Build run evidence
Link actual conditions, stage trajectories, material, tooling, and quality outcomes through one run identity.
PROCESS DIAGNOSIS · CONSTRAINED OPTIMIZATION
From run evidenceto the next recipe.
An open-source process diagnosis and optimization system that turns linked equipment, production, and inspection data into trustworthy evidence for the next recipe.
Traceable evidenceTestable causesReviewable recommendationsReusable conclusions
ENGINEERING DECISION · EVIDENCENext-recipe recommendation awaiting confirmation
Evidence for one real run
RECIPE OPTIMIZATION · RUN-042
Actual control42.0
Stage deviation+1.8σ
Tooling revisionTOOLING-A
Key differenceHolding stage
Valid runs12
Next recipeGenerated
Product illustration · facts, differences, uncertainty, and an actionable next step
FROM DATA TO DECISION
Normal production becomes the source of continuous optimization.
No experiment setup or manual recipe reclassification is required. The system links actual settings, process context, and quality outcomes from completed runs and creates an independent next-recipe recommendation when the evidence qualifies.
Form observations
Completed real recipe runs become reviewable optimization samples after quality and coverage admission.
Recommend the next recipe
Return one candidate recipe with prediction intervals and rationale inside objectives, safety boundaries, and observed coverage.
Keep learning from production
Engineers confirm through the existing production flow; each new run feeds the next recommendation.
ENGINEER IN THE LOOP
The system proposes. Engineers decide.
Ingot absorbs real runs, checks admission, computes, and exposes uncertainty. Engineers define objectives and safety boundaries and decide whether to adopt the next recipe. Recommendations are never dispatched automatically; controlled validation is created separately only for causal confirmation or extrapolation.
Define
question · variables · boundaries
Connect
protocols · points · units
Record
runs · trajectories · context
Qualify
quality · provenance · completeness
Optimize
observe · model · recommend
Confirm
adopt · learn · preserve
THE METHOD TOOLBOX
Confirm that the data are trustworthy before choosing an analysis or optimization method.
The system first checks that real runs are complete, comparable, and cover at least two actual recipes, then selects response-surface or constrained optimization methods by sample size, coverage, and safety constraints. Separate controlled validation is reserved for causal confirmation and extrapolation.
RUNSSTATISTICSMODELS
Confirm data usability
Check completeness, actual values, units, time, and provenance, and identify version changes or drift.
Process diagnosis
Use matching, robust statistics, stage trajectories, and context stratification to narrow candidates.
Form observations
Link actual recipes, process context, and reviewed quality outcomes automatically, with explicit exclusions and evidence scope.
Recommend the next recipe
Propose one candidate recipe inside objectives, safety boundaries, and the observed parameter envelope, with expected outcomes, risks, and rationale.
Data qualityReal runsOptimization observationsMultiple objectivesConstrained optimizationProcess knowledge
WORKS WITH YOUR EXISTING SYSTEMS
Connect existing systems without taking over production control.
A shared run identity links actual recipes, process data, and quality outcomes from existing systems into engineering evidence for process diagnosis and recipe optimization. Production execution, real-time control, and compliance approval remain with the systems and teams that own them.
MES · SCADA · HistorianReceive run and process facts without replacing execution, monitoring, or real-time control
LIMS · QMS · ELNLink inspection, review, and research records without replacing full quality or document management
DOE · response surfaces · Bayesian optimizationSelect methods by question and data instead of forcing one algorithm onto every scenario
Confirm · execute · learnThe system proposes the next recipe; engineers set boundaries and decide whether to adopt it through the existing production flow
STABLE CORE, EVOLVING METHODS
Optimization capabilities evolve. Evidence boundaries remain fixed.
A change in machine, product, or process requires new mappings, variables, objectives, constraints, and context. Run identity, evidence principles, independent recommendation records, and engineering authority remain stable.
Stays stable
Real data supports engineering judgment, and every conclusion traces to sources and remains testable
Configured per scenario
Equipment mappings, variables, stages, quality objectives, safety constraints, context, and mechanism knowledge
Continues evolving
Statistics, surrogate models, optimization strategies, page layouts, and language models
RUN IT YOURSELF
Open source across the complete recipe-optimization loop.
Ingot is Apache-2.0 licensed and self-hostable inside the plant. Field acquisition, run evidence, process diagnosis, real-run optimization, optional controlled validation, and knowledge preservation live in one repository; public validation protocols and results are independently reproducible.
Current maturity
The main software workflow is implemented and has automated tests; real-factory benefit validation remains incomplete. The system may be used for product evaluation and controlled pilots, but current evidence does not establish consistent reductions in experiments or development time.
git clone https://github.com/liuweichaox/Ingot.git
cd Ingot
cp .env.example .env
docker compose -f docker-compose.app.yml up -d --buildSTART WITH ONE REAL DATA LOOP
Begin with one real process problem.
Connect a set of real recipe runs, qualify actual settings and outcomes, and let the system form observations and return the first reviewable next recipe.