| name | synthesis-planner |
| description | Generate a good first-guess synthesis recipe for an inorganic material as the seed of a lab campaign. Sources the route by best available evidence: literature-validated recipes (Materials Project) first, then a route reasoned from solid-state chemistry and analogous known systems. Produces a structured, robot-middleware-ready recipe (precursors with forms, ordered steps with parameters), defines the tunable parameter search space for downstream Bayesian-optimization refinement, and can quantify precursor masses for a target batch size. Use this skill WHENEVER the user wants a synthesis route, recipe, or procedure; asks how to make / prepare / synthesize a compound or formula; needs precursors, calcination/annealing conditions, or a solid-state route; or when a DFT-validated candidate needs a first synthesis plan to hand to a robotic lab. It does NOT execute synthesis or approve autonomous runs — a separate experiment-orchestration skill plus a human gate own that — and it does NOT run the optimization loop itself (the active-learning skill does).
|
Synthesis Planner Skill
Produce a first-guess synthesis recipe for an inorganic target — the precursors,
method, and conditions most likely to work — packaged so two downstream skills can
take it forward:
- an experiment-orchestration skill that turns the recipe into
machine protocols and runs it (with a human gate), and
- an active-learning skill that refines the conditions via Bayesian optimization
over successive experiments.
So your output is not a final answer — it is the seed of an optimization campaign.
That framing drives every design choice below: a route a robot can execute, plus an
explicit parameter search space the optimizer can move in.
Input: a target formula (often a high-confidence candidate handed over after DFT),
with optional constraints (method, temperature ceiling, batch size).
Output: a structured recipe + a tunable parameter space + provenance and confidence.
Scope: where this skill stops
Be clear about the boundary — it keeps the pipeline clean and safe:
- You plan; you do not execute. Never mark a recipe "approved for autonomous
execution." Emit a recipe + a confidence level + a recommended human-review depth,
and hand off. The orchestration skill plus a human decide whether to run it.
- You seed; you do not optimize. Don't iterate conditions yourself. Define the
initial point and the bounds; the active-learning skill explores them.
This separation is why the output has both a concrete recipe (for the robot) and a
parameter space (for the optimizer), rather than a single take-it-or-leave-it route.
Tools
| Tool | Role |
|---|
mp_search_recipe | Literature-validated recipes from Materials Project. Try this first — it is the only sourcing tool. If it returns nothing, you reason the route out yourself (no tool for that tier — it's your chemistry judgment). |
synthesis_recipe_quantifier | Convert stoichiometry → precursor masses (g) for a target batch size. |
Upstream context often comes from candidate-screener / vasp / orca (a DFT-validated
formula to make). Downstream you hand to the experiment-orchestration and active-learning
skills.
Workflow
1. Intake
Establish the target and the campaign context:
- the target formula (normalize to a reduced formula, e.g.
LiCoO2, not Li1Co1O2);
- any constraints the user gave: method (solid-state / hydrothermal / sol-gel …),
temperature ceiling, time, atmosphere, batch size;
- the downstream intent — is this headed for a robotic run and a BO loop? (Usually
yes in this pipeline; it tells you to populate the parameter space carefully.)
2. Source the route — by best available evidence
Always start from literature; if there is none, reason the route out yourself. The
principle: a published recipe is proven; a route reasoned from solid-state chemistry
and close analogues is a strong, defensible starting point — so prefer the strongest
evidence available and be honest about which tier you landed on. Full algorithm,
tool arguments, and worked examples in
references/route-sourcing.md. In brief:
- Literature first —
mp_search_recipe(target_formula, format_routes=True). If it
returns routes, use them; this is the high-confidence path.
- Reasoned route — if there's no literature, design the route from
materials-chemistry principles and analogous known systems: pick precursors (common
carbonates, oxides, nitrates, or hydroxides that decompose cleanly to the product
and together supply every product element), a method suited to the phase, and
conditions (temperature, atmosphere, time) bracketed from compounds known to behave
similarly. Calibrate confidence to the analogues — medium when the chemistry is
well-precedented (e.g. a layered oxide or garnet near known ones), low when it is
novel or exotic. Name the analogues you reasoned from and require human review.
Careful chemical reasoning over close analogues is the heart of this tier — it is more
reliable than a black-box predictor, so invest in it rather than reaching for a shortcut.
3. Parameterize for optimization
A first guess that's a single fixed recipe wastes the lab loop. Identify the knobs the
active-learning skill should tune, each with an initial value and plausible
bounds, derived from the route you sourced. Typical knobs: calcination temperature,
dwell time, ramp rate, atmosphere, and stoichiometry offsets for volatile elements
(e.g. Li/Na excess). How to choose values and bounds per method is in
references/parameterization.md. This parameter space
is the single most important thing you add on top of the raw route.
4. Quantify (optional, when a batch size is known)
If the user or the orchestration layer specifies a batch size, call
synthesis_recipe_quantifier(recipes, target_batch_size_grams, target_formula) to turn
stoichiometry into masses to weigh out. Treat the masses as deterministic arithmetic
to be re-checked at the bench (verify hydrate/water-of-crystallization handling for
precursors like Ni(NO3)2·6H2O).
5. Hand off
Return the recipe in the structured shape below so the next skills can consume it
directly. Contracts and field meanings in
references/handoff-contracts.md.
Output shape
Return a single object with these parts (fill the fields; this describes intent and
structure, not a rigid literal — include what you have, omit what you don't):
{
"target_formula": "...",
"source": "literature | reasoned", # which evidence tier you used
"confidence": "high | medium | low",
"method": "solid_state | hydrothermal | sol-gel | ...",
"recipe": { # → experiment-orchestration skill
"precursors": [
{"compound": "Li2CO3", "role": "Li source", "form": "carbonate",
"stoichiometric_amount": 0.5, "mass_grams": null} # mass set if quantified
],
"steps": [
{"step": 1, "action": "mix_grind", "parameters": {}},
{"step": 2, "action": "calcine",
"parameters": {"temperature_c": 850, "time_h": 12, "atmosphere": "air",
"ramp_c_per_min": 5}}
]
},
"optimization": { # → active-learning (BO) skill
"objective": "phase purity by XRD (maximize)",
"variables": [
{"name": "calcination_temperature_c", "initial": 850, "bounds": [750, 1000], "type": "continuous"},
{"name": "calcination_time_h", "initial": 12, "bounds": [4, 24], "type": "continuous"},
{"name": "Li_excess_pct", "initial": 0, "bounds": [0, 15], "type": "continuous"},
{"name": "atmosphere", "initial": "air", "choices": ["air", "O2", "Ar"], "type": "categorical"}
]
},
"provenance": {"tool": "mp_search_recipe", "doi": "...", "n_literature_recipes": 206},
"review": "minimal | recommended | required", # human-review depth, NOT an execution approval
"warnings": ["..."]
}
Keep it honest: the confidence, review, and warnings fields must reflect which tier
sourced the route. See references/safety.md for the confidence
tiers, the review depth each implies, and the red flags that should raise caution.
Reference files
