Why simulate Dokra ageing?
Dokra (or Dhokra) is a family of lost-wax metal-casting traditions practised by artisan communities across India. Its value lies not only in the finished object, but also in the casting knowledge, surface treatment, motifs and cultural context carried by each piece. Preservation therefore means more than keeping an object visually attractive: it means protecting its material stability while documenting what should not be altered.
Reinforcement learning (RL) can help compare conservation strategies in a digital environment. It should not be used to experiment directly on an irreplaceable artefact. Instead, an agent can learn from historical observations, accelerated laboratory tests, expert rules and physics-informed simulations to estimate which storage or display conditions are least damaging.
This is a research workflow, not a shortcut. Conservation scientists must define the material risks, curators must set acceptable interventions, and artisans and community representatives should help interpret cultural significance. Teams new to applied AI can first build a small, reproducible prototype using guidance from machine learning portfolio projects for beginners in India.
Start with a defensible dataset
A useful system needs linked records for each documented object or test coupon. Capture:
- Object metadata: collection ID, approximate date, maker or community attribution where appropriate, dimensions, alloy information and casting method.
- Surface observations: colour, patina, corrosion products, cracking, deposits, abrasion and previous repairs.
- Environmental history: temperature, relative humidity, light exposure, pollutants, handling and storage changes.
- Scientific measurements: calibrated photographs, 3D scans, microscopy, mass, surface chemistry and non-destructive material readings.
- Conservation actions: cleaning, coating, packing, relocation and inspection dates, including who approved each intervention.
Document uncertainty rather than filling gaps with invented values. A photograph taken under inconsistent lighting cannot reliably support a colour-change label, and a single inspection cannot establish an ageing trend. Store provenance for every measurement, use consistent calibration targets, and separate training, validation and test objects by artefact rather than by individual image. This prevents the model from memorising one object’s appearance.
For a student or early-stage team, begin with a narrow research question such as: *How does relative humidity affect the probability of visible corrosion on a defined Dokra alloy under controlled conditions?* A small, well-labelled dataset is more valuable than a large archive with unclear provenance.
Represent the ageing environment
The simulation environment should describe both the object and its surroundings. A state might contain:
- current corrosion or patina indicators;
- temperature and relative humidity, including duration and rate of change;
- light dose and pollutant exposure;
- coating condition, packaging and display configuration;
- uncertainty estimates and the time since the last inspection.
Actions should be limited to decisions a conservator can actually implement, such as moving an object to a controlled case, changing inspection frequency, selecting an approved inert packing material or evaluating a conservation coating. Do not treat aggressive cleaning or untested chemical treatment as ordinary actions merely because the algorithm can model them.
A practical first environment can use a time step of one week or one month. Its transition model may combine empirical data with a corrosion or diffusion model, while a computer-vision model estimates changes in surface appearance. Where evidence is weak, use probability distributions rather than false precision. A digital twin should communicate confidence intervals and alternative scenarios, not present a single forecast as fact.
Design the reward around conservation ethics
Poorly chosen rewards can produce unsafe recommendations. A reward function should balance several goals:
- minimise predicted corrosion, cracking and irreversible surface change;
- maintain stable temperature and humidity rather than chasing short-term improvements;
- reduce unnecessary handling and intervention;
- respect approved conservation budgets and energy limits;
- preserve documentation quality and inspection coverage;
- penalise actions that exceed evidence, reversibility or institutional policy.
Use hard constraints for non-negotiable rules. For example, the agent should not recommend conditions outside a museum’s approved range, apply a coating without material compatibility evidence or optimise visual brightness at the expense of an authentic patina. Multi-objective optimisation is usually more appropriate than a single “preservation score”.
Before training, have conservators review hypothetical actions and rank their risks. Those judgements can shape the reward model and expose assumptions hidden in the data. Explainable baselines—such as threshold rules, regression and survival analysis—should be tested before RL. If a simple humidity alert performs as well as a complex policy, use the simpler system.
Train and evaluate safely
A suitable workflow is:
1. Build a baseline: predict corrosion probability or condition change without RL.
2. Create a simulator: combine measured observations, laboratory ageing studies and expert-defined constraints.
3. Train offline: use historical decision records or synthetic rollouts before allowing any operational recommendation.
4. Stress-test scenarios: vary humidity cycles, sensor failure, missing observations, uncertain alloy composition and extreme weather.
5. Compare policies: evaluate the RL policy against current practice, fixed rules and conservative alternatives.
6. Run shadow mode: generate recommendations for review without changing storage or display conditions.
7. Audit outcomes: track whether recommendations are understandable, reproducible and materially safe.
Offline RL and conservative algorithms are preferable when real-world exploration is unacceptable. Never reward the agent by exposing an artefact to harmful conditions. Accelerated ageing experiments should use representative, documented test samples and be approved by qualified specialists. Laboratory results also need careful interpretation: an artificial pollutant or rapid humidity cycle may not reproduce decades of museum storage.
Teams building the data and deployment layer can consult scalable machine learning infrastructure for developers, while those planning production inference should treat model versioning, sensor quality and rollback as conservation controls—not merely engineering details. A GitHub-based record of datasets, assumptions and evaluation results also makes the work easier to review; how to build a machine learning portfolio on GitHub offers a useful documentation pattern.
A realistic pilot for India
A museum, university laboratory or archive could begin with 20–50 documented objects and a small number of non-heritage metal coupons representing relevant alloys and finishes. Install calibrated temperature and humidity sensors in storage and display areas, photograph samples using a repeatable setup, and collect expert condition assessments at fixed intervals.
The first deliverable should be a dashboard showing environmental history, condition trends, uncertainty and recommended inspection priorities. Only after the dashboard has passed expert review should the team test a policy model in shadow mode. Include multilingual labels and training materials where staff or artisan partners prefer Indian languages. Data agreements should specify ownership, access, consent, attribution and whether culturally sensitive images may be published.
Limits, governance and success measures
Dokra objects are not chemically uniform, and visible patina is not automatically damage. Sparse historical records, changing display conditions and inconsistent terminology can make predictions unreliable. A model trained on one museum’s collection may fail on objects from another region or alloy tradition.
Measure success through conservation outcomes and accountability: fewer high-risk environmental excursions, earlier detection of deterioration, lower unnecessary handling, transparent decision logs and positive review by conservators and community stakeholders. Keep a human approval step for every intervention. Retire or retrain the model when new material evidence contradicts its assumptions.
RL is most valuable here as a decision-support method. Used cautiously, it can help Indian institutions compare long-term preservation strategies while leaving authority with the people responsible for the artefacts, the collections and the living traditions behind them.