Configuration

The GRAFT is configured through YAML files in app/configs/ and shared constants in app/configs/constants.py. No code changes are needed when tuning model parameters, switching data sources, or adjusting thresholds — edit the relevant file and restart the application.

Config File Overview

File	Purpose
database_config.yaml	WebSocket endpoint and SQLite database path
LinearRegression_config_NiceGUI.yaml	Linear regression model and evaluation settings
LSTM_config_NiceGUI.yaml	LSTM architecture, features, training, and evaluation settings
XGBoost_config_NiceGUI.yaml	XGBoost hyperparameters, features, and evaluation settings
constants.py	Runtime constants, directory paths, labels — set once, used everywhere

database_config.yaml

Controls how the application connects to the live data source and where the local results database lives.

# WebSocket configuration
websocket:
  # WebSocket endpoint for streaming experiment data
  uri: wss://maple.containers.wur.nl/ws

  # Base SQL query used by the WebSocket backend
  base_query: SELECT * FROM public.pioreactor

# Local database configuration
database:
  # SQLite database file used for metrics and predictions
  path: metrics.db

websocket.uri — The WebSocket server to connect to. Update this if your Pioreactor instance moves to a different host.

websocket.base_query — The SQL query sent to the WebSocket backend. The application appends WHERE clauses for experiment filtering and time pagination automatically — do not add these manually.

database.path — Path to the local SQLite file where prediction results and evaluation metrics are stored. Relative paths are resolved from the project root.

LinearRegression_config_NiceGUI.yaml

data:
  id_column: entity
  metric_column: metric
  plot_history_minutes: 20
  target_column: optical_density
  time_column: time

database:
  path: "metrics.db"

forecast:
  horizon_forecast: 15
  horizon_minutes: 1

training:
  window_train: 300
  window_test: 50

lags:
  number: 15

model:
  linear_regression_path: trained_models/linear_regression_model.pkl
  copy_X: True
  fit_intercept: True
  n_jobs: 1

preprocessing:
  fill_method: bfill_ffill
  resample_freq: 1min
  value_column: value
  feature_columns:
    - optical_density

db_config: database_config.yaml

evaluation:
  max_horizon: 6
  rolling_window_size: 30
  horizon_unit: "min"
  min_points_for_correlation: 20
  min_variance_for_correlation: 1e-4

Key Parameters

Parameter	Default	Description
forecast.horizon_forecast	15	Number of future time steps to predict
forecast.horizon_minutes	1	Duration of each time step in minutes
training.window_train	300	Number of most recent rows used for training
training.window_test	50	Number of rows held out for testing
lags.number	15	Number of lag features added to the feature matrix
preprocessing.resample_freq	1min	Resampling frequency applied before training
evaluation.rolling_window_size	30	Rolling window of predictions used for metric computation
evaluation.min_points_for_correlation	20	Minimum matched points before computing R² / Spearman

LSTM_config_NiceGUI.yaml

data:
  id_column: entity
  metric_column: metric
  plot_history_minutes: 20
  target_column: optical_density
  time_column: time

database:
  path: "metrics.db"

forecast:
  horizon_forecast: 15
  horizon_minutes: 1

lags:
  number: 60

gompertz:
  enabled: true
  adaptive_window: 300
  refit_every: 10
  min_points: 20
  use_residual_learning: true

model:
  sequence_length: 60
  lstm_path: trained_models/lstm_model.keras
  lstm_units: 100
  dropout: 0.3
  epochs: 5
  batch_size: 32
  optimizer: "adam"
  loss: "mae"
  max_rows: 10000
  patience: 20
  scale_target: false

preprocessing:
  fill_method: bfill_ffill
  use_log: true
  resample_freq: 1min
  value_column: value

feature_columns:
  - optical_density
  - temperature
  - agitation_rate
  - optical_density_roll3
  - optical_density_roll5
  - optical_density_roll10
  - optical_density_std3
  - optical_density_std5
  - optical_density_diff1
  - optical_density_diff_roll5
  - experiment_elapsed_min
  - tod_sin
  - tod_cos

training:
  window_train: 500
  window_test: 50

evaluation:
  max_horizon: 6
  rolling_window_size: 30
  horizon_unit: "min"
  min_points_for_correlation: 20
  min_variance_for_correlation: 1e-5

Key Parameters

Parameter	Default	Description
model.sequence_length	60	Look-back window (number of time steps fed into the LSTM)
model.lstm_units	100	Number of LSTM units per layer
model.dropout	0.3	Dropout rate applied during training
model.epochs	5	Maximum training epochs (early stopping via patience)
model.patience	20	Early stopping patience in epochs
model.scale_target	false	Whether to normalise the target variable before training
preprocessing.use_log	true	Apply log transform to optical density before training
gompertz.enabled	true	Whether to fit a Gompertz residual correction on top of LSTM
gompertz.adaptive_window	300	Recent points used for Gompertz fitting
gompertz.refit_every	10	Refit the Gompertz model every N predictions
lags.number	60	Number of lag features added to the input

Feature Columns

The LSTM uses 13 engineered features derived from raw sensor data:

Feature	Description
optical_density	Raw OD reading
temperature	Temperature (°C)
agitation_rate	Agitation speed (RPM)
optical_density_roll3/5/10	Rolling mean over 3, 5, 10 steps
optical_density_std3/5	Rolling standard deviation over 3, 5 steps
optical_density_diff1	First-order difference
optical_density_diff_roll5	Rolling mean of first-order differences
experiment_elapsed_min	Minutes since experiment start
tod_sin / tod_cos	Sine and cosine encoding of time-of-day

XGBoost_config_NiceGUI.yaml

db_config: database_config.yaml

forecast:
  horizon_forecast: 15
  horizon_minutes: 1

lags:
  number: 60

model:
  learning_rate: 0.01
  max_depth: 6
  n_estimators: 100
  objective: reg:squarederror
  tree_method: hist
  subsample: 0.8
  colsample_bytree: 0.8
  reg_lambda: 1.0

preprocessing:
  fill_method: bfill_ffill
  use_log: true
  resample_freq: 1min
  value_column: value

feature_columns:
  - optical_density
  - temperature
  - agitation_rate
  - optical_density_roll3
  - optical_density_roll5
  - optical_density_roll10
  - optical_density_std3
  - optical_density_std5
  - optical_density_diff1
  - optical_density_diff_roll5
  - experiment_elapsed_min
  - tod_sin
  - tod_cos

training:
  window_train: 500
  window_test: 50

data:
  id_column: entity
  metric_column: metric
  plot_history_minutes: 20
  target_column: optical_density
  time_column: time

database:
  path: "metrics.db"

gompertz:
  enabled: true
  adaptive_window: 300
  refit_every: 10
  min_points: 20
  use_residual_learning: true

evaluation:
  max_horizon: 6
  rolling_window_size: 30
  horizon_unit: "min"
  min_points_for_correlation: 20
  min_variance_for_correlation: 1e-4

Key Parameters

Parameter	Default	Description
model.n_estimators	100	Number of boosting rounds
model.max_depth	6	Maximum tree depth — increase for more complex patterns
model.learning_rate	0.01	Step size shrinkage to prevent overfitting
model.subsample	0.8	Fraction of rows sampled per tree
model.colsample_bytree	0.8	Fraction of features sampled per tree
model.reg_lambda	1.0	L2 regularisation term
gompertz.enabled	true	Fit a Gompertz residual correction on top of XGBoost
preprocessing.use_log	true	Apply log transform to optical density

XGBoost uses the same 13 engineered feature columns as the LSTM — see the LSTM feature table above.

constants.py

Defines application-wide runtime constants, directory paths, and display labels. These are set once and used across all modules.

Runtime Constants

Constant	Default	Description
RUN_INTERVAL	180	Seconds between automatic prediction runs
MAX_PLOT_POINTS	1000	Maximum data points rendered per chart
AVERAGE_FREQ	"1min"	Default resampling frequency for plots

Directory Paths

Constant	Path	Description
PACKAGE_DIR	Project root	Base directory for config file resolution
USER_DATA_DIR	~/.graft/	User-writable runtime data directory
MODEL_DIR	~/.graft/trained_models/	Saved model artefacts

Subdirectories results/, logs/, data/, and trained_models/ are created automatically inside USER_DATA_DIR on first run.

Labels

METRIC_LABELS maps internal column names to human-readable axis labels:

Key	Display Label
optical_density	Optical Density (OD)
growth_rate	Growth Rate (1/h)
agitation_rate	Agitation Rate (RPM)
temperature	Temperature (°C)

MODEL_METRIC_LABELS maps evaluation metric keys to display labels:

Key	Display Label
mse	Mean Squared Error (MSE)
rmse	Root Mean Squared Error (RMSE)
r2	R² Score
spearman	Spearman Correlation
mae	Mean Absolute Error (MAE)
mape	MAPE (%)
direction_accuracy	Direction Accuracy (%)
smape	sMAPE (%)
max_error	Maximum Error

Changing Parameters

You do not need to edit YAML files directly. All model parameters — training window sizes, forecast horizons, hyperparameters, and preprocessing settings — can be adjusted from within the dashboard itself on the Model Control page.

1. Open the dashboard at http://localhost:8080
2. Navigate to the Model Control page
3. Adjust the parameters using the input controls

4. Click Run to apply the new settings and retrain or re-run the model

The YAML files in app/configs/ serve as the default values loaded on startup. Any changes made in the app take effect immediately for that session.