The Science — Crush Dynamics Fermentation

Summerland, BC · Patented bioconversion process · FoodTech 500

Crush Dynamics fermenters convert grape pomace and other agricultural side-streams into premium multi-functional ingredients by activating polyphenols and fibres. The biology is non-trivial: feedstock varies batch-to-batch, microbial communities shift through the fermentation, and the desired endpoint sits inside a narrow window of cycle time, yield, and quality. Today operators steer the process from experience; this project gives them a co-pilot.

What we will learn (and what becomes scientifically novel)

A baseline of fermentation truth

Three years of historical batch data — recipe, feedstock lot, operator, lab analytics, energy submeters — consolidated into a single labelled dataset with documented baseline KPIs (cycle time, yield, batch failure rate, energy per kilogram). This is the first time CDI's process is quantified at this depth.

Deliverables: Baseline KPI specification (doc 07) · Three-year historical dataset, cleaned and harmonized · Documented baseline performance against which AI improvements are measured.

Pilot-trial design of experiments

A controlled DOE campaign at pilot scale to span the operating envelope, deliberately probe the failure modes, and produce labelled data points the soft-sensor and digital twin can train on. The DOE is engineered to maximize information per batch — every campaign run earns its place.

Deliverables: Pilot Trial DOE specification (doc 08) · Labelled pilot dataset across designed factor combinations · Quantified noise floor for every measured signal.

Soft-sensor validation against gold-standard analytics

The neural-network soft-sensors are only useful if they agree with the lab. We build a side-by-side comparison — model inference vs. wet-chem analytics — across pilot and industrial batches, and publish the validation envelope (where it works, where it doesn't, what to fall back to).

Deliverables: M06 model validation report · Soft-sensor accuracy bounds per analyte and operating regime · Documented failure modes and operator escalation criteria.

Industrial-scale demonstration with measured KPI lift

The proof point: a months-long industrial campaign with the AI in the loop, measured against the documented baseline. Yield, cycle-time, batch-failure, and energy consumption tracked daily, attributable to specific control actions. This is what we publish.

Deliverables: M09 industrial demo run · M10 performance validation report with statistically significant KPI deltas · Process audit trail tying every KPI improvement to AI control actions.

Where this fits in the broader fermentation science landscape

Most fermentation R&D in the agricultural-side-stream space is bench- or pilot-scale. The hard science problem is closing the loop at industrial scale — where feedstock variability, sensor drift, operator turnover, and recipe diversity stack up. By coupling CDI's industrial-scale fermenters to a hybrid digital twin and an MPC controller, we are testing a hypothesis that has only ever been partially evaluated in the literature: that supervisory AI can deliver measurable, sustained KPI improvement on a real production floor without disrupting the underlying process.

Publication of the validation results — including the negative or boundary findings — is part of the project commitments. The blog and reporting sections of the team site track that work as it lands.

See also

• The Technology — A47's AI/software stack that consumes this science.
• The Collaboration — how the science and tech teams pair to validate each other's rigour.
• Glossary — definitions for fermentation terms (SCI section), KPIs, validation methodology.