Research question and scope
The research problem is to choose the next feasible experiment from observations without treating an optimizer recommendation as permission to run chemistry. BayBE offers campaign-level recommendations and measurement feedback; BoTorch provides probabilistic models and acquisition-function building blocks; Ax manages search spaces, objectives and trials; Optuna uses objective functions and studies for general optimization. Compare their fit to your control and recordkeeping needs, not a claimed universal ranking.
Prerequisites
Create a safe, synthetic or previously measured example first. Define each variable with type, units and allowed range; distinguish controllable factors from recorded conditions. Specify objective direction, measurement uncertainty, batch size, costs, safety exclusions and a reviewer. Keep restricted experimental data on approved systems and review each package’s license.
Workflow
- Prepare a versioned table: trial ID, parameter values, units, timestamp, outcome, uncertainty, failure reason and data provenance. Reserve an immutable set of historical observations for retrospective evaluation.
- Select one tool branch. BayBE
Campaign.recommend()andadd_measurements()implement a measured feedback loop in its documentation. Ax represents candidates as trials within an experiment. BoTorch requires explicit surrogate and acquisition choices. Optuna’s study calls a user-defined objective; an objective that triggers a real experiment requires additional governance. - Generate a proposal for the next batch on synthetic data. Validate parameter types, hard constraints and duplicates; route every proposal to human approval. Only after an approved measurement should it be recorded as observed data. Never insert a simulated outcome as a laboratory result. No optimizer code was run for this article.
How to interpret and check the output
Log proposed versus executed conditions, the reason for rejection, and the measured result including uncertainty. A recommendation is a decision aid; compare it with a simple baseline under the same budget. If an experiment fails, preserve that status rather than treating missing outcome as success or zero.
Limits and common errors
Constraint handling and batch behavior vary by tool and version. Ax documentation notes that predicted constraints can still be violated by real outcomes. Optimization can overfit small, biased datasets; model uncertainty is not a safety guarantee. Neither a Recipe nor a CARS adapter for these four tools is asserted to exist.
Related ChemAI Atlas pages
Resources: BayBE, BoTorch, Ax, Optuna. Related guide: Evaluating AI workflows for chemical engineering.
Primary sources
- https://emdgroup.github.io/baybe/stable/index.html (official source; accessed 2026-10-13)
- https://emdgroup.github.io/baybe/0.13.2/userguide/campaigns.html (official source; accessed 2026-10-13)
- https://botorch.org/docs/models (official source; accessed 2026-10-13)
- https://ax.dev/docs/experiment/ (official source; accessed 2026-10-13)
- https://optuna.org/ (official source; accessed 2026-10-13)