Mechanical Reliability — CardiaNova CMU

Arch A · contact bearing

43.6%

R(5 yr) — bearing dominates

Arch B · maglev

60.7%

R(5 yr) — driveline dominates

Arch C · maglev + TET

72.4%

R(5 yr) — electronics dominate

Target

90%

R(5 yr) at 90% confidence

Scope and status

This is a design-exploration study, not validated device evidence. No physical hardware has been tested. Every input is a handbook material property, a physiological operating condition, or an explicitly labelled engineering estimate.

The numbers here are for ranking design options and sizing the test programme. They are not reliability claims and cannot support any regulatory or clinical statement. Section 8 lists exactly which inputs must be replaced with measured data before any of this becomes a claim.

01The question

"The device must last for years" has to become something we can design and test against. This study adopts a primary durability target of R(5 years) ≥ 90% at 90% confidence, with R(10 years) as a stretch goal, and asks one question throughout: what actually stops this device from reaching five years, and what should we do about it?

02Duty cycle — why ordinary fatigue data does not apply

A continuous-flow pump at 3,000 rpm accumulates load cycles at a rate that puts it outside the range where fatigue data normally exists.

Load spectrum	5 years	10 years
Rotor revolutions	7.9 × 10⁹	1.6 × 10¹⁰
Blade-passing events	4.7 × 10¹⁰	9.5 × 10¹⁰
Cardiac cycles (housing, seals)	2.0 × 10⁸	3.9 × 10⁸

Two consequences follow:

Two independent load spectra exist and differ by about 40×. The rotor and bearing see revolutions; the housing and seals see heartbeats. Sizing both against one number is wrong.
Standard S-N curves stop at 10⁷ cycles. Our rotor passes that in under 14 hours. Titanium alloys have no true endurance limit — strength keeps falling past 10⁷ — so the allowable stress must be extrapolated into the very-high-cycle regime.

Consequence

Designing to a "10⁷ endurance limit" would overstate the allowable stress by 1.9×. This is the most common error in this class of analysis.

Bar chart on a log scale comparing load cycles: standard S-N test truncation at 10 to the 7, cardiac cycles at 2.0e8, rotor revolutions at 7.9e9 for five years and 1.6e10 for ten years, and blade-pass events at 9.5e10. — **Fig 1.** Load cycles accumulated in service versus where fatigue test data conventionally stops.

03Rotor fatigue — not the problem

The Ti-6Al-4V impeller was analysed for centrifugal stress, blade-root bending, resonance, and very-high-cycle fatigue.

Centrifugal stress is negligible

σ_θ = (3+ν)/4 · ρω² · [ b² + (1−ν)/(3+ν) · a² ]

At 4,500 rpm this gives 0.33 MPa against an 880 MPa yield — a margin of roughly 2,500×. At this size and speed the rotor is nowhere near centrifugally limited.

Blade-root bending dominates, and is still small

Treating a blade as a cantilever carrying the full pump head across its face (deliberately conservative), with a stress-concentration factor of 2.0 at the root fillet, gives 3.1 MPa.

Fatigue assessment (Basquin + Goodman)	5 years	10 years
Cycles	7.9 × 10⁹	1.6 × 10¹⁰
Allowable alternating stress	162 MPa	152 MPa
Applied (Goodman-equivalent)	3.1 MPa	3.1 MPa
Safety factor	52×	48×

Resonance is clear

Blade first bending mode sits at 26,200 Hz against blade-passing excitation of 450 Hz — a 58× margin. On magnetic suspension the rigid-body mode is 168 Hz against a maximum operating frequency of 75 Hz, so the rotor runs subcritical with 2.2× separation.

Log-log S-N curve for Ti-6Al-4V showing allowable stress falling from 316 MPa at 10 to the 7 cycles to 162 MPa at the five-year cycle count, with the applied 3.1 MPa alternating stress far below both. — **Fig 2.** The allowable stress must be extrapolated to the actual cycle count. Even after a 1.9× knockdown, the applied stress sits far below it.

Finding 1

Rotor structural fatigue is not the life-limiting element. Safety factor ≈ 50 even after very-high-cycle extrapolation. Effort spent thickening the rotor buys nothing.

Caveat: this assumes defect-free material and good surface finish. In the very-high-cycle regime, failures initiate from subsurface inclusions rather than the surface — so this conclusion depends on inclusion control in the supply chain. That is a materials-qualification task, not a stress-analysis one.

04Bearing — the real mechanical limit

Over five years the journal contact slides 148,700 km — 3.7 times around the Earth — at 0.94 m/s, continuously, immersed in blood, with no possibility of maintenance.

Wear life is unpredictable by two orders of magnitude

Archard wear law

\begin{array}{rcl} V & = & \frac{k W s}{H} & worn volume \\ h & = & \frac{V}{A_{c}} & linear wear depth \\ A_{c} & = & 2 r L & projected contact area \end{array}

Failure when h \geq c

Where

Symbol	Meaning	Value	Basis
$V$	Volume of material worn away	computed
$k$	Archard wear coefficient — dimensionless; the fraction of contacts that shed a wear particle	10⁻⁹ – 10⁻⁷	ESTIMATE
$W$	Radial load pressing the journal into the bush — net hydraulic + magnetic side load	2 N	ESTIMATE
$s$	Sliding distance, $s = 2 π r N$ over $N$ revolutions	1.487×10⁸ m @ 5 yr
$H$	Hardness of the softer surface — alumina/zirconia class	20 GPa	HANDBOOK
$h$	Linear wear depth — how far the surface has receded	computed
$A_{c}$	Projected journal contact area	24 mm²	DESIGN
$r$ , $L$	Journal radius and length	3 mm, 4 mm	DESIGN
$c$	Design radial clearance — the gap wear is allowed to consume before the bearing seizes	30 µm	DESIGN

Worked, nominal k at five years

\begin{array}{rcl} h & = & \frac{k W s}{H A_{c}} \\ = & \frac{(10^{- 8}) (2 N) (1.487 \times 10^{8} m)}{(20 \times 10^{9} Pa) (2.4 \times 10^{- 5} m^{2})} \\ = & 6.2 µm \leq 30 µm \end{array}

Wear coefficient k	Wear @ 5 yr	Predicted life	Verdict
10⁻⁹ optimistic	0.6 µm	242 yr	PASS
10⁻⁸ nominal	6.2 µm	24 yr	PASS
10⁻⁷ pessimistic	62.0 µm	2.4 yr	FAIL

The wear coefficient for blood-immersed ceramic pairs with adsorbed protein layers spans roughly three decades in the literature. That spread propagates directly into a 100× spread in predicted life — straddling the target.

There is also a thermal path: 0.19 W of friction dissipated into a 24 mm² contact is 0.8 W/cm² locally. Blood proteins denature above about 42 °C, and local hot spots are a recognised thrombus nucleation site. This is a blood-compatibility failure mode that exists even when the bearing is mechanically within spec.

Log-log plot of predicted bearing life against Archard wear coefficient, showing life falling from 242 years at k equals 10 to the minus 9 to 2.4 years at 10 to the minus 7, crossing below the five-year target line. — **Fig 3.** The three-decade uncertainty in the wear coefficient produces a hundredfold spread in predicted life, straddling the five-year target.

Finding 2

A contact bearing cannot be certified, because its life cannot be predicted. The problem is not that the nominal answer is bad — 24 years is fine. It is that the honest error bar spans 2.4 to 242 years. You cannot demonstrate R(5 yr) ≥ 90% on a mechanism whose governing coefficient is unknown to two decades.

Elimination

	Contact bearing	Full magnetic levitation
Wear mechanism	Yes — unbounded uncertainty	None
Local heating	0.19 W at contact	None
Added power	—	~1.5 W
Failure mode	Progressive wear → seizure	Control fault → touchdown
Life predictable?	No (2 decades)	Yes — becomes electronics

Magnetic levitation is not more reliable by magic. It converts an unpredictable tribological problem into a predictable electronic one — and electronics reliability is something we know how to model, test, and demonstrate. That is the real argument, and it is the direction the field moved: HeartMate II used mechanical bearings, HeartMate 3 is fully levitated.

Recommendation 1

Adopt full magnetic levitation. Retain a touchdown bearing for fault conditions only. Budget approximately 1.5 W and the associated control complexity.

05Seal reliability — two very different problems

Hermetic enclosure — solvable, and specifiable

Moisture ingress into the electronics cavity, driven by the water-vapour partial pressure of saturated tissue at 37 °C:

Helium leak rate	Time to critical moisture
10⁻⁸ atm·cc/s	6.9 yr	MARGINAL
10⁻⁹ atm·cc/s — proposed spec	68.9 yr	PASS
10⁻¹⁰ atm·cc/s	689 yr	PASS

Maximum allowable leak rate for the five-year target is 1.4 × 10⁻⁸ atm·cc/s. The proposed 10⁻⁹ spec carries 14× margin and is verifiable on every unit with a standard helium fine-leak test.

Log-log plot of time to critical moisture against helium leak rate, with the proposed 10 to the minus 9 spec giving 69 years, well above the five-year target line. — **Fig 4.** The proposed hermeticity spec carries roughly fourteen times margin against the five-year target.

Finding 3

The hermetic seal is not a limiting risk, provided we specify and test it. Laser-welded titanium with brazed ceramic feedthroughs routinely achieves 10⁻⁹. Make it a 100%-inspection requirement.

Percutaneous driveline — the one that actually bites

The cable crossing the skin is not a mechanical sealing problem. It is a chronic infection pathway: the skin barrier is permanently broken, and driveline infection is one of the most significant sources of adverse events and readmission in long-term circulatory support. No amount of mechanical sealing fixes it — the failure is biological.

Recommendation 2

Treat elimination of the percutaneous driveline (transcutaneous energy transfer) as a reliability requirement, not a convenience feature.

06System reliability — where the bottleneck moves

Analysing subsystems individually is misleading. The device is a series system: any subsystem failing fails the device. The model below is a Monte Carlo over 200,000 virtual devices with a Weibull life model per subsystem.

Architecture	R(5 yr)	R(10 yr)	Median life	Dominant early failure
A — contact bearing	43.6%	3.6%	4.5 yr	Contact bearing (41%)
B — maglev	60.7%	24.9%	6.3 yr	Percutaneous line (53%)
C — maglev + TET	72.4%	38.4%	8.2 yr	Controller electronics (46%)

Reliability versus years in service for three architectures, showing 44 percent, 61 percent and 72 percent at the five-year mark, all below the 90 percent target. — **Fig 5.** Removing the bearing lifts reliability substantially, but no architecture reaches the target on current subsystem estimates.

Three small-multiple bar charts showing the dominant cause of early failure shifting from contact bearing to percutaneous line to controller electronics across the three architectures. — **Fig 6.** The dominant failure mode moves each time the previous one is fixed.

Finding 4

Bearing elimination is necessary but not sufficient. Eliminating the bearing lifts R(5 yr) from 44% to 61% — real and worth doing. But the bottleneck immediately moves to the driveline, and removing that moves it again to controller electronics. Reliability work on this device is a sequence of moving bottlenecks; fixing one in isolation produces a much smaller system-level gain than intuition suggests.

Turning the target into per-subsystem requirements

For a series system of eight subsystems to reach R(5 yr) = 0.90, each must average R(5 yr) ≥ 0.9869. Inverting the Weibull gives the required characteristic life:

Subsystem	η now	η required	Gap
Controller electronics	25 yr	185 yr	×7.4
Implanted battery	20 yr	44 yr	×2.2
Maglev suspension	40 yr	75 yr	×1.9
TET coil / link	60 yr	110 yr	×1.8
Hermetic seal	30 yr	44 yr	×1.5
Motor / drive coil	45 yr	55 yr	×1.2
Touchdown bearing	90 yr	28 yr	met
Rotor structure	180 yr	21 yr	met

Horizontal bar chart of required improvement factor per subsystem, with controller electronics needing 7.4 times and rotor structure and touchdown bearing already meeting their allocation. — **Fig 7.** Required improvement in characteristic life per subsystem. Nothing on the rotor.

Recommendation 3

The controller electronics are the binding constraint, not anything mechanical. A ×7.4 improvement will not come from better components alone — it needs redundancy (dual-channel controller with automatic failover), which is an architecture decision to make now, not later.

07Accelerated life testing

Mechanism	Model	Condition	AF
Thermal ageing	Arrhenius, Eₐ = 0.7 eV	37 → 60 °C	6.1×
Wear	Inverse power law, n = 2	1.5× rpm	2.2×
Combined			13.7×

Sizing the demonstration

n = ln(1−C)/ln(R) · ( t_target / (AF · t_test) )^β

Test approach	Units for a 1-year test
Real time (AF = 1)	546
Speed only (2.2×)	108
Thermal only (6.1×)	15
Combined (13.7×)	3

Log-scale plot of units required against test duration for four acceleration strategies, from 546 units at real time down to about 3 with combined acceleration. — **Fig 8.** Acceleration is what makes the durability test affordable.

Why "3 units" is not a real test plan

That formula assumes the acceleration factor is known exactly. It never is — AF depends on an activation energy and a power-law exponent, both estimated. Propagating a log-normal uncertainty on AF:

Units	P(test actually delivers R ≥ 0.90)
3	51.8%	COIN FLIP
6	85.2%	SHORT
12	97.9%	ADEQUATE
20	99.7%	ADEQUATE

Bar chart of the probability a test delivers its claimed reliability by sample size, rising from 52 percent at three units to 98 percent at twelve units. — **Fig 9.** Once acceleration-factor uncertainty is included, the textbook sample size delivers its claim only about half the time.

Recommendation 4

Test 12 units for one year at combined acceleration (≈13.7×, so roughly 14 equivalent field-years each). Robustness alone needs 8; raise it to 12 so a Weibull shape parameter can actually be fitted from the data rather than assumed.

Run a subset to destruction to observe real failure modes. A zero-failure test that passes teaches you nothing about how the device fails. ISO 14708-1 and ISO 14708-5 govern this device class — a real durability programme must be written against them. This study sizes the engineering test; it does not satisfy the standard.

08What to measure first

Parameter	Life range	Swing
Archard wear coefficient k	2.4 – 242 yr	×100
Radial load W	9.7 – 48.4 yr	×5
Clearance	16.1 – 40.3 yr	×2
Hardness H	14.5 – 33.9 yr	×2
Speed	16.1 – 36.3 yr	×2
Bearing radius	no effect	×1

Tornado chart of predicted bearing life sensitivity, with the wear coefficient spanning a hundredfold range and every other parameter spanning five times or less. — **Fig 10.** The wear coefficient dominates every other input by a factor of twenty.

Two things worth keeping:

k dominates everything else by 20×. If a contact bearing stays in the design, bench-measuring k on the actual material pair in a blood analogue is the single highest-value experiment available.
Bearing radius cancels out entirely. Sliding distance scales with r and contact area scales with r, so wear depth is independent of it — worth knowing before anyone spends a week optimising that dimension.

Inputs that must be replaced with measured data

Input	Current basis	How to fix
Archard k	Literature, 3 decades	Pin-on-disc in blood analogue
Radial load W	Estimate, 1–5 N	CFD + bench force measurement
Basquin σ'f, b	Representative Ti-6Al-4V	Coupon fatigue, production finish
Subsystem Weibull β, η	All placeholders	The 12-unit ALT above
Acceleration factors	Assumed Eₐ, n	Multi-stress ALT to fit them
Friction coefficient	Estimate 0.05–0.25	Tribometer, protein medium

09Conclusions

Rotor fatigue is not the constraint — safety factor ≈ 50 even in the very-high-cycle regime. But extrapolate the S-N curve properly; the 10⁷ endurance limit overstates strength by 1.9×.
The contact bearing is uncertifiable — not because its nominal life is bad, but because its life cannot be predicted within two orders of magnitude. Eliminate it; adopt magnetic levitation.
Hermetic sealing is a solved problem at 10⁻⁹ atm·cc/s with 14× margin. Make it a 100%-inspection spec.
The percutaneous driveline becomes the dominant failure mode the moment the bearing is fixed. Plan transcutaneous energy transfer now.
Controller electronics are the binding constraint at ×7.4 — solve with redundancy, an architecture decision that must be made early.
Test 12 units for one year at ~14× acceleration. The textbook answer of 3 units delivers the claimed reliability only about half the time.

Priority order

Magnetic levitation → driveline elimination → controller redundancy → battery. Nothing on the rotor.

Every number in this document was produced by the analysis pipeline (params.py · analysis.py · figures.py · run_all.py); none were entered by hand. To challenge any conclusion, change one value in params.py and re-run — the whole chain updates.

CardiaNova · Chiang Mai University · Revision A · 3 August 2026
Design-exploration study. Not validated device evidence. Not a reliability claim.