How to Use Reinforcement Learning for Terracotta Cooling Systems

Why combine reinforcement learning with terracotta cooling?

Terracotta can support passive cooling through shading, thermal mass, evaporative effects and controlled air movement. But its performance depends on geometry, porosity, wall thickness, orientation, water availability, local humidity and occupant behaviour. A design that works in a dry Rajasthan summer may perform poorly in a humid coastal city.

Reinforcement learning (RL) is useful when a design or control problem involves repeated decisions under changing conditions. An agent tests actions in an environment, receives a reward, and gradually learns a policy. For terracotta cooling, those actions might include opening vents, changing airflow paths, adjusting shading or selecting among façade configurations.

The goal is not to make AI replace architects or building scientists. It is to search a large design space efficiently while keeping comfort, cost, water use and embodied carbon within clear limits.

Start with a precise design question

Avoid beginning with “use AI to optimise cooling”. Define the decision, time horizon and constraints first. Practical project questions include:

Which terracotta block geometry delivers the best cooling with a fixed material volume?
How should vents open during the day to reduce indoor temperature without increasing dust or noise?
What combination of cavity depth, surface area and airflow minimises cooling energy in a selected climate?
Can a façade maintain comfort during power cuts while using limited water for evaporative cooling?

Separate design variables from operational actions. Design variables are generally fixed after construction: module dimensions, clay composition, porosity, cavity depth, orientation and shading depth. Operational actions can change over time: vent position, fan speed, irrigation rate or window state. This distinction determines whether you need an optimisation workflow, an RL controller, or a combination of both.

Teams new to machine learning can establish the data and evaluation workflow through machine learning portfolio projects for beginners in India before attempting a building-scale agent.

Build the environment before training the agent

An RL agent is only as credible as its environment. A useful environment should represent the building, terracotta system, weather and occupants at an appropriate time step—typically 5, 15 or 60 minutes depending on the decision being made.

Include:

Outdoor dry-bulb temperature, relative humidity, wind speed, solar radiation and rainfall.
Indoor air temperature, operative temperature, humidity, CO₂ and surface temperatures.
Terracotta properties such as conductivity, density, specific heat, moisture response and surface emissivity.
Room geometry, glazing, occupancy, lighting, plug loads and ventilation rates.
Equipment limits, including fan capacity, actuator response, water availability and electrical tariffs.

For an early prototype, a reduced-order thermal model can make experiments faster. For final design decisions, compare promising solutions with a calibrated simulation tool or physical test cell. Use weather files relevant to the target location rather than generic annual averages. Indian projects may require separate experiments for hot-dry, warm-humid, composite and temperate conditions.

Sensor data should be cleaned, time-synchronised and labelled with operating conditions. Track missing readings and sensor drift; do not silently fill long gaps and then treat the resulting dataset as ground truth. A clear data pipeline is more valuable than a sophisticated algorithm trained on unreliable measurements.

Define the action space and reward carefully

The action space should reflect what the real system can actually do. Examples include:

Select one of several terracotta module geometries during parametric design.
Set vent openings to discrete positions.
Adjust fan speed within a safe operating range.
Choose a shading configuration based on solar exposure.
Control intermittent water delivery for evaporative surfaces.

A reward function should balance comfort and resource use rather than reward low temperature alone. One practical structure is:

Reward = comfort score − energy penalty − water penalty − carbon penalty − constraint penalty

Comfort can be measured using operative temperature or an accepted comfort model, with stronger penalties when conditions leave the project’s target range. Add penalties for excessive humidity, condensation risk, rapid temperature changes, actuator wear, noise and unsafe air quality. If the project is off-grid, include battery state and peak load. If water is scarce, water use must be a first-class constraint, not an afterthought.

Normalise reward components so that one metric does not dominate accidentally. Test the reward with architects, mechanical engineers and occupants before training. A mathematically neat reward can still produce an unacceptable building—for example, a controller that saves electricity by keeping vents closed and allowing stale air.

Select an algorithm and training strategy

For a small, discrete action space, Q-learning or a small Deep Q-Network may be sufficient. Continuous controls such as fan speed or vent angle can use algorithms such as Soft Actor-Critic or Proximal Policy Optimisation. The choice should follow the control problem, not the popularity of an algorithm.

Use a staged workflow:

1. Create a baseline. Compare the proposed system with conventional ventilation, fixed schedules and a simple rule-based controller.
2. Train in simulation. Vary weather, occupancy, sensor noise and equipment performance so the agent does not memorise one scenario.
3. Use conservative constraints. Prevent actions that could cause condensation, overheating, structural stress or unsafe humidity.
4. Evaluate unseen cases. Hold out weather periods, occupancy patterns and climate zones for testing.
5. Deploy in shadow mode. Let the agent recommend actions while the existing controller remains in charge.
6. Move to supervised control. Require hard safety limits, manual override and automatic fallback to a known rule set.

This workflow benefits from reproducible experiments, versioned datasets and scalable compute; guidance on scalable machine learning infrastructure for developers is relevant when simulations and experiments begin to grow.

Optimise the terracotta design, not only the controller

RL can explore design combinations, but each candidate must be assessed across multiple objectives. Useful outputs include annual discomfort hours, peak indoor temperature, cooling electricity, water consumption, terracotta mass, embodied carbon, maintenance burden and estimated cost.

A multi-objective approach is usually better than a single “best” design. Generate a Pareto set showing trade-offs such as lower embodied carbon versus higher peak temperature or lower water use versus reduced evaporative performance. Architects can then select a solution appropriate to the site and budget.

Consider manufacturability early. A theoretically optimal module may require tolerances, firing temperatures or moulds unavailable to local producers. Include breakage rates, transport distance, repairability and end-of-life reuse in the assessment. Terracotta is not automatically low-carbon: firing energy, replacement frequency and transport matter.

Validate with a physical test cell

Simulation should narrow the options, not end the project. Build a small test cell or façade mock-up with the same module geometry, mortar or dry-joint system, coatings and sensors proposed for construction. Instrument both the terracotta surface and occupied zone.

Compare modelled and measured temperature, humidity, airflow and energy use across contrasting weather conditions. Calibrate uncertain parameters, then rerun the agent with the updated model. Report error ranges and failure cases instead of presenting one impressive average.

For Indian deployment, test dust accumulation, monsoon moisture, power interruptions, water quality, insects, vandalism and maintenance access. A passive system that performs well in a clean laboratory but clogs or develops mould during monsoon season is not a sustainable solution.

Common failure modes

Training on one climate: produces policies that fail when weather or building use changes.
Rewarding energy savings alone: can sacrifice comfort and indoor air quality.
Ignoring uncertainty: gives false confidence in material properties and occupancy forecasts.
Overusing deep learning: adds complexity where a rule-based controller or Bayesian optimiser would work.
No baseline comparison: makes it impossible to quantify the AI benefit.
Unsafe deployment: allows an experimental policy to control equipment without hard limits.
Weak documentation: prevents architects, fabricators and facility teams from understanding the result.

Use transparent dashboards to show comfort, energy, water and carbon outcomes. A well-designed visualisation workflow—such as the one discussed in AI tools for data visualisation design—can help stakeholders inspect trade-offs rather than accept a single score.

A practical 2026 project plan

Begin with a two- to four-week baseline study and site survey. Next, build a calibrated digital model and collect representative weather and occupancy data. Run a small design sweep before introducing RL; this reveals whether the problem is truly sequential and whether the agent adds value. Train offline, evaluate against fixed schedules and rule-based control, then validate the top designs in a test cell.

Set success thresholds before deployment—for example, a target comfort range, maximum water use, peak electrical demand and acceptable model error. Publish the assumptions, reward function, safety constraints and test scenarios with the project results. This makes the work useful to future builders instead of turning it into an opaque demonstration.

For a broader sustainability application of AI optimisation, compare the same principles with AI route optimisation for sustainable EV charging in India: define constraints, model local conditions, test alternatives and validate against real operations.

Conclusion

Reinforcement learning can help design and operate terracotta cooling systems when the project has a credible thermal model, reliable data, explicit constraints and physical validation. Its strongest role is exploring climate-responsive designs and adaptive controls that would be difficult to tune manually. Start with a measurable building problem, establish a non-AI baseline, keep humans in control, and judge success by comfort, energy, water, carbon and maintainability together.

Apply for AI Grants India