Method

The test setup

We didn't test with water. We wanted real-world conditions:

• Fluid: 30% glucose syrup — viscous, representative of actual bioprocess feed media
• Backpressure simulation: Flow disturber installed on the discharge line to mimic real-world line resistance (filters, restrictors, bioreactor headpressure)
• Pump: QB PUMP — peristaltic pump module
• Scales: QB MULTISCALE unit — source scale (0.1 g resolution) for loss-in-weight control, destination scale (5 g resolution) for independent verification
• Controller: QB EDGE running QB Control software
• Service: pre-built gravimetric feeding service template with PID-controlled mass flow

Experiment scope:

• 6 batches with full process data (4.4M+ data points)
• Flow rates tested: 1, 2, 5, 10, 20 g/min
• Batch sizes: 100–500 g
• Longest continuous run: 8.5 hours
• 249 PID parameter changes in the tuning campaign
• One deliberate "worst case" batch with pump control instability

Challenge 2

PID tuning with viscous media and backpressure

The flow disturber and glucose syrup changed the pump's response characteristics compared to water. PID tuning had to account for this.

Tuning campaign results (249 parameter changes):

• Kp = 5.0 → pump spiked to 87 RPM (instability boundary)
• Ki = 1.5 → integral windup, sluggish response
• Ki = 0 → stable, but 22% steady-state bias at 5 g/min
• Eventually → breakthrough result

A small integral term eliminated the steady-state error caused by the viscous media and backpressure — without causing windup.

Challenge 4

Multi-setpoint recipe execution

Real processes don't run at a single flow rate. Perfusion protocols adjust feed based on cell density. Fed-batch processes ramp over days.

Batch 6 — sequential multi-flow validation:

All transitions clean. PID reinitialized properly per phase. No cumulative errors carried between runs. 400+ grams dispensed in 55 minutes — through viscous glucose syrup with simulated backpressure.

Here's where it gets really interesting. After each batch, we exported the complete process data — mass flow, pump RPM, scale weight, PID parameters, audit trail — directly from QB Control's Backup & Export module. A few clicks, and we had the full dataset ready for analysis.

We then fed those 4.4 million data points into an AI analytics pipeline that:

• Correlated PID parameter changes with flow accuracy across all 249 tuning adjustments
• Identified the optimal Kp/Ki/Deadband/FilterTC combination per flow rate regime
• Detected the Ki=0.12 breakthrough — something that took 5 batches to find manually, but the AI flagged from cross-batch pattern analysis
• Generated per-phase accuracy metrics for the multi-setpoint batch, with flow bias breakdowns at each setpoint
• Produced recommended parameter tables covering 0.5–20+ g/min with evidence traceability back to specific batches

The traditional approach: run a batch, manually review trends, adjust one parameter, repeat. For 249 parameter changes, that's weeks of painstaking review.

Our approach: Run batches → export from Backup & Export → AI analytics → optimized recommendations. What would take days of manual review was done in minutes.

The result: a validated parameter set (Kp=0.8, Ki=0.12, Deadband=0.2) with quantified accuracy expectations for every flow rate — ready to deploy as service defaults.

This is the feedback loop that makes software-defined processing practical: rapid experimentation + instant data export + AI-driven analysis = tuning cycles measured in hours, not weeks.

When setup takes an hour instead of weeks, and PID optimization takes minutes instead of days, you can afford to run 6 experimental batches, deliberately break things, and build a validated knowledge base for your specific fluid and hardware combination.