Table of Contents
Machine learning optimises phosphate beneficiation by predicting concentrate quality (P2O5, MgO) and recovery from feed, reagent and operating data before lab assays arrive, and by suggesting setpoints for metallurgists to test. It works best as a soft sensor and decision aid alongside online analysers and the lab, not as an autopilot.
Key takeaways
- Beneficiation removes clays, silica, carbonates and other impurities to upgrade the phosphate content; the steps depend on the deposit type and its impurities (IFA and UNEP, 2001).
- Magnesium, much of it carried by dolomite, raises the viscosity of phosphoric acid slurries and reduces their filterability, which is why MgO in the concentrate is a key target (Boukhsib et al., 2019).
- Published machine learning work on phosphate flotation is still mostly lab-scale, and industrial reports rarely publish accuracy figures.
- A simple baseline can be hard to beat: in one iron ore flotation study, holding the last lab value outperformed several tree-based models on plant data (Basilio et al., 2026).
- The project starts with data: lab assays with true sampling times, reagent dosages, feed information and tonnages.
What does phosphate beneficiation remove?
Beneficiation upgrades the phosphate content of the ore by removing contaminants and barren material before chemical processing. A few ores need no concentration at all. The impurities depend on the type of deposit (sedimentary or igneous), the associated minerals and the extent of weathering (IFA and UNEP, 2001).
Most of the world's phosphate resources are sedimentary marine phosphorites, and the largest sedimentary deposits are in northern Africa, the Middle East, China and the United States. In the United States, more than 95% of mined phosphate rock is used to make wet-process and superphosphoric acid, and the USGS defines marketable product as beneficiated rock with a P2O5 content suitable for phosphoric acid or elemental phosphorus production (USGS, 2026).
| Impurity | Typical removal methods |
|---|---|
| Clays and other fines | Crushing or grinding, scrubbing, washing, screening, hydrocyclones |
| Silica (sand) | Froth flotation |
| Carbonates (calcite, dolomite) | Calcination followed by slaking; flotation |
| Iron minerals (magnetite, hematite, goethite) | Scrubbing and size classification, or magnetic separation |
Source: IFA and UNEP, 2001.
The same guide notes that carbonates "may cause downstream processing problems and may reduce the quality of the end product". Each step in the table is a place where operating decisions, such as reagent dosage, pH, cyclone pressure or kiln temperature, change the grade and the recovery.
Why does MgO matter so much?
Magnesium is one of the undesirable impurities in wet-process phosphoric acid. It increases the viscosity of the liquid phase, slows gypsum crystal growth and contributes to fouling and phosphorus losses. In laboratory pilot tests, higher magnesium reduced the filterability of the phosphoric slurry and increased its apparent viscosity (Boukhsib et al., 2019).
At Djebel Onk in Tébessa, dolomite is abundant. A 2026 study of dry tailings from the Somiphos beneficiation plant found a high proportion of dolomite, which the authors call a limiting factor for direct enrichment. In laboratory tests, they tuned flotation reagents to aim for concentrates above 28% P2O5 with MgO below 1%, and obtained about 30% P2O5 with MgO below 1%. Calcination at 1000 °C reached 33.51% P2O5, but the authors note that its high energy cost limits large-scale use, and they see flotation as easier to integrate into existing circuits (Meniai et al., 2026). These are laboratory results on tailings, not plant figures.
For a model, this means the target is rarely a single number. A concentrate must meet several limits at once, usually a minimum P2O5 and a maximum MgO, while recovery should stay as high as possible.
What is the Algerian context?
The USGS estimates Algeria's marketable phosphate rock production at 2.0 million tonnes in both 2024 and 2025, with reserves of 2.2 billion tonnes (USGS, 2026).
Capacity is growing. Argus reported in July 2026 that Somiphos expects to complete a 1 Mt/yr expansion at Djebel Onk by mid-2027. The Integrated Phosphate Project, a partnership between Sonatrach and Sonarem, is scheduled to enter production in the first quarter of 2027, and its Bled El Hadba mine is already operating and building up stocks (Argus, 2026).
New circuits start with little operating history, and expanded ones see new ore. Both points shape what machine learning can do: a model trained on one ore and one flowsheet must be checked again when either changes. For the wider picture across exploration, mining and processing, see our guide to AI in mining.
Where can machine learning help in a beneficiation plant?
| Use case | What the model does | Main inputs | Reference for validation |
|---|---|---|---|
| Soft sensor for concentrate quality | Estimates P2O5 and MgO between lab assays | Reagent dosages, pH, flows, levels, feed information | Lab assays |
| Reagent and setpoint optimisation | Relates dosage, pH and time to grade and recovery; proposes trials | Plant history and planned tests | Plant trials and mass balance |
| Feed blending | Predicts concentrate quality for a proposed blend | Stockpile or mine assays, online analyser data | Concentrate assays |
| Feed chemistry measurement | Supplies continuous feed composition to the other models | Online analyser on the conveyor | Lab assays of the feed |
Soft sensor for concentrate quality. A paper at the SME 2025 Annual Meeting describes an AI-based flotation prediction model that forecasts P2O5 grade from real-time operational data, to help operators adjust flotation parameters. The abstract gives no accuracy figures and does not name the plant (AlMuhaisen et al., 2025). The method is the same as in other process plants; our guide to soft sensors explains how to build one.
Reagent and setpoint optimisation. In a laboratory study on silicate-rich phosphate tailings with 21.6% P2O5, researchers varied polymer depressant type and dosage, pH and flotation time, then trained neural networks on the results. The best test reached 84.6% recovery at 28.6% P2O5, and the networks predicted different optimum pH and flotation time for each depressant (Alsafasfeh et al., 2022). In a plant, the same idea means using history to choose which settings to test, then confirming them in trials. Our comparison of design of experiments and Bayesian optimisation covers how to plan those trials.
Feed blending. Blending decides what the plant receives. A conference paper on mineral blending states that OCP Group "primarily employs simple blending formulas" and that growing ore variability has made it harder to maintain concentrate quality (Xu et al.). The authors propose reinforcement learning for blending and routing; in two synthetic scenarios, their solver outperformed a traditional optimisation benchmark on average 73.5% of the time (Xu et al., abstract). This is a research result, not a deployed system.
Online analysers. Prompt gamma neutron activation analysis (PGNAA) measures bulk composition continuously on the conveyor belt, and Thermo Fisher lists phosphate among the materials where its PGNAA analysers have been deployed (Thermo Fisher). An analyser gives the feed chemistry that grade models and blending models need as inputs. It does not replace the model, and the model does not replace the analyser.
The logic resembles ore grade prediction for grade control upstream: cheap, frequent measurements estimate a costly one, and assays remain the reference.
What data do you need?
- Lab assays with the time the sample was taken, not the time it was analysed or entered. Include P2O5, MgO and every other element your quality specification limits.
- Tonnages and flows for feed, concentrate and tailings, so recovery can be calculated from a mass balance rather than assumed.
- Reagent dosages as delivered, not only setpoints, along with pH, pulp density and water source.
- Feed information: which pit, stockpile or blend was fed, and feed assays or analyser readings.
- Events: shutdowns, reagent batch or supplier changes, instrument calibrations and ore changes, so these periods can be labelled or excluded.
Most of this sits in the historian and the LIMS. Our guide to using process historian data for machine learning covers extraction and alignment, and industrial data quality problems lists what to fix first.
How do you build and validate a model, step by step?
- Define the target and the trade-off. Agree with the metallurgists which variable matters (P2O5, MgO or recovery) and what a point of grade or recovery is worth.
- Audit sampling and assays. A biased sample gives a biased model. Check where and how samples are cut, and how sample IDs link to times.
- Align the data in time. Shift each input by the residence time between where it is measured and where the concentrate is sampled.
- Set baselines. Use the last lab value held until the next one, and a linear model. In the iron ore flotation study cited above, the naive baseline reached R² = 0.64 on industrial data, ahead of a decision tree (0.55), random forest (0.57) and XGBoost (0.58). The best model there, an LSTM (0.75), still showed bias in its residuals (Basilio et al., 2026).
- Split by time. Train on older data and test on a later period that includes ore changes and upsets.
- Run in shadow mode. Show predictions to operators next to lab results for several weeks before anyone acts on them.
- Test setpoint changes in planned trials with the metallurgists, with a mass balance to measure recovery, before changing standard operating values.
- Monitor drift. New ore, new reagents or a new circuit will move the model away from the lab. Our guide to model drift monitoring covers what to track.
What are the limits?
The published evidence on machine learning in phosphate beneficiation is thin. The studies cited here are a lab study, a conference abstract without figures and a research paper on synthetic scenarios. None reports a validated plant result that you could reuse as a benchmark.
A model learns only the conditions in its history. If the plant has never processed a given ore blend, the model cannot be trusted on it, which matters for new deposits and expansions. Models also do not change the metallurgy: whether dolomite is removed by flotation, calcination or another route remains a decision based on testwork and costs.
What we have seen at KABAS AI
KABAS AI has not published a phosphate project, so we have no plant results to share. What we can describe is how our offer fits the approach above.
On our mining solutions page, Virtual Lab is positioned for process optimisation: testing scenarios virtually before changes are made in the field. That matches step 7: use the model to narrow down which reagent and setpoint changes are worth a plant trial. In formulation R&D, a different use case, Virtual Lab helped ENAD Shymeca reduce physical testing cycles by 40%. DALILE runs on-premise without an internet connection and is positioned to connect production data, equipment records and safety documentation at remote sites.
Our practical advice: start with one quality variable that costs money when it is late, usually MgO or P2O5 in the concentrate; check that the assays and sampling times can support a model; and keep the data and models on infrastructure you control. Our data science team can help with the data audit and the first model, and our post on AI at remote sites covers deployment where connectivity is weak. You can also contact us.
Sources
- USGS, Mineral Commodity Summaries 2026: Phosphate Rock, February 2026
- IFA and UNEP, Environmental Aspects of Phosphate and Potash Mining, 2001
- Boukhsib, Khamar, Kadiri and Omari, Effects of Magnesium Content on the Rheological Behaviour of Phosphoric Acid and on the Performances of his Industrial Manufacturing by the Dihydrate Wet Processes, IJESRT 8(7), 2019
- Meniai et al., Management and development of mining waste from the dry process of the Djebel Onk phosphate ore beneficiation plant (Somiphos), Tébessa, Journal of Mining and Environment 17(4), 2026
- Argus, Algeria to accelerate Annaba phosphate port expansion, 13 July 2026
- AlMuhaisen et al., Unlocking Optimal Flotation Outcomes: Leveraging AI to Enhance Phosphate Recovery Through Predictive Modeling, SME Annual Meeting 2025
- Alsafasfeh, Alagha, Alzidaneen and Nadendla, Optimization of flotation efficiency of phosphate minerals in mine tailings using polymeric depressants: Experiments and machine learning, Physicochemical Problems of Mineral Processing 58(4), 2022
- Xu, Arief, Elghali and Caers, Adaptive mineral blending and routing under uncertainty using reinforcement learning, IIMP proceedings, Track 1: AI and Data-Driven Decision Making (abstract)
- Xu, Arief, Elghali and Caers, same paper, background section
- Thermo Fisher Scientific, 40 years of process control with PGNAA
- Basilio et al., Silica grade forecast in flotation processes: Evaluation and Statistical Diagnosis of Machine Learning Methods, SBSI 2026 extended proceedings
Frequently asked questions
What is the difference between phosphate beneficiation and phosphoric acid production?
Beneficiation upgrades the mined ore into a concentrate by removing clays, silica, carbonates and other impurities, using steps such as washing, screening, flotation and sometimes calcination. Phosphoric acid production is the chemical step that follows, where the concentrate is attacked with sulphuric acid. The quality of the concentrate, especially its magnesium content, affects how well the acid plant runs.
Can machine learning replace the laboratory in a phosphate plant?
No. Lab assays are the reference used to train a model, to check it before use and to correct it once it is live. A model can estimate P2O5 or MgO between assays, but if lab sampling is cut too far, the model loses the data it depends on.
Which machine learning model is best for predicting phosphate concentrate grade?
There is no universal answer. Start with a naive baseline, such as the last lab value held until the next one, and a linear model. In a 2026 study on iron ore flotation, the naive baseline beat several tree-based models on industrial data, so a more complex model must prove it does better on your own plant data.
Does machine learning help with dolomitic phosphate ores?
It can help operate the circuit that removes dolomite, for example by estimating MgO in the concentrate between assays or by relating reagent dosage and pH to grade and recovery. It does not change the mineralogy. Whether flotation, calcination or another route is used remains a metallurgical and economic decision based on testwork.
How much data do we need to start?
The limit is usually the number of lab assays with reliable sampling times, not process data. You also need the reagent dosages, flows, pH and feed information for the same periods, and the history should cover the different ore types the plant treats. A data audit before modelling tells you whether this exists.



