When the Environmental Model Is Wrong, the Answer Is Not on the Screen — It Is Underground

Imagine trying to find out where the smoke goes inside a dark house by taking only two flash photographs a year. It sounds reckless, but that is exactly how most environmental projects monitor water and soil contamination today: teams go into the field every six months, collect manual samples from a handful of wells and try to guess what happened on the other 363 days of the year.

With the arrival of artificial intelligence and automation in environmental engineering, a dangerous illusion has emerged: the idea that good software or an intelligent algorithm can "predict" the behavior of a site even when data is scarce. In fact, the opposite happens. And understanding what to do when the data modeler (Data Modeler — whether a human specialist or an AI agent) gets it wrong is what separates a safe project from a multimillion-dollar liability.

The problem of seeing the subsurface in slow motion

Beneath our feet, in the layers that ECBYTS organizes from the deep subsurface (-Z), through the surface where people work (0Z), up to the atmosphere (+Z), the environment never stops. A heavy rain on a Tuesday afternoon can push a contaminant closer to a river; a change in air pressure can make toxic gases rise from the soil into a warehouse within hours.

In physics and telecommunications, there is a basic rule called the Nyquist-Shannon Sampling Theorem: if a phenomenon changes every day and you only measure it every six months, you lose the entire film and create an optical illusion. In hydrogeology, the scientists Keith Beven and Andrew Binley gave this trap another name: equifinality. When we have few measurements, ten specialists (or ten AI agents) can produce ten different maps that look "right" for the past but point to completely opposite paths for the future of the risk.

What the law requires: the Conceptual Site Model (CSM) as a confidence thermometer

In the State of São Paulo — the regulatory reference for Brazil through CETESB DD No. 038/2017/C — every contaminated site study must build and update a Conceptual Site Model (CSM).

To simplify: the CSM is the "technical story" of the site. It explains where the contamination was born, how it travels and whom it can reach. The CETESB standard itself determines that, right at the start (CSM 1), the study receives a confidence grade (A, B or C) and that the project only advances to the next phases (CSM 2 and CSM 3) as the doubts decrease.

The problem is that, outside an integrated system, this doubt is usually treated subjectively. At ECBYTS, uncertainty has a number, a color and a direct consequence.

How ECBYTS turns the modeler's error into real protection

Inside the Uncertainty Hub (UNC-001) of ECBYTS, no number exists alone. The system runs thousands of mathematical simulations (the Monte Carlo method and standard uncertainty propagation per ISO/GUM) to calculate the exceedance probability — that is, the real chance that contamination exceeds the legal safety limit.

When that uncertainty grows, the system applies what we call the Band Widener: the lower the quality and the certainty of the data (measured by the κ index), the larger the financial and environmental risk margin calculated for the project.

In our vision of the future for highly automated studies, two scenarios must immediately trigger a safety lock on the Conceptual Site Model (CSM):

1. When the modeler (human or AI) gets the prediction wrong. If the person responsible for modeling — or the autonomous agent of the system — makes an estimate for an area and the next measurement shows that it missed, the system must not "adjust the curve" to hide the error. The error proves that we still do not understand how the subsurface of that site works. The result must be automatic: rigor on the CSM increases and any hasty decision is blocked until the site has been better investigated.
2. When a lot of new data comes in and doubt increases. It sounds contradictory, but it happens all the time in geoscience (a phenomenon known in classical geostatistics through Journel as the proportional effect): you run dozens of new boreholes on the site and, instead of the map becoming simple, you discover that the soil is full of cracks, mixed layers and hidden pathways.

When the modeler receives a flood of data and the ECBYTS risk card jumps from Green (less than 5% chance of violation) to Yellow or Red (above 20%), the old Conceptual Site Model is no longer valid. More data revealed a more complex problem, demanding a far stricter approval filter.

The solution: replacing guesswork with real-time sensors (IoT)

If the modeler was wrong or if the excess of data revealed an unpredictable site, how do we actually reduce uncertainty? The Value of Information (VoI) theory — developed at Harvard by Raiffa and Schlaifer and applied to environmental engineering by Allan Freeze — gives the exact answer: doubt is only overcome by buying new information in the right place.

Instead of waiting another six months for a manual sampling campaign, the failure of the mathematical model should trigger a request to install connected sensors (IoT) directly at the critical points of the site:

• Where the models disagree: if two scenarios in the ECBYTS comparator fight over which way the groundwater is flowing, an automatic level and pressure sensor is installed at that exact point.
• Where the risk is immediate: if there is doubt about gases rising into an area with workers (0Z), a continuous vapor sensor is installed and sends alerts every minute.
• Where the chemistry changes fast: automatic probes measuring acidity (pH), conductivity and oxygen in the water warn immediately if a containment barrier has stopped working.

At ECBYTS, every time one of these sensors sends a reading, the entire model recalculates itself. The Model Owner, the technical manager and the auditor can see through the ECBYTS Seal exactly who validated each step and where uncertainty still needs to be fought.

Less guesswork, more engineering

Automating environmental studies does not mean letting the computer decide alone in the dark. It means using artificial intelligence and statistics to keep watch over our own ignorance about the site.

When the modeler fails or the data contradict each other, the system fulfills its noblest role: it tightens the belt on the Conceptual Site Model, clearly shows the size of the risk and demands that engineering put sensors in the ground to hear, in real time, what nature is doing.

The Scientific Foundation Behind This Vision (For those who want to go deeper)

• The danger of measuring too little over time: Shannon, C. E. (1949) — _Communication in the presence of noise_. It explains why rare manual sampling campaigns miss critical variations in the environment.
• Why several models look right but miss the future: Beven, K., & Binley, A. (1992) — _The future of distributed models: model calibration and uncertainty prediction_ (concept of equifinality).
• The rules of the game in Brazil: CETESB (2017) — DD No. 038/2017/C and ABNT NBR 16210. It defines the evolution of the Conceptual Site Model (CSM 1, 2 and 3) based on the reduction of uncertainties.
• Uncertainty standard used at ECBYTS (UNC-001): JCGM 100:2008 (GUM) — _Evaluation of measurement data — Guide to the expression of uncertainty in measurement_ & Monte Carlo Simulation.
• When it is worth installing a sensor (Value of Information): Freeze, R. A. et al. (1990) — _Hydrogeological decision analysis_ and Eidsvik, J. et al. (2015) — _Value of Information in the Earth Sciences_.