3. Initial design¶

Before the optimizer has any results to learn from, it needs a first set of experiments. This page generates that set, queues it so you can record the results, and exports it.

A Latin hypercube design of 10 points with seed 42. Stage queues the points for recording; CSV saves them.

Choose a method¶

Most sessions start with a space-filling design, which spreads points over the whole space without assuming a model. Classical and screening designs have a fixed structure for fitting a polynomial or finding the variables that matter. The theory is in Design of experiments. The Method menu groups the methods as shown.

For the space-filling methods you choose the number of points. For the others the design decides: with four real variables and the default settings, Full Factorial gives 17 runs, Fractional Factorial 9, CCD 26, Box-Behnken 25, Plackett-Burman 9 and GSD 8. Box-Behnken needs at least three continuous variables. Fractional Factorial, CCD, Box-Behnken and Plackett-Burman do not support categorical variables; the panel warns you and disables Generate Design. Use Full Factorial, GSD or a space-filling method when a variable is categorical.

An optimal design, which targets a model you name, has its own panel: see Optimal design. If a variable is hard to change, a Split-plot toggle appears above Method; see Split-plot designs.

1. Generate a design¶

1. In Initial Design (DoE), pick a Method.
2. For a space-filling method, set Points. Seed (opt) is shown for LHS, Sobol and Random; leave it blank and each click draws a new design, or type a whole number to get the same design every time. Halton and Hammersly are fixed sequences and have no seed.
3. For LHS, LHS Criterion is Maximin, Correlation or Ratio. For CCD, Alpha is Orthogonal or Rotatable, and Face is Circumscribed, Inscribed or Faced.
4. Choose Generate Design. The panel reports "10 points generated" and lists them in a table.
5. Choose Stage to queue the points. The Add Point... button in the Experiment Data panel then shows the count, and steps through the staged points so you can enter each result; see Run and record. Choose CSV to save the points as initial_design_<method>_<N>pts.csv, for example initial_design_lhs_10pts.csv. In Chrome and Edge on https or localhost a save dialog lets you pick the folder; other browsers, and plain http addresses, download the file.

Generating a design does not add anything to the session's data. Stage records each point with a reason such as "Initial DoE (LHS)", and writes the method and its settings to the audit log. Linear input constraints, if you set any, apply to every method; see Constraints.

Defaults¶

From Python or REST¶

from alchemist_core import OptimizationSession

session = OptimizationSession()
session.add_variable("temperature", "real", bounds=(200.0, 300.0), unit="°C")
session.add_variable("pressure", "real", bounds=(20.0, 80.0), unit="bar")
session.add_variable("h2_co2", "real", bounds=(2.0, 4.0))
session.add_variable("whsv", "real", bounds=(1.5, 18.0))

points = session.generate_initial_design("lhs", n_points=10, random_seed=42)
print(len(points))
print(points[0])

for method in ("full_factorial", "ccd", "plackett_burman"):
    print(method, len(session.generate_initial_design(method)))

POST /api/v1/sessions/{session_id}/initial-design
Content-Type: application/json

{"method": "lhs", "n_points": 10, "random_seed": 42, "lhs_criterion": "maximin"}

POST /api/v1/sessions/{session_id}/experiments/staged/batch
Content-Type: application/json

{"experiments": [{"temperature": 271.2, "pressure": 55.3, "h2_co2": 3.3, "whsv": 7.4}], "reason": "Initial DoE (LHS)"}