one-second duration to the proof mass center using the focused infrared laser.
- Log the instantaneous resonant frequency deviation at ten-thousand samples per second throughout the heating pulse and subsequent cooling relaxation.
- Compare the reconstructed nodal temperature trajectories against the recorded frequency response to verify state observer convergence time.
Calibrating state observers against room-temperature pulse data reliably predicts high-temperature transient behavior when thermal conductivity scaling curves incorporate acoustic phonon boundary scattering limits.

Yield
Commercial production of micro-resonators relying on multi-node state estimation demands strict control over etch facet repeatability and package cavity pressure. Wet anisotropic etching with potassium hydroxide provides smooth sidewall surfaces with low surface defect densities, reducing anchor acoustic losses. However, batch wet chemical baths suffer from chemical depletion and bath composition drift, altering sidewall angles by fractions of a degree across an eight-inch wafer.
A 0.5-degree variation in the 54.74-degree {111} plane angle shifts suspension beam thermal resistance by up to eight percent, moving state-space model parameters away from nominal calibration.
Deep reactive ion etching delivers tight vertical dimensional control with typical sidewall tolerances within 0.2 degrees. Scallop depth and surface roughness generated by the Bosch process introduce localized phonon scattering sites that depress thermal conductivity non-uniformly across process runs. Wafer-level capping technologies enclose the resonator die in high vacuum before wafer sawing, controlling gas conduction while getter films absorb outgassed molecules.
MIL-STD-883 Method 1014 condition B2 rejects packages exhibiting helium leak rates above 1.0e-8 standard cubic centimeters per second.

Wafer Etch Economics and Uniformity
Wet chemical etching tanks process batches of twenty-five wafers simultaneously, achieving high throughput and low tooling amortization costs. Maintaining etch bath chemistry demands continuous monitoring of water evaporation and silicate reaction byproduct concentrations. Etch stops utilizing heavily doped boron buried layers create lattice strain due to covalent radius mismatch between boron and silicon atoms.
This mismatch introduces residual stress gradients that warp suspended plates after release etching, altering thermal contact resistance at anchor points.
Plasma-based deep reactive ion etching processes wafers individually in single-wafer chambers. Inductively coupled plasma etch tools achieve high etch rate uniformity across the wafer surface through multi-zone gas distribution rings and dynamic electrostatic chuck temperature management. Capital depreciation and process gas consumption, specifically sulfur hexafluoride and octafluorocyclobutane, elevate the raw fabrication cost per square millimeter of silicon by a factor of 2.5 to 3.8 compared to wet chemical etch lines.
High production volumes balance this expense through improved spatial uniformity and tighter thermal model parameter distributions.

Package Hermeticity and Getter Depletion
Long-term retention of vacuum inside the cavity governs the stability of thermal observer coefficients over operating lifespans exceeding ten years. Micro-scale cavity hermeticity relies on metallic eutectic bonding, such as gold-silicon or aluminum-germanium, or glass frit bonding between the device wafer and cap wafer. Non-evaporable getter thin films deposited on the cap cavity ceiling absorb outgassed water vapor, hydrogen, carbon monoxide, and nitrogen through chemical chemisorption after thermal activation at 350 to 400 degrees Celsius.
Procurement documents for high-stability timing components incorporate strict verification mandates to screen out structural and thermal parameter drift:
- Wafer Sidewall Angle Tolerances restricted to within 0.3 degrees of nominal design across all wafer production lots.
- Residual Cavity Pressure Limits maintained below 0.05 millibar throughout twenty-year simulated operational lifetimes.
- Active Getter Surface Coverage specified across at least forty percent of the interior cap wafer cavity area.
- Temperature Coefficient Tracking Accuracy held within five percent across the full military temperature range from minus 55 to plus 125 degrees Celsius.
Standard purchase agreements incorporating AEC-Q100 grade 1 qualification impose mechanical shock and thermal cycling screens that reveal whether micro-scale thermal anchoring interfaces undergo delamination or stress relaxation under repetitive thermomechanical loading.




