🧭 FusionGuide

AI experiment planner for fusion materials.
Tells you which irradiation experiments to run next — with the fewest possible reactor slots.

Problem · Quick Start · How It Works · Facilities · FusionMatDB · FusionUQ

---

The Problem

A slot at HFIR costs ~$500k and has an 18-month queue. You have 15 experiment budget slots to characterise V-4Cr-4Ti yield strength across dose and temperature.

Random sampling needs ~25 experiments to reach acceptable uncertainty. FusionGuide uses Bayesian optimisation with physics-informed priors to reach the same threshold in ~9 experiments — cutting reactor time and cost by 60%.

---

Quick Start

pip install -e .

# Get experiment recommendations for a vanadium alloy campaign
fusionguide recommend --material vanadium_alloy --n 3

# What facilities can take 20 samples within 12 months?
fusionguide facilities --timeline 12 --samples 20

# Load real data from FusionMatDB as prior
fusionguide recommend --material vanadium_alloy --n 3 \
  --prior-parquet ../fusionmatdb/data/export.parquet

---

How It Works

Built on BayBE (Merck, Apache 2.0) + BoTorch (Meta). The fusion-domain adaptation is the novel contribution:

from fusionguide.campaign import FusionCampaign
from fusionguide.spaces import VanadiumAlloySpace
from fusionguide.targets import YieldStrengthTarget

# Start a campaign with physics-informed prior
campaign = FusionCampaign(
    name="V-4Cr-4Ti characterisation",
    parameters=VanadiumAlloySpace(dose_dpa_range=(0.1, 20.0)),
    target=YieldStrengthTarget(),
)

# Get recommendation — tells you which dose/temp/spectrum to test
rec = campaign.recommend(n=1)
# → dose_dpa=8.3, irradiation_temperature_C=412, neutron_spectrum=fission

# Report result, update the GP
measurement = rec.copy()
measurement["yield_strength_MPa"] = 647.0
campaign.add_measurement(measurement)

# Save and resume later
campaign.save("campaign.json")

What FusionGuide adds on top of BayBE

Feature	BayBE	FusionGuide
Fusion parameter spaces (V-4Cr-4Ti, ceramics)	❌	✅
Irradiation facility scheduler	❌	✅
Physics prior (DBH radiation hardening model)	❌	✅
FusionMatDB prior loader	❌	✅
Uncertainty map visualisation	❌	✅

---

Pre-configured Spaces

Space	Material	Parameters
VanadiumAlloySpace	V-4Cr-4Ti	dose_dpa, temperature, Cr/Ti wt%, neutron spectrum, test temp
CeramicInsulatorSpace	Al₂O₃, MgAl₂O₄, AlN, SiC, BN, ZrO₂	ceramic class, additive wt%, sintering temp, dose, test temp

---

Facility Scheduler

from fusionguide.facility_scheduler import recommend_facility

recs = recommend_facility(timeline_months=12, n_samples=20, budget_relative=200.0)
# → HFIR (fidelity=85%, queue=12mo), ion_beam (fidelity=30%, queue=1mo)

Facility	Queue	Fidelity	Spectrum
HFIR	12 mo	85%	Fission
ATR	18 mo	85%	Fission
BOR-60	24 mo	90%	Fast
BR2	18 mo	80%	Fission
Ion beam	1 mo	30%	Ion
Proton	2 mo	60%	Proton

---

Physics Prior

The RadiationHardeningPrior encodes the Dispersed Barrier Hardening (DBH) model as a GP prior mean:

Δσ_y = A × (1 − exp(−B × dpa)) × f(T)

The GP starts with known physics and learns corrections — far more data-efficient than a black-box GP.

---

Related Projects

⚛️ FusionMatDB	The database — provides GP prior data
🧭 FusionGuide	This repo — experiment planning
🔬 FusionUQ	Uncertainty quantification for ML potentials