Secure Access Required
Select devices to register
Your bioreactor needs to be connected to the AI cloud service. Download and run the setup wizard on the computer running D2MS.
Download Setup Wizard (54 MB)D:\D2MS Native\OAJ\D2MS_Gateway\TunnelSetup.exe and select "Run as administrator"Requirements: Windows 10/11 (64-bit), Administrator privileges
您的生物反应器需要连接到AI云服务。请在运行D2MS的电脑上下载并运行安装向导。
下载安装向导 (54 MB)D:\D2MS Native\OAJ\D2MS_Gateway\TunnelSetup.exe 选择"以管理员身份运行"系统要求： Windows 10/11（64位），管理员权限
Need help? Email: support@tjxbio.com 需要帮助？邮箱：support@tjxbio.com
Register your bioreactor
Ownership Verification - Step 1 of 2
Please check the current Set Point (SP) value for:
--
Find this value in your D2MS Software and enter it below.
Ownership Verification - Step 2 of 2
Now, please change the Set Point (SP) for:
--
to a different value (must differ by more than 10%), then enter the new value below.
After changing the value in D2MS Software, enter it here.
Verification Successful!
Create a password and agree to the terms to complete registration.
End User License Agreement
Key Points:
| Parameter | PV | SP | Source | AUTO | PROFILE |
|---|
Select parameters to use as model inputs.
Comma-separated list (e.g., 60, 120, 240 for 1h, 2h, 4h)
Based on current configuration, you need:
Copy this reactor's config to another reactor, or load config from another reactor.
Click "Refresh Preview" to see the model architecture based on current configuration.
No predictions available. Configure and train the model first.
⚠️ Predictor not configured
Please configure the predictor first by selecting input features and output parameter.
Choose a run and time range to use for model training. The predictor requires sufficient data points within the selected range.
Each training creates a new model version. You can load a previous version to use for predictions.
Loading model history...
Compare model predictions against actual historical data to evaluate accuracy.
Previously run control loops. Click "Run" to start a loop with the same parameters.
All triggers (active and inactive). View fire history and manage triggers.
Previously run supervision tasks. Click "Rerun" to start a task with the same objective.
Historical experiments with high-frequency data logging. View graphs, download data, or delete experiments.
No bioreactors available.
Contact your administrator to request access.
Loading users...
New users are created with observer role (read-only access). You can promote them later.
Register a new bioreactor by verifying ownership through SP parameter source address.
Ownership Verification
Please enter the Source Address for the parameter:
--
Find this value in your D2MS Software, under Control:SP Source Specifications.
✓ Verification Successful!
The bioreactor has been verified and added to the system.
Loading users...
Configure the AI models used for chat and knowledge extraction. Changes take effect immediately for new conversations.
Language for system prompts and context. Independent of AI provider.
Leave empty to use environment variable (OPENAI_API_KEY)
Lightweight model recommended for frequent small calls
Leave empty to use environment variable (DEEPSEEK_API_KEY)
Manage LLM prompts for different AI providers. OpenAI prompts are in English, Moonshot prompts are in Chinese.
| Name | OpenAI Prompt (English) | Moonshot Prompt (Chinese) | Actions |
|---|---|---|---|
| Loading prompts... | |||
Executable analysis methods (Python modules). These are dynamically loaded and can be updated without server restart.
Formula documentation describing calculation steps. For reference and review.
Define expected ranges to validate results
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.
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 | Process variables (PVs) and setpoints (SPs) used as inputs to the model. Select parameters that influence your target output. |
| Output Parameter | The target variable to predict (e.g., titer, biomass). This is what the model learns to forecast. |
| Output as Input (AR) | Autoregressive mode - uses past output values as additional input features. Lag steps controls how many past values to include. |
| Model Type | GRU - Faster training, fewer parameters. LSTM - Better for very long sequences, more memory. |
| Avg Window (sec) | Downsamples training data by averaging points. Higher = less data, faster training. 30s (6x reduction) is a good default. |
| Min Sequences | Minimum training sequences required. Lower = less data needed but potentially less accurate. 50-200 typical. Directly affects data requirements. |
| Auto-Retrain (min) | Automatically retrain the model at this interval. Set to 0 to disable. Uses "Quick Train Epochs" for faster training. |
| Sequence (min) | How many minutes of history the model sees at each step. Longer = more context but more data needed. 60 min is typical. |
| Hidden Size | Number of neurons in each recurrent layer. Larger = more capacity but slower. 64-256 typical. |
| Layers | Number of stacked recurrent layers. 1-2 usually sufficient. More layers = more complex patterns but harder to train. |
| Dropout | Regularization to prevent overfitting. 0.1-0.3 typical. Higher if you see overfitting. |
| Learning Rate | How fast the model learns. 0.001 is a good default. Lower if training is unstable. |
| Batch Size | Samples processed before updating weights. 16-64 typical. Larger = faster but uses more memory. |
| Max Epochs | Maximum training iterations for manual "Train Now". Early stopping may end training sooner. |
| Quick Train Epochs | Epochs used for auto-retrain (fewer for faster background training). |
| Early Stop Patience | Stop training if loss doesn't improve for this many epochs. Saves time and prevents overfitting. |
| Horizons (min) | Comma-separated list of future times to predict. E.g., "60, 120, 240" predicts 1h, 2h, 4h ahead. Max horizon affects data requirements! |
| Log Predictions | Which horizons to log to the data history for later analysis and charting. |
Loading tips...