Protein Industries Canada proposal • Crush Dynamics + Atomic47 Labs

High-performance fermentation control, explained for everyone.

A fermentation explorer for everyone from board members to process engineers. Dig into the science, interrogate real run data, and run the models to understand how AI soft sensors, a hybrid digital twin, and model-predictive control turn a variable bioprocess into a measurable, auditable production system.

Project number in proposal: Ai26.10Working context: ai16-10 / Crush Dynamics PICAudience: non-technical + technical reviewers

Proposal targets at a glance

Targets are proposal objectives and must be validated against M01 baseline and later industrial demonstration data.

45 → 10–15day cycle target
≥95%yield target
<1%failure-rate target
~70%energy reduction target

The opportunity axis

The opportunity is not “AI for its own sake.” The opportunity is to turn a variable, slow, manually adjusted bioprocess into a predictable, measurable, auditable production system.

1. Waste becomes value

Grape pomace is fibre-rich and polyphenol-rich. Controlled fermentation can turn an underused by-product into high-value functional ingredients.

2. Time becomes capacity

Reducing cycle time from roughly 45 days toward 10–15 days increases effective capacity without building new tanks.

3. Data becomes control

Historical and pilot data become the ground truth for soft sensors, digital twins, and closed-loop control decisions.

Important framing for stakeholders: the spreadsheet and analysis included here demonstrate the evidence pattern and modelling workflow. The formal M01 milestone requires a complete three-year historical data package, CDI baseline KPI calculations, A47 re-computation, reconciliation within ±2% or a signed divergence statement, and formal acceptance.

Full science section

A plain-language and technical explanation of what is happening in the fermentation process, why the measurements matter, and how the AI control layer uses them.

What fermentation is doing here

Fermentation is a controlled biological conversion process. Microbial and enzymatic activity transforms grape pomace chemistry over time. The production goal is not simply to “let it ferment,” but to reach target acidity, quality, and consistency as quickly and reliably as possible.

Industrial fermentation is sensitive to oxygen transfer, temperature, substrate availability, moisture, pH, and mixing. Small differences early in the batch can compound into different final acidity, yield, and cycle time.

Why grape pomace is hard

Pomace is biologically rich but variable. Feedstock lots can differ by grape variety, harvest timing, moisture, skin/seed composition, sugar, phenolics, and microbial load. That variability is why a static recipe or manual checklist cannot always hold the process in the optimum zone.

TA: titratable acidity

TA is a key process-progress and quality KPI. It represents the acid content measured by titration. In this project, TA is treated as a central endpoint and state variable: the system needs to know where the batch is now, where it is heading, and when it will cross target.

TA(t) ≈ measured acid concentration
dTA/dt = instantaneous acidification rate
ETA-to-target ≈ (TA_target − TA_now) / predicted dTA/dt

pH versus acidity

pH measures hydrogen ion activity, while TA measures total titratable acid capacity. They are related but not interchangeable. A process can have similar pH but different TA depending on buffering and acid composition. That is why AI soft sensors should learn from multiple inputs rather than relying on pH alone.

Oxygen transfer and kLa

For aerobic or oxygen-influenced fermentation, cells and reactions depend on oxygen delivery. The oxygen transfer coefficient kLa summarizes how effectively oxygen moves from bubbles/headspace into the liquid/solid matrix.

OTR = kLa × (C* − C)
where OTR = oxygen transfer rate,
C* = saturation concentration,
C = current dissolved oxygen concentration

Oxygen uptake rate: OUR

OUR estimates how quickly the biological system consumes oxygen. A mismatch between oxygen transfer and oxygen demand can create stress, slow conversion, or quality drift.

Oxygen gap = OTR − OUR
If gap < 0: risk of oxygen limitation
If gap ≫ 0: possible energy waste / over-aeration

Temperature and kinetics

Temperature changes reaction rates and biological activity. Too low can slow fermentation; too high can stress organisms, shift pathways, or damage quality. The controller’s job is to manage temperature as a dynamic input, not a fixed background condition.

Rate ∝ exp(−Ea / RT)
Small temperature shifts can materially change reaction velocity.

Mixing and power

Mixing affects mass transfer, heat distribution, and homogeneity. More mixing is not always better: it may improve oxygen transfer but also consume energy or create shear. The control problem is a trade-off between biological performance and energy efficiency.

Agitation power ∝ N³D⁵
N = impeller speed, D = impeller diameter

Full science & algorithm documentation

The complete technical reference: fermentation kinetics, governing equations, dataset schema, data-cleaning algorithm, model fits, soft-sensor derivation, and a plain-English "Othertist" brief for non-technical reviewers.

Open Science & Algorithms page →

System architecture

The AI layer supervises the existing PLC/SCADA environment. It predicts, recommends, validates, and eventually updates setpoints inside safety and quality limits; it does not remove existing interlocks or operator authority.

FermenterpH • DO • temp • airflowmixing • lab TA • yield PLC / SCADAsafety interlocksoperator overrides Edge IPCinference <100 mslocal buffering AI Modelssoft sensorsanomaly detection Cloud Twintrainingsimulation Actuatorsair valves • mixer VFDtemperature jacket Model-Predictive Controlforecast → optimize → constrainvalidated setpoint updates Operator DashboardETA-to-target • alertstraceability • reporting

Soft sensor

Predicts unmeasured state variables such as TA, yield trajectory, and fermentation progress.

Digital twin

Combines mechanistic fermentation equations with ML residual correction.

MPC

Chooses setpoints that optimize future outcomes under constraints.

Anomaly detection

Flags sensor drift, contamination risks, unexpected kinetics, and actuator issues.

Evidence pack: raw spreadsheet + analysis

The uploaded spreadsheet is converted into cleaned tables and run-level metrics, then used to demonstrate how fermentation evidence becomes model-ready data.

18 runs • 316 rows • 92 quality flags

What the current evidence shows

The workbook contains run sheets with time, acidity, pH, Brix, ethanol, temperature, aeration, and volume measurements. The analysis transforms those sheets into long-format observations and run-level summaries.

18runs analyzed
316observations
23_978highest endpoint acidity

How this supports the proposal

This data package demonstrates the analysis pipeline required for M01 and later model training: inventory, cleaning, feature extraction, quality flagging, KPI computation, and stakeholder review.

For formal M01 completion, this same pattern must scale to the full three-year CDI dataset across in-scope tanks and data systems.

Run explorer

Chart controls

Interactive modelling sandbox

Use the controls to understand the economics and process dynamics behind the proposal. These are illustrative calculators, not final validated project KPIs.

Facility impact model

The proposal uses a representative 1,000 tonne/year facility and targets cycle-time reduction, yield uplift, failure reduction, and energy savings. Adjusting inputs shows sensitivity.

Acidification trajectory model

A simple process model shows how control quality changes the time to target acidity. Better control raises the effective acidification rate and reduces drift.

Soft-sensor algorithm

  1. Synchronize sensor and lab data by batch time.
  2. Engineer rates and lag features.
  3. Train model to predict TA and yield.
  4. Validate on held-out batches.
  5. Deploy to edge for real-time inference.

Digital twin algorithm

  1. Compute mechanistic estimates: OTR, OUR, kLa, temperature response.
  2. Estimate residual error with ML.
  3. Forecast future states under candidate setpoints.
  4. Compare forecast against biological and quality constraints.

MPC algorithm

  1. Read current state estimate.
  2. Generate candidate aeration/mixing/temperature actions.
  3. Simulate each action through the twin.
  4. Select lowest-cost safe trajectory.
  5. Send validated setpoint to PLC/SCADA.

M01: data readiness and baseline establishment

M01 is the foundation for credible before/after measurement. Without a clean baseline, later claims about cycle time, yield, energy, and failure reduction cannot be defended.

Inventory

Catalogue SCADA, lab, batch, energy, and maintenance sources.

Quality

Find gaps, drift, tag changes, and batch-label inconsistencies.

Package

Extract historical telemetry, lab records, metadata, energy records.

Compute KPIs

Cycle time, yield, failure rate, energy per kg.

Reconcile

CDI and A47 compare; resolve divergences over ±2%.

Acceptance means

Historical package ingests cleanly, data quality is sufficient, baseline KPI calculations converge or divergence is signed, facility access is committed, and formal acceptance is issued.

Open decisions before M01 starts

Yield denominator basis, energy submeter scope, lab data format, data redaction scope, and tank scope must be resolved before baseline work begins.

Proposal workplan and milestones

The project progresses from historical baseline to pilot data, AI model development, control integration, industrial-scale demonstration, and commercialization readiness.

Business case and commercialization

The proposal combines ingredient-product economics with a scalable AI platform model.

B2B ingredient sales

CDI and future platform users can sell more consistent premium fermented ingredients to food manufacturers seeking sustainable, clean-label, ESG-aligned supply.

SaaS licensing

Atomic47 can package the fermentation optimization stack as a configurable platform for processors using related fermentation processes.

Integration services

Large or complex facilities may need paid customization, data engineering, and industrial AI integration support.

Representative ROI from proposal

Estimated capital investmentCAD $700k–800k
Annual operating benefit~CAD $288k
Payback, 1,000 t/year facility2.5–3.0 years
Payback, ≥5,000 t/year facilities<18 months

Defensible differentiation

Closed-loop supervisory control, industrial PLC/SCADA integration, 40,000 L validation target, RSM-grounded performance targets, proprietary CDI process knowledge, and A47 AI control IP create a combined process-and-platform advantage.

Governance, security, and IP

The project is designed for a high-consequence food-manufacturing environment where data ownership, auditability, and operational boundaries matter.

Data security

AES-256 at rest, TLS 1.3 in transit, RBAC, MFA for admin access, audit logging, and Canadian sovereign cloud/edge infrastructure.

Human oversight

AI operates within validated safety and quality constraints with operator override and existing PLC/SCADA interlocks retained.

IP structure

CDI retains fermentation process and operational data background IP; A47 retains AI models, digital twin, MLOps, and platform foreground IP subject to agreed project use rights.

Downloads and review order

Use this package to brief non-technical stakeholders first, then let technical reviewers inspect the raw analysis and data.

Recommended review order

  1. Open this site for the story and interactive models.
  2. Read the audience brief for board-level context.
  3. Review the full technical analysis.
  4. Open the raw spreadsheet and derived CSV files.
  5. Use the proposal text to map claims to milestones.