Designing a 38-gram sensor without compromise.

Thirty-eight grams was not a marketing target. It is the point past which the sensor stops feeling like an intrusion on the lift and starts behaving like dead weight that changes the bar's dynamics. The mass constraint is real, but it is not the only constraint. The sensor also has to survive high-g impacts, stay locked to the bar under violent acceleration, and deliver lab-grade motion data at 200 Hz with velocity resolved to three decimals in m/s. Every design decision is a trade between those four demands.

Before we committed to any geometry, we measured drop heights and impact orientations from realistic gym scenarios: a bar rolling off a rack at roughly waist height, a plate stack nudging the mount, a direct vertical drop onto a platform, and a barbell snapping into the rack after a missed catch. Those became the boundary conditions for every simulation, test, and material decision that followed. An arbitrary safety factor is not engineering; it is a guess with a number attached. Every structural decision downstream traces back to a measured or conservatively bounded load case.

The housing material could not be chosen on stiffness or impact resistance alone. It had to be chosen against the full set of constraints the simulation was surfacing: strength-to-weight ratio under dynamic load cases, fatigue behavior over repeated impacts, thermal stability across a garage in winter to a hot car in summer, and the machining tolerances required to protect the IMU assembly inside.

We went through the selection together. Nicolas brought candidate geometries and I brought the coupled FEA results. We ruled materials in or out based on what happened under a transient impact, not on datasheet properties. A material that looks strong in a static pull test can still underperform once you factor in how it behaves after thermal cycling or under a high-g impact. Precision-sintered PA-12 nylon emerged as the only material that cleared every load case at the mass we required. It has the toughness to absorb a barbell impact instead of transmitting the shock straight into the IMU board, the fatigue resistance to take that hit thousands of times without micro-cracking, the dimensional stability to hold the tolerances that keep the sensor assembly aligned, and it is radio-transparent, so the Bluetooth link is never fighting its own housing.

The cheaper option was a commodity injection-molded ABS shell. It would have been faster to tool and far less expensive per unit. It also would have gone brittle and cracked on the first season of hard drops. Metal went the other way: a machined alloy case would have been stiffer on paper, but it transmits impact energy directly into the electronics, adds mass we could not spend, and detunes the antenna. We did not choose the cheap material or the impressive-sounding one. We chose the one that survives.

A sensor that slips is worse than a sensor that does not exist. If the magnet gives up during a clean or a heavy jerk, the data is not just wrong, it is dangerous, because the athlete might keep training on a number that did not come from the bar. We specified N52 grade neodymium, the highest magnetic grade available in a reliable, repeatable supply chain, and arranged four magnets to deliver a combined 200 N of pulling force.

That number is not arbitrary. 200 N is well beyond what the bar can transfer to the sensor during the most violent lifts we modeled, including drop-loaded cleans and snatches where the bar decelerates under load. We did not design for the average rep. We designed for the rep where the athlete is losing control and needs accurate feedback the most. For non-ferromagnetic bars, and for cases where a mechanical guarantee is preferred, the compact strap mount is included in every kit. But the magnetic mount is the primary workflow, and it had to be uncompromising.

The housing and magnets protect the device. The IMUs are the reason the device exists. We chose two independent research-grade 6-axis inertial measurement units and fused them with a Kalman filter running on the device. Each IMU samples internally at 32 kHz, anti-aliased and fused, then streams motion data at 200 Hz over Bluetooth Low Energy 5.0. Velocity is resolved to three decimals in m/s.

Two IMUs matter because a single IMU is vulnerable to drift, noise, and mounting bias. With two independent sensors, the gyroscope and accelerometer channels cross-check every rep, so noise is filtered out and drift is eliminated. A single-axis tethered tool can only measure in the axis it is pointed; it silently filters out bar tilt, drift, and rotation. ORCA reconstructs the true 3D path of the bar because that is what the athlete actually moved.

The IMUs are research-grade because the difference between a consumer-grade MEMS sensor and a research-grade one is visible in the third decimal. If the data is not reliable enough to stand up to a coach's decision or a lab's peer review, it is not useful to either of them.

Once Nicolas had a candidate geometry, I ran finite element analysis across several coupled regimes rather than treating this as a single static stress problem:

• Structural, to check peak stress concentrations in the housing and mounting points under the worst-case impact orientation.
• Dynamic, because a barbell impact is a transient event, not a static load. The response of the housing over milliseconds, including how internal components resonate against their mounts, determines whether the PCB or battery sees damaging acceleration even if the housing itself survives.
• Buckling, on the thinner wall sections where minimizing mass pushes wall thickness down to the point where local instability, not yield stress, becomes the governing failure mode.
• Thermo-mechanical, since the housing also has to tolerate temperature cycling without inducing stress at material interfaces that a purely mechanical model would miss.

Running these as coupled multiphysics problems, rather than four separate checks, is what let us find failure modes that would not show up if we had only run a static structural case at room temperature and called it done. Every time a simulation flagged a governing failure mode, that went back to Nicolas as direct input to the next geometry revision.

Simulation tells us where a design should fail first and at what load. It does not replace physical validation; it directs it. Every housing revision that passed simulation went through physical drop testing at the same load cases the FEA was built around, specifically to catch anything the model's assumptions did not capture, like manufacturing tolerances in the actual machined part or adhesive behavior that is hard to characterize analytically.

If a physical test failed at a load the simulation said should be safe, that is not a testing problem, it is a modeling gap, and we go back and figure out what boundary condition or material property was wrong before touching the geometry again. We also tested the magnetic hold on real bars under real loads, because the FEA does not know how chalk, bar knurling, or impact bounce affect the magnet interface.

A number we put on a spec sheet has to survive three separate tests: it has to hold up in simulation, it has to hold up if we derive it by hand on a whiteboard as a sanity check against the model, and it has to hold up when the physical part is actually dropped, impacted, or lifted with. If any one of those three disagrees with the others, the design is not finished, regardless of how good it looks in the other two.

That is the discipline that gets a 38-gram sensor to survive years of gym use without becoming fragile or over-built. It is also the discipline that prevents us from cutting corners on the components inside it. The PA-12 nylon case, the N52 magnets, and the dual research-grade IMUs are not premium upgrades. They are the baseline we chose because the cheaper alternatives were not good enough for the job.

Léo Fortin Dionne works on the advanced measurement systems and engineering simulation behind ORCA Strength Systems.