R&D Campaign
Define controls, objectives, hard bounds, quality constraints, experiment cost, and stopping rules.
AI PROCESS R&D · CLOSED-LOOP OPTIMIZATION
Make every experiment converge on a better process.
Ingot turns equipment trajectories, actual recipes, and quality outcomes into learnable experiments, combining physical priors with Bayesian optimization to choose the next run within safe process boundaries.
Multi-objectiveSmall-data learningSafety constraintsCross-product transfer
NEXT RUN · RECOMMENDATIONSAFE TO EXECUTE
Recommended next run
MANUFACTURING R&D · CAMPAIGN-042
Control variable A42.0
Control variable B11.8
Control variable C0.72
Primary quality metric0.21 ± 0.06
Constraint pass probability94%
Information valueHIGH
Product UI illustration · every recommendation includes intervals, risk, and rationale
THE COMPLETE PRODUCT
From R&D target to a stable process window, in one system.
This is not an algorithm tab attached to data collection. Data, models, decisions, and knowledge are organized around one question: what experiment should run next?
Automatic observations
RunKey joins setpoints, realized trajectories, process features, inspections, and validity.
Next-run recommendations
Return one or many candidates with intervals, safety probability, information value, and rationale.
Engineer decision desk
Compare candidates, lock variables, adjust bounds, approve execution, and preserve evidence.
Convergence monitoring
Identify process windows, repeatability, and drift; recommend stopping when marginal value falls.
Knowledge transfer
Turn validated relationships and model posteriors into warm-start priors for similar products.
THE SELF-IMPROVING LOOP
Each real outcome changes the next decision.
Every run becomes evidence in the loop. The system continuously updates its process belief instead of periodically producing a model report that never learns again.
Define
objectives · constraints · variables · cost
Observe
setpoints · trajectories · quality
Learn
physical mean + GP residual
Suggest
safe · multi-objective · cost-aware
Verify
execute · inspect · update · converge
THE NUMERICAL CORE
An optimization brain built for expensive, small-data experiments.
PyTorch, GPyTorch, and BoTorch make numerical decisions. An LLM can parse intent, structure knowledge, and explain recommendations, but never invents process settings.
PYTORCHGPYTORCHBOTORCH
Trajectory surrogate
GP₁ learns how controls create realized temperature, pressure, displacement, and other trajectories.
Quality surrogate
GP₂ combines settings, trajectories, and context to predict optical objectives and safety outcomes.
Decision policy
qLogNEI, qLogNEHVI, batching, and cost-aware strategies balance exploration and exploitation.
Trust and safety
Hard bounds, outcome constraints, calibration, drift detection, and fallback policies work together.
Multi-output objectivesNoise modelingPhysical mean functionsPending-point avoidanceParallel batchesPluggable acquisition
ONE OPTIMIZATION SYSTEM
Start with any industrial experiment. Do not bind the core loop to equipment or process.
Equipment, sensors, manufacturing execution, and quality systems are data sources—not platform boundaries. For another process, replace mappings, feature definitions, objectives, constraints, and physical priors; the experiment model and optimization engine remain.
Industrial data mappingMap equipment, sensors, manufacturing systems, and quality systems into one experimental observation
Feature & objective specConfigure stages, trajectory features, controls, objectives, constraints, and cost
GP · BO · physics priorUse one core for small data, multiple objectives, noise, and safety boundaries
Transferable priorCarry validated relationships into the next model, material, or process scenario
FROM ONE PROCESS TO THE NEXT
Change the scenario, not the optimization system.
The reusable asset is not an address map for any device. It is the complete capability for sequential experiment optimization.
Stays the same
Campaign workflow, experiment model, GP/BO core, review, and convergence logic
Configured per process
Device mappings, stage features, quality objectives, constraints, and physical priors
Compounds over time
Cross-product priors, material knowledge, process windows, and explainable evidence
RUN IT YOURSELF
The complete loop, not just an algorithm sample.
MIT licensed. Edge acquisition, experiment platform, optimizer, engineering workbench, and documentation live in one repository. Evaluate with historical replay, then connect real equipment.
git clone https://github.com/liuweichaox/Ingot.git
cd Ingot
cp .env.example .env
docker compose -f docker-compose.app.yml up -d --buildBUILD THE FIRST CLOSED LOOP
Start with one real process. Build optimization that transfers.
Replay history first, assist the next run, then let every new process begin with knowledge from the last.