Scenario
When each experiment is expensive and the design space is large, a fixed DOE spends most of the budget in uninformative regions. Active learning instead proposes the most informative next experiments, guided by a campaign planner that respects constraints and budget — finding good candidates fast.
Worked example. A coating formulation must hit a hardness target across resin, filler, and additive levels within 20 experiments.
Agent workflow
- Experiment campaign planner — define variables, constraints, objectives, and budget.
- Active-learning loop — fit a surrogate and rank candidate experiments by expected information gain / acquisition value.
- Iterate: propose, (the lab runs), fold results back, repeat until the objective or budget is met.
Demo output
A ranked set of proposed experiments each round and a convergence trace toward the target within budget. The magic moment: a strong candidate emerges after a handful of runs, long before a full DOE would have finished.
Deterministic vs LLM
- Deterministic — the Bayesian optimization and campaign planning.
- LLM — the discovery-strategy narrative.
Every proposal is a validated calculation. The agent reads and routes; the engine decides.
Limits
Screening-level discovery, dependent on the quality of the measured feedback. It guides the search; the lab confirms the winner.
Skills orchestrated by this use case
This scenario routes an AI agent through the following curated skills.
-
Active Learning Loop
Autonomous LabsRun a closed-loop active-learning optimization that, given results from completed experiments, updates a surrogate model, proposes the next most informative experiments via an acquisition strategy, and decides whether to continue or stop — orchestrating Paramus ML/optimization tools.When to use: a user wants the system to choose the next best experiments automatically.Multi-tool WORKFLOW with acquisition-strategy judgment. -
Experiment Campaign Planner
Autonomous LabsPlan an autonomous experimental campaign by translating a research goal and constraints into a concrete experiment set, selecting a design strategy, and defining the design–make–test–analyze (DMTA) loop with clear stop criteria, orchestrating Paramus DOE and planning tools.When to use: a user wants to set up a closed-loop experimental campaign.Multi-tool WORKFLOW with experimental-strategy judgment.
Tools it reaches for
- Active-learning / Bayesian optimization (BoFire)
- Experiment campaign planning
- Constraint and budget handling
Browse the full deterministic layer in the tool browser.