By the time you see it, you've already lost the material
Formation and aging mean your quality signal arrives weeks after the upstream cause. Every event carries hidden committed cost.
Built for
Manufacturing EngineeringProcess EngineeringQuality EngineeringCell EngineeringProduction Leadership
The problem
Three problems we hear every time we talk to a plant team
Defect risk becomes expensive when plant teams can't turn scattered process, lot, equipment, and quality history into timely action.
Your data exists. It just doesn't talk to itself.
MES, historian, SPC, QMS, shift notes — the data is there. Getting it to line up around a specific defect event is the actual engineering problem.
Half the investigation is just finding the starting point.
Before your team can investigate, they have to reconstruct what happened across five systems. That's hours per event, every time.
Pilot motion
First signal in 5 days
Lychee scopes a tight first pilot — one line, one defect family, one KPI — so plant teams can evaluate fast.
0 days
From historian credentials to first prediction
Historian, MES, and lot context connected, validated, and surfacing risk on real lots.
0 weeks earlier
Than your current process catches them
Defects flagged from coating-and-drying signals before formation confirms — the gap where scrap and cycle time compound.
01
Ingest
Connect historian, MES, and lot context — AVEVA PI, Aspen InfoPlus.21, Siemens Opcenter, Ignition. No new data collection required.
02
Predict & Explain
Ranked-risk queue across lots and chemistries, with upstream driver attribution and calibrated confidence intervals. Every flag ships with an interval — not a score, an interval.
03
Close the Loop
Investigations feed back to the model. Each resolved case sharpens the next flag.
02 — From research wins to factory wins
Why battery manufacturers choose Lychee
Lychee beats the field’s published gold standard (Severson 2019, Nature Energy) — and outperforms standard ML across every Li-ion chemistry where generic foundation models like AWS Chronos break down on real, fragmented production data. Six public datasets, five labs, four chemistries (LFP, LCO, NMC, Na-ion) — one hybrid, no retraining per launch. For plant teams: defect risk surfaces weeks earlier, across every chemistry on your line.
Early defect detectionMIT · Stanford · SLAC benchmark
Spot defective cells in 100 cycles — weeks before downstream QC confirms it
Beats Severson 2019 (Nature Energy) on the field's primary cycle-life benchmark — 8.9% vs 9.1% published¹
- Dataset
- Severson primary (LFP, n=42, MIT-Stanford-SLAC)
- Task
- Cycle-life prediction · pooled b1+b2 alternating split
- Comparator
- Lychee hybrid vs. Severson 2019 published vs. physics-only baseline
8.9%
Lychee hybrid · median error¹
9.1%
Severson 2019 (Nature Energy)¹
24.7%
Physics-only baseline
Plant teams today learn a cell is defective from lifetime testing — months after coating, drying, and formation. Lychee reads the early-cycle signature: variance in how voltage-capacity curves warp between cycles 10 and 100 surfaces defect risk weeks before downstream QC confirms it. Validated on the canonical MIT-Stanford-SLAC benchmark — Lychee's hybrid edges the field's published Nature Energy result (8.9% vs 9.1%) and runs 15.8 percentage points tighter than physics-only baselines. Same hybrid now validated on six public benchmarks across four chemistries (LFP, LCO, NMC, Na-ion) and five labs.
“Using discharge voltage curves from early cycles yet to exhibit capacity degradation, we apply machine-learning tools to predict cycle life.”Severson et al. — Nature Energy, 2019 →
124 commercial Li-ion cells · A123 LFP chemistry · public benchmark
Built for battery physicsChronos · AWS
Lychee beats generic AI by up to 150ײ
Same hybrid across LFP, LCO, NMC where generic AI breaks down — no retraining per launch²
- Dataset
- Severson + CALCE + NASA + Oxford + BatteryLife · LFP, LCO, NMC, Na-ion across five labs
- Task
- Cycle-life prediction across chemistries · same model, no retuning
- Comparator
- Lychee hybrid vs. Chronos zero-shot (AWS)
5–12%
Lychee hybrid · median error across 5 benchmarks
400–760%
Chronos zero-shot · breaks down on 3 of 4 tested²
30–150×
Tighter than Chronos zero-shot²
Plant teams hear 'AI for batteries' from every vendor. Most pitch generic time-series models — the same class as AWS Chronos, the leading foundation model for forecasting. On battery degradation, Chronos breaks down at 400–760% error: it forecasts flat capacity and can't resolve the slow per-cycle drop where defects hide. Lychee uses the same architecture class — a senior Lychee advisor led Chronos at AWS — but adds physics grounding so the model holds where generic AI flatlines. Under 10% error on four of five tested benchmarks, across LFP, LCO, NMC, and Na-ion — five labs, same hybrid, no retraining per launch. For plant teams: ranked-risk cells weeks earlier, every chemistry on the line.
“Chronos models trained on a corpus of public datasets have comparable or, occasionally, superior zero-shot performance on new datasets, relative to methods trained specifically on them.”Ansari et al. — Chronos paper, AWS 2024 →
Lead Chronos architect (AWS) advises Lychee Labs
Team and advisors from
Reads from
AVEVA PI System · Siemens Opcenter · Aspen InfoPlus.21 · Ignition by Inductive Automation · custom historian and CSV exports
Data security
Process data stays in your environment. On-premises and private-cloud deployments supported.
01 — Economic stakes
Why earlier visibility matters
Scrap, delayed visibility, and small yield movements become material at factory scale.
~€10M/yr
Per 1-point gain
At 40 GWh/year scale, even a 1-point improvement can be worth about this much annually.
70-80%
Early-ramp scrap
Hazardous waste and lost value can reach this level in early battery-factory ramp-up.
3 wks
Cycle-time delay
Formation and aging mean defects are often recognized only after weeks of added time and value.
32%
Cell-cost impact
Battery quality failures do not just hit inspection. They hit a large share of total cell cost.
Delayed discovery is a capital-efficiency problem.
Mineral efficiency
Yield improvement is also a critical-mineral-efficiency lever. Each percentage point of yield at 40 GWh scale reduces lithium, cobalt, nickel, and graphite waste per cell — directly relevant to EU Battery Regulation traceability and UK supply-chain competitiveness goals.
The workflow panel
What plant teams see
A live operations view of the line: ranked-risk queue across lots and chemistries, drift-flag chart with confidence and lead-time, likely upstream drivers, and a recommended next step. Investigations open with structure, not a blank screen.
Lychee Labs
Live
Line A·LFP / Graphite·Operator: J. Walsh
Cycle 42 / 100
Capacity drift · LOT-0427
Flagged 3 weeks before formation confirms
Confidence
0.84
Detected
Cycle 30
Recommended
Quarantine LOT-0427 batch
Estimated value
€41k saved
Risk queue · 3 active
Updated 14:42
LOT-0427LFP
Formation outlier risk
High
0.91
LOT-0419LFP
Coating drift
Elevated
0.74
LOT-0398NMC
Drying variance
Monitor
0.53
Likely upstream drivers
Pump fluctuation
Slot-die line A2
Slurry viscosity shift
Mixer 03
Coating head imbalance
Line A · 14:32
Why now
Global gigafactory expansion. EU and China regulation tightening.
Manufacturing capacity scaling in the UK, Europe, and China simultaneously — with regulatory requirements tightening across all three. Earlier defect visibility is the lever that compounds across every dimension.
Global gigafactory expansion
CATL, BYD, Envision AESC, and Tata Agratas are all ramping simultaneously — Erfurt, Sunderland, Shenzhen. Yield and ramp speed decide whether they reach nameplate capacity.
EU Battery Regulation, 2027
Chapter VII obligations require per-cell process and lifecycle data most plants don't yet capture. Affects every European manufacturer — and every Chinese producer exporting to the EU.
Na-ion scale-up
CATL, BYD, and HKUST-GZ are scaling Na-ion to market. New chemistry means new quality profiles. Lychee's 4.3% APE on the BatteryLife Na-ion benchmark is the first published validation on this architecture.³
Critical-mineral efficiency
Each yield point at 40 GWh reduces lithium, cobalt, nickel, and graphite waste. Supply-chain competitiveness from Sheffield to Shenzhen.
Lychee LabsBattery process flow
Lychee learns from upstream signals at coating, drying, and calendaring — flagging defect risk weeks before formation testing surfaces the same loss.
Lychee detection window
Confirms
01
Coating
Slot-die thickness, pump speed
02
Drying
Solvent residual, web tension
03
Calendaring
Density, thickness variance
04
Formation
Capacity, coulombic efficiency
~3 weeks earlier
Why upstream
Coating, drying, and calendaring carry the earliest physical evidence of the defects formation later confirms.
Why it compounds
Every cycle past detection adds material, labor, and downstream cost — the cost of finding it later.
“We started Lychee because every battery engineer we talked to lost weeks to investigations that should've taken hours. If that's your reality, I want to hear from you.”
Linda Hong Cheng — Founder & CEO
Next step
Working on battery yield, scrap, or diagnosis speed?
Scope a manufacturing pilot
Start with one line, one defect family, and one KPI.
Scope A PilotDownload 1-pager
Share the short overview with technical, manufacturing, or investment stakeholders.
Download 1-pager1 Severson 2019 (Nature Energy) primary split: Lychee’s 8.9% median APE beats the published 9.1%. On the secondary split (held-out novel protocols), Severson 2019 reaches 8.6% with the full feature set; Lychee runs 11.9% on that harder generalization regime — reported transparently on /benchmarks.
230–150× tighter than Chronos zero-shot (AWS) applies to the three of four datasets where Chronos was tested — Severson primary, Severson secondary, and CALCE CS2. NASA PCoE (n=4) shows small-sample variance in Chronos’s median; Oxford NMC was not run against Chronos in this benchmark cycle.
3 BatteryLife Na-ion benchmark (HKUST-GZ, n=31 cells, 3 charge protocols, LOO-CV across cells): Lychee median APE 4.3% — first known published result on a public Na-ion cycle-life benchmark. Methodology and dataset details on /benchmarks.