Process Control System for In-Situ Uranium Mining

We designed a control and analytics system for in-situ uranium mining - for one of the largest in-situ leaching operators in the world. The task sounded simple: see what is happening inside the reservoir and make decisions faster. In practice it is about geology, hydrodynamics, ecology and economics all at once. Here is a calm, to-the-point account: what was required, how the solution is built and what it gives the operator.

What in-situ mining is and why it is hard to control

In-situ leaching (ISL) is a way of mining uranium without shafts or open pits. A leaching solution is injected into the ore-bearing formation through injection wells, it dissolves the uranium right underground, and recovery wells lift the productive solution to the surface for processing. The ore is not extracted - it is chemically recovered in place.

The advantage of the method is a lighter footprint on the surface and the absence of waste dumps. The difficulty is that the core technological process takes place where it cannot be seen directly: at depth, in a porous formation, surrounded by aquifers. The operator manages the process indirectly - through flow rates, pressures and solution chemistry at the wellheads of hundreds of wells.

The task: four loops that cannot be separated

The operator framed the requirements not as "we need a dashboard", but as four interconnected control loops. Each of them is solvable on its own, but value only appears when they are brought together into a single picture.

These loops conflict. You can raise injection intensity and speed up recovery - but reagent consumption and the risk of the solution escaping the contour will grow. You can tighten the regime for the sake of ecology and economy - but drag out the mining of the block. Managing ISL is a constant search for balance, not the optimisation of a single parameter.

How the system is built

We did not try to replace the process engineer with automation. The goal was different: to give the person responsible for the block a complete and timely picture of the reservoir and well-grounded decision options. The system is assembled from three layers.

Layer 1. Well monitoring

For each well - injection, recovery, observation - telemetry is collected: flow rates and injectivity, wellhead pressures, solution chemistry (pH, redox potential, concentration of uranium and accompanying components), reagent consumption. The data is aligned to a single time base and spatial reference: what matters is not a value at an individual wellhead, but the balance across the cell and the block.

• merging heterogeneous telemetry sources into a single stream with data quality control;
• the "injection - recovery" balance for each cell as the first indicator of a leak or crossflow;
• automatic anomaly detection: pressure drift, falling flow rate, rising concentrations at observation wells.

Layer 2. Digital reservoir model

A hydrodynamic model of the block runs on top of the telemetry. It links measurements to geology and shows what is not measured directly: the actual sweep of the ore body, the movement of the solution front, under-leached zones and residual reserves. The model is not static - it is calibrated against incoming data, so over time it describes the specific block more accurately rather than an averaged formation.

Layer 3. LLM analytics layer

This is the part the whole project was started for. The reservoir model and the stream of anomalies amount to a great many numbers and maps that a process engineer physically cannot work through across hundreds of wells. The LLM layer takes on the interpretation: it links scattered signals into a hypothesis, explains it in natural language and proposes courses of action with a reference to the data the conclusion is based on.

• turns a set of anomalies into a statement: what is most likely happening in the block and why;
• ranks events by their consequences - ecology and solution losses rank higher than a local drop in flow rate;
• prepares a draft decision (change the injection regime, switch a well, increase observation), but the final word stays with the process engineer.

What it gives the operator

We measure the effect not in a single nice number, but in how work with the block changes. A "before and after" comparison by the nature of decisions:

Aspect	Before	After
Reservoir picture	periodic reports, scattered spreadsheets	a single block model in real time
Response to an anomaly	noticed after the fact, following a round	signal and hypothesis at the moment it arises
Contour control	manual reconciliation of observation wells	automatic solution balance across cells
Engineer's decisions	rely on experience and fragments of data	rely on the model and verifiable reasoning

For the operator this comes down to three practical things. First - fewer hidden losses: solution and reagent are not wasted on under-leached or, conversely, bypass zones. Second - manageable ecology: deviations in the contour are visible before they become a problem. Third - retained expertise: the logic of experienced process engineers is captured in the system and available to the whole shift, rather than living only in the heads of individual specialists.

The main result is not automation but transparency. The process engineer makes the same decisions as before, but sees the reservoir as a whole and understands what each conclusion is based on.- from the project principles

What we took away from the project

In-situ mining is a special case of a task that comes up in many complex production settings: the process takes place where it cannot be observed directly, and it has to be managed through indirect signals. In such tasks the value is created not by a single model and not by a single dashboard, but by the combination: quality telemetry, a physically correct model of the process and an analytics layer that translates it into decisions in the operator's language.

Frequently asked questions

Does the system control the wells itself?

No. It collects data, builds a reservoir model and proposes decision options with reasoning. Changing the injection regime or switching wells is done by the process engineer - responsibility stays with the human.

Why an LLM here, if there is a hydrodynamic model?

The model produces numbers and maps across hundreds of wells. The LLM layer links these signals into an understandable hypothesis, ranks events by their consequences and frames a conclusion in natural language. This removes the routine of interpretation that a human physically cannot keep up with across the whole well field.

How is ecology controlled?

Through a continuous "injection - recovery" balance for each cell and monitoring of observation wells on the mining contour. Deviations that may indicate the solution escaping the boundary are detected automatically and raised as a priority.

Does this approach suit other production settings?

Yes - for any processes that run "out of sight" and are managed through indirect data: chemistry, metallurgy, energy, infrastructure. The three-layer architecture carries over; what changes is the physical model in the middle.