Add New Rule

Existing Rules

Add New Control Parameter

Existing Control Parameters

Current Predictions

Current Run

Phase History

Reference Run Candidates

Runs marked as golden candidates in D2MS Runs. Select which ones to use for training.

Input Features

Output Parameters (prediction targets)

Model Architecture

Pre-filled from short-horizon predictor config. Adjust independently — golden models train once on full runs, so larger networks are fine.

Training Parameters

Train Golden Model

Train a frozen model on the selected reference run(s). Configure features and architecture in the Configure tab. The short-horizon predictor (auto-retune) can be used alone, but the golden batch model requires a trained short-horizon predictor for Mia to evaluate short-term impacts.

Test Against Golden Reference

Score a historical run against the golden reference to see how well it matches.

Correction Rules

Parameter-specific correction rules guide the L6 Supervisor's adjustments. Auto-generated from the golden trajectory, editable by lab managers.

Model Templates

Browse published golden batch models or publish yours for other reactors.

Overview

The Titer Predictor uses a recurrent neural network (GRU/LSTM) to forecast future values based on historical process data. The model learns patterns from your bioreactor's time-series data and predicts values at configurable time horizons.

📊 Data Requirements Calculation

The minimum raw data points needed for training is calculated as:

Example: With Sequence=60, Max Horizon=480, Min Sequences=100, Avg Window=30s:
Downsampled = 60 + 480 + 100 = 640 points
Raw = 640 × 6 = 3,840 points (~53 hours at 5-second intervals)

Input Features

Data Configuration

Model Structure

Training Control

Prediction Horizons

💡 Tips for Reducing Data Requirements

• Reduce Min Sequences: Try 50 instead of 100 to halve the sequence requirement (may reduce accuracy)
• Reduce Max Horizon: Use 60, 120, 240 instead of 480 to cut requirements by ~40%
• Reduce Sequence Length: Try 30 instead of 60 if your process doesn't need long history
• Increase Averaging Window: 60s instead of 30s halves the raw data needed
• Use Predictor Management: Select a larger time range from historical runs