Skip to content

Strategy 4: Feedforward Prediction

Overview

Use a fast probe on a small sample to predict the parameters for the full batch.

Feedforward prediction combines speed and accuracy: take a quick measurement on a tiny sample, use a pre-trained model to predict the optimal parameters, then execute the main task at high speed without further iteration.


When to Use

  • Speed is critical (fast throughput required)
  • Accuracy is high (±1–2% target)
  • Value is high (expensive reagents; want to minimize waste)
  • Fast characterization tool available (rheometer, spectrophotometer, analyzer)
  • Model is well-validated (trained on many materials; generalizes well)
  • Material varies batch-to-batch (but predictably, based on easy-to-measure properties)

Common applications: - Pharmaceutical formulation (dose 10 mg of API quickly and accurately) - High-value chemical synthesis (use 1 mL to predict parameters for 100 mL) - Cosmetic/food manufacturing (characterize ingredients, predict mixing times) - Bioprocessing (measure cell culture viscosity in a 2 mL sample, predict fermentation parameters)


How It Works

Step 1: Probe the Sample

Take a small, representative sample (~1–10 mL for liquids/semi-solids; ~1–10 g for powders).

Measurements (30 seconds – 5 minutes per sample):

  • Rheometer: Measure viscosity, yield stress, shear-thinning index
  • Particle analyzer: Measure size distribution, density
  • Spectroscopy: Measure concentration, purity
  • Moisture analyzer: Measure water content
  • Flowability meter: Measure Hausner ratio, angle of repose

Step 2: Apply a Predictive Model

Use the sample measurements as inputs to a model that predicts the main-task parameters:

\[ \text{parameters} = f(\text{sample\_measurements}) \]

Model types:

  • Linear regression: flow_time = 0.5 + 0.1 × viscosity
  • Polynomial: pressure = 100 + 2 × τ_y + 0.05 × τ_y^2
  • Machine learning: Random forest, neural network trained on 20–50 past materials
  • Physics-based: Navier-Stokes solver for particle-size-dependent powder flow

Step 3: Execute at Predicted Parameters

Program the robot/pump with the predicted parameters and execute the main task (dispense, extrude, fill, handle) in one pass, no feedback loops.

Duration: 5–60 seconds (material-dependent)

Step 4: (Optional) Verify & Update

After main execution, measure the outcome: - If within tolerance: store as validation data - If outside tolerance: retrain model or escalate to real-time correction


Advantages

Aspect Benefit
Speed Probing takes 1–5 min, but main task is one pass (~30 sec). Fast for batches.
Accuracy Combines measurement (±1%) with predictive model → ±1–2% overall
Robustness Model learned from diverse materials; generalizes to new ones
Minimal waste Small probe sample (1–10 mL) vs. iterative trials (50+ mL)
Value preservation Best for expensive reagents or fragile cell cultures

Limitations

Challenge Issue
Model dependence If model is wrong or trained on different domain, predictions fail
Probe representativeness Tiny sample may not represent batch (stratification, settling)
Equipment cost Rheometer, analyzer, etc. are expensive (€10k–€100k)
Validation overhead Model must be trained and validated on many materials upfront
Generalization gap Model trained on water-based liquids may fail on organic solvents

Real-World Examples

Example 1: Viscosity-Dependent Liquid Dispensing

Task: Dispense 10 mL of various solvent mixtures to a reaction vessel, 50 times per day

Probe Protocol (2 minutes): 1. Withdraw 2 mL of the reaction mixture 2. Insert into Anton Paar MCR 302 rheometer 3. Measure viscosity at 25°C, shear rate 100 s⁻¹ 4. Read result: η = 3.2 cP

Model:

Trained on 30 different solvent mixtures
Linear fit: aspirate_time (s) = 0.05 + 0.012 × viscosity (cP)
Predicted aspirate_time = 0.05 + 0.012 × 3.2 = 0.088 s ✓

Execution:

Set syringe pump: aspirate 10 mL at 0.088 s air-gap time
Dispense → bottle is filled
Time: 30 seconds, accuracy ±1.5%, waste: 2 mL (probe only)

Daily workflow: - 9:00 AM: probe viscosity (2 min) + dispense 50 batches (25 min) = 27 min total - Alternative (iterative optimization): 30 min probe + calibration + 50 batches = 90 min - Time saved: 63 minutes per day


Example 2: Powder Flowability-Dependent Dosing

Task: Dispense 5.0 g of pharmaceutical powders from 10 different suppliers, each uniquely

Probe Protocol (3 minutes per powder): 1. Measure 10 g of powder 2. Use powder rheometer (FT4 Freeman Technology) to measure: - Basic Flowability Energy (BFE) - Stability Index (SI) - Hausner Ratio 3. Results: BFE = 250 mJ, SI = 120 mJ, HR = 1.18

Model (trained on 20 pharmaceutical powders):

Decision tree / random forest:
IF BFE < 300 and HR < 1.15:
  → Use auger speed 35 rpm, high confidence ✓
ELSE IF BFE 300–500 and HR 1.15–1.25:
  → Use auger speed 28 rpm with real-time weight feedback
ELSE IF BFE > 500 or HR > 1.3:
  → Use fluidized-bed feeder, model unreliable; fall back to iterative optimization

Prediction: Supplier A powder → use auger at 35 rpm

Execution: Dispense 5.0 g → actual weight 4.98 g ✓

Time savings: - Probe: 3 min per powder type (once) - Per-batch dispensing: 30 seconds (one-pass) - Alternative (iterative): 15–20 min per batch - For 100 batches: 1 probe (3 min) + 50 min total vs. 25 hours total


Example 3: Machine Learning Model for Extrusion

Task: 3D print with various bio-inks (cells + matrix), each with different viscosity and yield stress

Probe Protocol (4 minutes): 1. Rheometer sweep on 0.1 mL sample: - Viscosity at γ̇ = 1 s⁻¹: η₁ = 2.5 Pa·s - Viscosity at γ̇ = 100 s⁻¹: η₁₀₀ = 0.8 Pa·s - Yield stress (oscillation): τ_y = 15 Pa - Temperature drift (repeat at 30 min): Δη = 0.1 Pa·s

  1. Feature vector: [2.5, 0.8, 15, 0.1]

Neural Network Model:

Input: [η₁, η₁₀₀, τ_y, ΔT]
Hidden layers: [64, 32, 16]
Output: [extrusion_pressure_kPa, print_speed_mm/s, nozzle_temp_°C]

Trained on 50 bio-inks with known optimal parameters
Validation accuracy: ±2.5% on test set of 5 held-out inks

Prediction:

Input: [2.5, 0.8, 15, 0.1]
Output: [250 kPa, 5 mm/s, 37 °C]

Execution: 3D print 10 mL scaffold at 250 kPa, 5 mm/s → completes in 20 minutes, cell viability ±2% vs. manual optimization

Time savings: 4 min probe + 20 min print vs. 90 min manual tuning + print


Model Training Workflow

Offline (One-Time)

  1. Collect training data (weeks of lab work):
  2. Test 20–50 representative materials
  3. For each, measure properties (viscosity, size, etc.)
  4. Measure optimal handling parameters (via iterative optimization)
  5. Record failures and edge cases

  6. Build dataset:

    Material | Viscosity | Yield Stress | Optimal Speed | Optimal Pressure
    ---------|-----------|--------------|---------------|------------------
    A        | 1.2 cP    | 0 Pa         | 8 mm/s        | N/A
    B        | 50 cP     | 10 Pa        | 4 mm/s        | N/A
    C        | 500 cP    | 50 Pa        | 2 mm/s        | 150 kPa
    ...      | ...       | ...          | ...           | ...
    

  7. Train model:

  8. Linear regression (if clear trends): speed = f(viscosity, τ_y)
  9. Neural network (if complex): train on 80%, validate on 20%
  10. Random forest (if non-linear): good generalization

  11. Validate on held-out materials:

  12. Test model on 3–5 materials not in training set
  13. Verify accuracy is within specification (e.g., ±3%)

Online (Per-Use)

  1. Probe sample (2–5 min)
  2. Predict parameters using trained model (< 1 sec)
  3. Execute (30 sec – 10 min)
  4. (Optional) Log outcome for future model retraining

Model Maintenance

  • Seasonal retraining: Every 3 months, retrain on new data collected since last training
  • Domain shift detection: If new materials are outside training domain, trigger human review
  • Degradation monitoring: Track prediction errors over time; if trend is upward, retrain

Implementation Checklist

  • [ ] Identify which material property to measure (viscosity, particle size, flow energy, etc.)
  • [ ] Select a probe device (rheometer, analyzer, scale)
  • [ ] Collect training data: 20–50 materials, properties + optimal parameters
  • [ ] Split data: 80% training, 20% validation
  • [ ] Train model (linear, polynomial, tree-based, or neural network)
  • [ ] Validate on held-out test set
  • [ ] Implement prediction model in robot control software
  • [ ] Test on 5–10 production materials
  • [ ] Document model inputs, outputs, and expected accuracy range
  • [ ] Set up retraining schedule (quarterly or as-needed)
  • [ ] Create fallback logic (if prediction confidence is low, use real-time correction instead)

Comparison with Other Strategies

Strategy Probe Time Main Task Time Total Time Accuracy Model Requirement
Characterize, then act 2–5 min 30 sec 2.5 min ±5% First-principles
Iterative optimization — 15–30 min 15–30 min ±1% Trial-and-error
Real-time correction — 5–30 sec 5–30 sec ±2% Sensor + control
Feedforward prediction 2–5 min 30 sec – 10 min 2.5–15 min ±2% ML model
Simulation generalization — 30 sec 30 sec ±5% Simulator + ML

References

  • Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
  • Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning (2nd ed.). Springer.
  • Rastegardoost et al. (2026). Feedforward prediction for autonomous material handling. Lab on a Chip.
  • Anton Paar. Application Report: Rheology in Liquid Dispensing. (Free online resource)