The platform
Three pillars.One closed loop.
A precision cycler, lab software with an AI co-researcher, and a machine-learning model that calls cycle life early. Built together, so each one makes the others better.
The cycler that measures like an instrument.
An 8-channel precision cycler with per-channel power and measurement boards, integrated thermal control, and safety limits enforced in firmware. Installed, calibrated, and maintained by Cyclone, as a service.
8 independent channels
Each channel has its own power and measurement driver board. No multiplexing, no cross-talk between cells.
24-bit measurement, 1 kHz
0.015% of full scale on voltage and current, sampled up to 1,000 times a second. Model quality inherits measurement quality.
Thermal control in the loop
Integrated thermal control, with per-cell thermal fixtures for active heating and cooling, not just ambient.
Hard limits in firmware
Five absolute limits on voltage, current, and temperature are enforced on the board itself, microseconds from the cell, with or without a network.
- Model
- CBT-15
- Channels
- 8, independent
- Voltage range
- 0 – 5 V
- Current ranges
- 1.5 mA · 150 mA · 1.5 A · 15 A
- Measurement resolution
- 24-bit, V and I
- Control resolution
- 16-bit
- Voltage accuracy
- 0.015% of full scale
- Current accuracy
- 0.015% of full scale
- Current control accuracy
- 0.02% of full scale
- Sample rate
- up to 1,000 Hz
- Current response
- 0 → 90% in < 1 ms
- Temperature
- PT100 3-wire · K-type option
- Communication
- Ethernet 802.3 · USB-B
- Enclosure
- 533 × 514 × 178 mm · 4U rack
- On-channel EIS
- 10 mHz – 100 kHz
- Self-calibration on device
- 2 – 3 min, in process
- Parallel channel current
- up to 30 A, 2 channels
- Load-cell measurement
- 4-wire, 24-bit
Runs with local authority
The Eye bench app is offline-first. Protocols execute from the bench, telemetry lands in local storage first, and cloud sync is a separate, best-effort queue. A network outage never stops a test or loses a point.
- Offline-first
- Auto-discovery on the lab network
- HPPC and drive-cycle capable
- Drives Neware and other cyclers too
Lab software with a co-researcher in it.
Sentinel is the lab’s source of truth and its second pair of hands. It runs the fleet, versions every protocol, and puts an AI co-researcher next to the evidence: one that designs experiments, watches every channel, and investigates whatever drifts.
- 01
Operate the whole fleet
Vendor-agnostic drivers run Eye and incumbent cyclers from one place, with live floor telemetry and audit trails on every run.
- 02
Protocols you can trust a year later
Immutable versioning with history, branch and restore. Every run pins the exact version it executed.
- 03
Intent → DOE → protocol
State the decision, hypotheses, and constraints. The co-researcher quantifies factors and levels, builds the condition matrix, and emits one executable protocol per chemistry and condition.
- 04
Physics before cells
Chemistry-specific voltage windows from OCV–SOC curves, a per-protocol validation checklist, and electrochemical simulation before anything touches a cell.
- 05
Guardrails on every running test
Deterministic detectors watch thermal, voltage, and replicate divergence and open an investigation the moment something drifts.
- 06
Investigations, not chats
Evidence, transparent code, ranked hypotheses, and a human decision, saved as a recipe that reruns on the next dataset.
- 07
Traceable, cell ID to report
Auto-generated post-test reports with full lineage. Asset allocation matches each condition to available channels for the campaign’s duration.
- 08
Ask the lab anything
The Lab Assistant answers over live dashboard data: what is blocking the lab, which cyclers are offline, what changed overnight.
Machine learning that calls cycle life in weeks, from tens of cells.
Sentinel+ is the machine-learning layer of the platform. Its models read the first hundred cycles of a cell, call its life with a calibrated interval, and abstain when a cell is outside what they have learned. The channel moves on to the next variant months sooner, and every cell you run trains the model further.
A forecast that knows its own limits.
The Sentinel+ model forecasts cycle life from early cycles and ships every number with a calibrated interval that tightens as more cycles are observed. Illustrative NMC cell · 25 °C · 1C.
- Knee estimate
- 1,180 cycles
- 90% interval
- 961 – 1,399
- Interval width
- ±219
- Intervals are conformal: a 90% interval is designed to cover 90% of held-out cells. Coverage is reported with every ML model version, not assumed.
- Benchmarked on the public Severson et al. dataset (Nature Energy, 2019). Validated within chemistry and protocol family; re-measured on your cells in a pilot.
- Outside the validated envelope, Sentinel+ returns an actionable refusal and suggests the anchor condition that would extend it.
Your chemistry, your fleet · how the model learns
- Day one0 cells
Automate with guardrails from the start. Every cycle you run becomes training data for the model.
- ~10cells to end of life
First useful accuracy, verified on public nickel-rich sets.
- ~25cells
About 2.5× tighter than guessing the fleet average.
- ~90cells
Benchmark-grade: about 9% error on Severson 2019 LFP.
Owning the cycler, the protocol, and the model closes the loop.
- 01
Eye
Measures at 24-bit and 1 kHz with per-channel telemetry. Clean, dense, poolable data by construction.
- 02
Sentinel
Turns intent into versioned protocols and runs them on any cycler. Standardized capture, guardrails built in.
- 03
Sentinel+
Machine learning that trains on every cycle, calls life early, and orders the next test. The next DOE starts sharper.