From plant-floor telemetry
to closed-loop economic value.
Axiom Process Technology works on one question: can this operation prove what it believes about itself? We establish what the instruments underneath a decision are actually capable of supporting, we say plainly where the evidence will not carry the weight being put on it, and we build control, detection and reporting only on what survives that test. Every controller and every alarm earns its place against numbers the operating team can defend, every shift, on the screens that actually run the plant.
Operating excellence is an
intelligence problem first.
We close the distance between raw process data and the high-value decisions that move throughput, recovery, and unit cost. Every engagement opens on a signed baseline and closes on a system the operating team owns, runs, and continues to evolve.
An industrial intelligence consultancy
built on the plant floor.
Axiom Process Technology is a South African industrial intelligence firm operating at the absolute intersection of mining, metallurgy, process control, data science, and machine learning. Boutique by design, accountable by engineering discipline, and committed to leaving every operating team more capable than we found it.
Axiom was founded at the intersection that generic AI vendors and traditional control houses both miss. AI vendors treat process operations as a commodity data problem and stall on the plant floor. Traditional control houses lack the modern machine-learning depth to convert sensor abundance into closed-loop economic value. The intersection is where the work actually lives, and the first question we ask there is not which model to build. It is whether the instruments the decision rests on are telling the truth, because a control layer built on an unverified measurement inherits the error and then hides it behind a confident number. Once that is established, what we build is a five-layer control architecture: a healthy DCS at the foundation, gated by a formal Control Readiness check; an anomaly detection layer guarding it against sensor dropout, sticky valves, and dead-banded transmitters; soft sensors converting opaque process states into continuously controllable signals; an Advanced Process Control layer above them; and an AI Optimisation Layer at the top, recommending setpoints the operating team accepts or rejects.
The architecture is applied where it earns its place. The APC layer stabilises metallurgical operations: crushing, milling, comminution, flotation, and pyrometallurgy circuits, where multivariable control and feedforward soft sensors compound into measurable throughput, recovery, and energy gains. In energy and utilities, the same optimisation layer schedules specific energy consumption against tariff structure, demand limits and self-generation, treating energy as a constraint inside the process problem rather than as a separate system beside it. In manufacturing, the soft sensors and APC layer hold quality, yield and throughput across the production line, most closely where the process is continuous or batch and the physics resembles the circuits this practice was built on. Reliability engineering runs in parallel on the asset side of every operation, contributing probabilistic Remaining Useful Life forecasting, condition monitoring, computer-vision inspection, and risk-weighted CAPEX planning.
Every engagement begins with a site team's question, not an algorithm's offer. We build soft sensors from the plant's most valuable existing asset, its historian, alongside the SCADA tags, periodic laboratory assays, and maintenance records the operation already produces. No CAPEX push means we never propose instrumentation to make our own product work. Where the physics of a control objective requires a measurement that does not exist, we say so plainly. We attempt the inferential route first and report honestly whether the estimate clears our published accuracy floor. If it does not, we scope a smaller control objective in preference to forcing the spend. The decision to instrument is always the operation's, and it is made against a number we have published rather than a claim we have asserted. We cannot sell you a sensor, a mill, a filter press, grinding media, a flotation cell or a DCS, because we make none of them, and there is no product line, distribution agreement or commission arrangement anywhere in this business against which a recommendation could be written. The deliverable is never a slide deck. The deliverable is a system the operating team owns and continues to evolve.
Multidisciplinary,
by deliberate design.
The Axiom leadership team operates at the exact intersection of mining, metallurgy, multi-disciplinary engineering, machine learning, and applied artificial intelligence. Senior practitioners with hands-on plant credentials, not consultants briefed at the airport. Deliverables carrying an engineering judgement are signed by a named engineer.
A chemical engineer with fifteen years of operating experience across the southern African mining industry. Ditebogo has held metallurgical roles at major PGM, base metal, and diamond producers, including an appointment as Technical Assistant to a Group Chief Operating Officer. As a senior consultant she led commodity market studies, techno-financial modelling, scenario analysis, and capital-grade business case development. She is also the founding Managing Director of Marula MM Resources.
Leads the technical function at Axiom, owning engagement scoping through to the models and control strategies that go live in the control room. His work sits at the intersection of metallurgy, process control, machine learning, and data science. Physics-informed soft sensors feeding fuzzy logic and model predictive control across milling, flotation, and furnace circuits. Work he has led lifted PGM recovery by three percentage points in 2025, measured against a signed baseline, and stabilised energy intensity across multi-circuit chrome and PGM operations.
Nineteen years across flotation, smelting and plant operations in the southern African mining industry. Helly started as a field metallurgist on reagent performance and flotation optimisation, then moved to the reagent supply side in vendor management. At Rio Tinto Richards Bay Minerals he held process metallurgy roles across the smelter and the final product plant, covering furnace operation, electrode consumption, slag chemistry and freeze lining integrity, before appointment as plant superintendent with responsibility for metallurgical accounting, bottleneck resolution and shutdown planning. He owns the pyrometallurgy practice at Axiom, and holds a senior technical role at Masingita Group.
Six sectors,
one engineering philosophy.
From the open pit through the concentrator and the furnace to the tailings facility that holds what is left, the platform meets each sector with discipline calibrated to its physics, its risk profile, and its operating cadence. We do not generalise across industries. We specialise inside them.
Mining
Mineral Processing
Tailings Storage Facilities
Pyrometallurgy and Smelting
Energy and Utilities
Manufacturing
Vision, mission, purpose:
anchored in the word itself.
An axiom is a fundamental, self-evident truth. A rock-solid starting point from which every other truth can be derived. We chose the name because the discipline embedded in it describes precisely what we build for industrial operations: an unshakeable foundation of data-driven truth from which every downstream decision becomes mathematically justifiable.
The definitive global foundation for absolute, data-driven operating truth in heavy industry.
To establish Axiom as the standard against which industrial intelligence is measured. A future in which every heavy-industrial operation, from the open pit to the smelter to the manufacturing line, runs on a continuously verified, physics-informed source of operating truth. From that foundation, self-optimising, zero-loss operations stop being an aspiration and become an engineering inevitability.
Strip uncertainty out of complex industrial operations through first-principles engineering and hybrid process intelligence.
We deploy fundamental engineering first-principles, hybrid process intelligence, elite control loops, and digital twin frameworks to convert raw process data into closed-loop economic value. Mass, energy, and hydrodynamic balances govern every soft sensor we ship. Fuzzy Logic and Model Predictive Control hold targets that move throughput, recovery, and energy intensity. Every claim is held against a countersigned baseline. Unit costs go down, throughput and energy targets are hit, and the operating team owns the resulting system.
Complex industrial variables can be mastered through indisputable, mathematically stable control.
The foundational why of the firm. We believe the most intricate thermodynamic, mechanical, and pyrometallurgical processes are not mysteries to be managed reactively. They are systems to be modelled, instrumented, and controlled with the same rigour as a structural engineering calculation. When industrial operations are governed by axiomatic truth rather than operator intuition, the boardroom-level outcomes (predictable cash flow, defensible unit costs, structurally lower risk) follow without negotiation.
Our Framework
Understand the business.
Then the data. Then build.
Every Axiom engagement runs through the same six phases, in the same order, whether the asset is a concentrator, a furnace or a tailings storage facility. The order is deliberate and it is not negotiable, because each phase exists to prevent a specific and entirely predictable failure in the one after it.
Most operations are data-rich,
yet insight-poor.
Generic data science firms do not understand the difference between a rougher and a cleaner circuit. They treat a flotation cell like an e-commerce funnel and ignore the hydrodynamic and kinetic realities that actually govern recovery. Conventional Advanced Process Control is frequently brittle. Rigid, linear models demand constant manual retuning whenever ore mineralogy shifts, and they sever the plant from enterprise data at the precise moment integration would deliver the most value.
Axiom resolves this tension. We speak both languages with equal fluency: deep domain physics and modern machine learning. We build physics-informed models that embed foundational physical boundaries directly into the learning architecture: comminution kinetics, flotation hydrodynamics, multiphase slag-melt thermodynamics, and continuum mechanics. By constraining the learned components to respect those balances, we integrate them with existing SCADA, DCS, and enterprise automation layers. Operational data becomes continuous, verifiable, and automated economic gain.
The backbone is CRISP-DM, the Cross Industry Standard Process for Data Mining, developed in the late 1990s and published as version 1.0 in 2000. It remains the most commonly cited reference process for analytics projects, and we use it for that reason rather than inventing a proprietary method with a memorable name. What we do add is the adaptation it never had, for an environment where the model writes setpoints to live equipment and being wrong costs production, safety and licence to operate.
Two things change in that adaptation. Commissioning becomes its own phase, because proving a model on a running plant with operators in the loop is a completely different activity from building it. And value monitoring becomes permanent rather than a closing report, because an industrial process keeps moving underneath a model long after the project team has gone home.
- Skip business understanding and you optimise the wrong thing beautifully
- Skip data understanding and the model inherits every bias in the instrument
- Skip the build discipline and you get a notebook, not a system
- Skip commissioning and the operators switch it off in week three
- Skip hand-over and you have bought a dependency, not a capability
- Skip value monitoring and the gain is claimed, never proven, and it quietly decays
One discipline,
two very different assets.
Each phase below is set out twice, because a concentrator and a tailings facility fail in different ways and need different evidence, even though the discipline that gets us there is identical.
What we establish in this phase 12 checkpoints
- Metallurgists, plant manager, operators and C&I engaged separately, because they do not tell the same story
- What has been tried, what failed, and where the frustration actually sits
- A red-pen walk of the circuit against the drawings, producing a corrected as-built
- A constraint and bottleneck register: what limits this operation today, and the evidence for each entry
- An installed-but-idle asset register, because the fastest value is often something already paid for
- Feed provenance: whether you can reconstruct what material ran when, and from where
- Engineer of Record, Responsible Tailings Facility Engineer and site custodian engaged directly
- Design report, raise method and deposition plan read before the first site visit
- The failure modes already named in your risk assessment, and the TARP that sits over them
- A physical walk of the wall, toe, decant and delivery line against the drawings
- What instrumentation exists, who reads it, and whether anyone trusts it
- How the water balance is currently understood, and where it does not close
What we establish in this phase 11 checkpoints
- A six-point audit of every critical tag: calibration, installation geometry, signal path, deadband, sampling rate, physical condition
- Loop health established rather than assumed: auto mode, valve response, cascade integrity, silent overrides
- Controller integrity judged from the historian, not from the colour of a faceplate
- Whether two years of history covers two years of operating envelope, or two years of one ore type
- Laboratory and sampling integrity, because recovery and grade are laboratory numbers first
- The route data takes out, agreed with IT and OT security in week one
- The same audit applied to piezometers, weirs, level instruments and survey control
- Whether the network resolves the sections your Engineer of Record considers governing, or only the ones that were easy to drill
- Survey continuity: a common datum, and a stage-storage curve that still describes the facility
- Whether the water balance closes, and by how much it does not
- Where instrumentation is absent altogether, which is often the honest finding
What we establish in this phase 10 checkpoints
- Grey-box construction: a first-principles core with data-driven residual layers on top
- Soft sensors for the states the plant cannot measure, controllers for the loops it cannot hold
- Validation on held-out future periods, never on the data the model was trained on
- Performance reported stratified by mode and feed type, with calibrated uncertainty
- A digital twin exercised across the operating envelope before anything touches the plant
- The five-layer detection stack configured to your facility rather than to a template
- Physical balances closed and thresholds set from your own data, not a vendor default
- Detection thresholds mapped onto your TARP levels with the Engineer of Record in the room
- Phreatic surface reconstruction validated against the design surface and manual dip readings
- False-alarm rate estimated on history before a single alarm reaches a person
What we establish in this phase 10 checkpoints
- Controllers run in advisory mode, with each recommendation accepted or rejected explicitly
- Performance tracked live against the baseline, visible to the shift and not only to management
- Training on how the system reasons, not just which button to press
- Alarm rationalisation, because an alarm nobody can respond to is worse than none
- A site acceptance test against criteria agreed in advance
- Detection runs in parallel with the existing inspection regime, never instead of it
- Every alarm reviewed with the Engineer of Record before the system is given escalation authority
- False-alarm rate measured in the field and thresholds tuned before TARP linkage is enabled
- Instruments verified in situ against manual readings across a wet and a dry period
- Reporting templates agreed with the people who will have to sign them
What we establish in this phase 10 checkpoints
- Closed loop activated on your team's call, with a documented reversion path
- Scale-out from the template line to the rest of the circuit, then across the plant
- Runbooks, model registry, retraining procedures and escalation path transferred
- You own the models, the code and the right to change them
- Support model agreed and staffed, with response times in writing
- TARP linkage enabled, each escalation carrying an owner and a required response time
- Instrumentation extended from the governing sections to full facility coverage
- Governance reporting automated for the Engineer of Record, the RTFE and the Accountable Executive
- A field maintenance schedule, because sensors that are not maintained become the risk they were installed to detect
- Your team trained to add, replace and commission instruments themselves
What we establish in this phase 10 checkpoints
- Value measured against the baseline, normalised, and reviewed on a fixed cadence
- Drift monitoring on inputs and predictions, with champion-challenger evaluation governing promotion
- Scheduled retraining, plus a documented trigger for unscheduled retraining
- Calibration and loop health re-audited on a defined interval. Phase 02 is not a one-off
- Adoption reported alongside the metallurgy: acceptance rate, uptime, and which constraints bound
- Water recovered, deposition efficiency and remaining storage life tracked against the baseline
- Detection performance reviewed openly: what was caught, what was missed, what was noise
- Instrument network health reported as a first-class metric alongside the facility itself
- Annual reporting pack assembled automatically and reconciled against the independent review
- Thresholds revisited after every raise, because the geometry the models were tuned on has changed
How the six phases
are packaged commercially.
The technical discipline is the six phases above. The commercial shape wrapped around it is fixed and public. It opens with a complimentary audit, formalises a countersigned baseline, proves the approach in a bounded pilot on one circuit, furnace or facility, and only then scales with ownership transferred to your team. Nothing is invoiced against a claim the baseline cannot test.
Nine practices,
one integrated framework.
Phase 03 draws on nine tightly coupled engineering practices. Each is governed by a single MLOps plane and a single accountability standard. The right tool for each loop and each facility, never an off-the-shelf product.
A 90-minute review can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
Mineral Processing Optimisation
Crushing, milling, flotation, leaching, adsorption, dewatering, and the water that returns. The full circuit, the whole circuit, controlled.
Mineral processing rewards the operation that controls variance, not the one that chases the mean. Recovery is a non-linear function of feed grade, mineralogy, particle size, pulp density, reagent suite, gas holdup, and froth phase behaviour, and these variables are coupled across the comminution, flotation, and water reticulation circuits. Axiom engages the full sequence as the single coupled system it physically is, never as a collection of unit operations to be tuned in isolation.
Mineral Processing Challenges
Mineral processing operations face a wide spectrum of structural challenges that erode throughput, recovery, and stability. The set below names the most common ones we engage on a typical concentrator, not the full taxonomy. Each is invisible to conventional control, and each yields to an Axiom engineered interface: a soft sensor that turns the unobserved into a continuously controllable signal.
The Axiom control hierarchy, five layers, one contract.
A control stack delivers value only when each layer does its own job, and only its own job. The Axiom architecture is five layers, from the DCS upward. Each layer operates at the timescale its physics demands, and hands a well-defined signal to the layer above.
AI Optimisation Layer
Recommends economic setpoints for the whole circuit. Every recommendation is surfaced to operators for explicit accept or reject before it is handed down. The layer carries the operation's regulatory and safety limits as hard constraints alongside the process and equipment constraints. Discharge quality, emissions, occupational exposure, water licence conditions and permit boundaries define the edge of the feasible region. They are not terms weighed against economics.
An economic optimiser without them will eventually find value on the wrong side of a limit, and it will find it faster than a human would. This is not a compliance disclaimer. It is a statement about what an unconstrained objective function does when it is allowed to run.
APC Layer
Coordinates multivariable moves against process constraints and disturbances, holding the plant on target through the noise that regulatory control alone cannot handle.
Physics-Informed Soft Sensors
Estimates the variables the plant cannot measure directly. First-principles physics with machine-learning corrections, calibrated against laboratory reality.
Anomaly Detection Layer
Watches the plant's healthy operating envelope and flags any departure. Overrides are not failures, they are training signal for the next revision.
Healthy DCS
Stable, well-tuned regulatory control on the plant's installed PLC and DCS. No layer above is engaged until the foundation passes a formal Control Readiness check.
What happens when a measurement fails
Every signal carries a health state, and the controller's response to each state is defined in writing before commissioning rather than decided during an incident.
Before an estimate is allowed to move a setpoint
Every soft sensor clears a published floor: an R² of at least 0.85 on held-out future periods, with RMSE and MAE both reported. Three further conditions apply, and all three exist because the floor on its own can be cleared by a model that is quietly unfit for control.
A 90-minute Circuit Data Audit can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
Hydrometallurgy
The circuit where time is the reagent you cannot buy more of.
Comminution decides what can be recovered. Hydrometallurgy decides what is. The two are separated by a dewatering stage and joined by a constraint most circuits never write down: the time available for the reaction to finish.
Throughput, delivered density and installed tank volume together set residence time, and of the three only tank volume is fixed. Yet residence time is almost never carried as a measured state, which means an operation can lose a third of it to a dewatering setpoint that drifted, and find out from a residue assay two shifts later.
The architecture below applies to cyanidation and carbon adsorption, to acid and alkaline leaching of copper oxides, to uranium leaching with resin or solvent recovery, and to atmospheric and pressure circuits on nickel and cobalt. The chemistry differs by commodity. The control discipline does not.
Hydrometallurgy Challenges
Leaching and adsorption circuits fail quietly. There is no froth to watch, no power draw that spikes, no obvious signature when a circuit begins running below what the ore could give it, only an assay that arrives after the conditions have already changed. The set below names the most common structural challenges we engage, not the full taxonomy. Each is invisible to conventional control, and each yields to an Axiom engineered interface: an estimator that turns an unobserved state into a continuously controllable signal.
A 90-minute Circuit Data Audit can establish what your leach data can actually support.
Complimentary, no commitment. We walk the assay cadence, the residence time assumptions and the accounting against the decisions each one is currently being used to make, and you leave with a documented short-list.
Dewatering and Water
The most consequential number on the plant is the one nobody owns.
Dewatering sits between the concentrator, the hydrometallurgical circuit and the tailings facility, and it sets the terms for all three. The density it delivers fixes residence time in the leach, water returned to the plant, and the dry tonnage that becomes rate of rise on the wall. It is usually managed as plumbing, by people measured on none of those things.
A calibration drift on a density gauge is a geotechnical error, not an accounting error.
Slurry density and solids fraction are related by a mixing rule that is not in dispute:
Rearrange for the volumetric solids fraction and differentiate, and a relative error in measured slurry density does not pass through to solids fraction one for one. It is amplified, by a factor of ρslurry / (ρslurry − ρwater). The amplification is worst exactly where most tailings streams run, because the denominator is the part of the density that the solids contribute, and at low density there is very little of it.
So a density gauge reading half a percent high at a specific gravity of 1.50 does not put your solids concentration out by half a percent. It puts it out by roughly one and a half percent, and dry tonnage with it.
Dry tonnage sets rate of rise. Rate of rise governs whether pore pressure can dissipate as fast as load is applied. That is the mechanism, and it means an instrument technician's calibration interval is quietly a geotechnical control. Nobody on the plant sees that chain, because it crosses three departments and two disciplines.
What has to be true before a controller is built on a slurry stream
Every item below is checked and written down before any model is trained. None of it is exotic. All of it is routinely assumed rather than verified, and each one fails in a direction that flatters the number rather than exposing it.
There is one test that settles whether the instrumentation on a slurry stream can carry a model. Integrate flow times density times solids specific gravity over a period, and compare it against surveyed deposited volume times measured in-situ dry density over the same period. Two independent measurement chains, one physical quantity. They either agree to a defensible tolerance or they do not.
If they do not, one of them is wrong, and every economic claim, every soft sensor and every rate-of-rise number downstream inherits the error without any indication that something is missing. We do not proceed to build on an open balance. We report the residual, we bound it, and we scope the remediation that closes it. That is frequently a smaller and less interesting piece of work than the one we were asked for, and we say so anyway.
The circuit runs on water it can never fully get back.
Mineral processing is water-intensive, and much of the water fed to the plant leaves locked in the tailings rather than in the concentrate. No operation recovers all of it: water is lost to entrainment in the tailings, to evaporation, and to seepage, so every concentrator runs on continuous makeup water. Where water is scarce, that makeup supply, not the mill, is the real limit on throughput.
Dewatering and reticulation are usually treated as fixed plumbing, yet thickener density, filtration, and reclaim are all manipulated variables that move recovery, reagents, and energy at once. The aim is straightforward: water can only be managed once it can be seen. Axiom instruments the full network as a live mass balance, every stream metered and reconciled, so the APC layer can hold water usage as a controlled objective and return makeup water to throughput.
Treated as a controlled objective rather than a consequence, dewatering density stops being a thickener setting and becomes what it physically is: a manipulated variable of the circuit two units downstream, constrained by the thickener's own limits but optimised against the recovery and the rate of rise it actually governs.
A 90-minute review can establish whether your solids balance closes.
Complimentary, no commitment. A Tailings Evidence Review if the question is the facility, a Circuit Data Audit if the question is the plant. Either way it ends with a documented short-list of where the measurement chain is thin, and neither creates an obligation.
Pyrometallurgy Optimisation
Kilns, furnaces, electrode columns, refractories. The slow physics that no PLC was ever designed to manage.
A pyrometallurgical operation is a coupled electrical reactor, thermo-chemical reduction vessel, and structural refractory system, fed by an upstream pelletising or sintering circuit that determines whether the furnace ever stood a chance. The architecture below applies equally to ferrochrome submerged-arc furnaces, copper flash and reverberatory smelters, nickel and PGM smelters, and the rotary kilns and pelletisers that feed them. Operations differ in chemistry, the control discipline does not.
Pyrometallurgy Challenges
Pyrometallurgical operations carry a wide spectrum of challenges spanning electrical, thermal, refractory, and chemistry domains. The set below names the most common ones we engage across submerged-arc furnaces, electric arc furnaces, flash and reverberatory copper smelters, and the kilns and sinter strands that feed them. The list is representative, not exhaustive. Each yields to an Axiom engineered interface: a soft sensor that turns the unobservable into a continuously controllable signal.
The same five-layer hierarchy, applied to the furnace
The control hierarchy that governs the concentrator governs the smelter. The layers, the cadences, and the contracts between them are identical. The variables, the constraints, and the consequences of failure are not.
AI Optimisation Layer
Recommends economic setpoints for the whole circuit. Every recommendation is surfaced to operators for explicit accept or reject before it is handed down. The layer carries the operation's regulatory and safety limits as hard constraints alongside the process and equipment constraints. Discharge quality, emissions, occupational exposure, water licence conditions and permit boundaries define the edge of the feasible region. They are not terms weighed against economics.
An economic optimiser without them will eventually find value on the wrong side of a limit, and it will find it faster than a human would. This is not a compliance disclaimer. It is a statement about what an unconstrained objective function does when it is allowed to run.
APC Layer
Coordinates multivariable moves against process constraints and disturbances, holding the plant on target through the noise that regulatory control alone cannot handle.
Physics-Informed Soft Sensors
Estimates the variables the plant cannot measure directly. First-principles physics with machine-learning corrections, calibrated against laboratory reality.
Anomaly Detection Layer
Watches the plant's healthy operating envelope and flags any departure. Overrides are not failures, they are training signal for the next revision.
Healthy DCS
Stable, well-tuned regulatory control on the plant's installed PLC and DCS. No layer above is engaged until the foundation passes a formal Control Readiness check.
What happens when a measurement fails
Every signal carries a health state, and the controller's response to each state is defined in writing before commissioning rather than decided during an incident.
A 90-minute Circuit Data Audit can locate the single highest-value constraint on the furnace.
Complimentary, no commitment. We review the furnace data landscape, walk the historian against the operating narrative, and surface the most economically significant intervention available in the next 60 days.
Tailings Storage Facility Management
Your slimes dam is built every hour of every shift. Measure it that way.
A tailings facility is designed by geotechnical engineers and then built, continuously, by a thickener, a pump and a pipeline. Axiom instruments that whole chain and turns pore pressure, movement, seepage and water balance into live measured states instead of monthly readings. Your Engineer of Record keeps every engineering decision. What changes is that the evidence never stops.
Most tailings monitoring tells you whether the sensor is talking
The harder question, and the one an alarm you can trust depends on, is whether it is telling the truth. Network liveness and device health answer the first. Nothing answers the second, because knowing that an instrument has failed and knowing that a structure is moving are different problems needing different methods, and a system that treats them as one produces alarms nobody acts on. Layers one to four establish that the data is real. Layer five, which generic anomaly detection does not contain, asks what real data says about the physical state of the facility.
A facility measured once a month is unobserved for the other twenty-nine days.
A monthly piezometer reading gives you twelve numbers a year. Hourly telemetry gives you 8,760, and with them a response function rather than a single value. You stop asking what the head is and start asking how it responds to rainfall, to rate of rise, to a change in underflow density. A response that changes is a warning. A response that disappears is either a drifting instrument or a clogging tip.
None of those questions exist at monthly resolution. This is an argument about resolution, not about technology, and it is the honest reason to instrument a facility. Every decision taken between two readings rests on an assumption.
Two facilities where the interval was the problem
Both are public record and both are stated as reported. Neither is offered as a sales argument. They are here because each one turned on evidence that existed but was not continuous, which is the specific failure the resolution argument above describes.
How it works
Six steps, running continuously, from the sensor on the wall to the report on the executive's desk.
Most monitoring starts at the wall. Ours starts at the thickener underflow.
What we install
No facility gets all of it. The plan is sized against your risk assessment and the failure modes your Engineer of Record has identified, and every item is justified in writing against a decision it will inform. We name the requirement, never the supplier. We make no instrument and hold no distribution agreement, so the specification below cannot be written toward a product we would be paid for. The choice of manufacturer is yours, and it is normally made against your existing site standards and support arrangements.
We therefore treat it as a screening layer, not a monitoring solution, and we produce a site-specific InSAR coverage map as a standard deliverable: where measurement points can actually be retrieved on your facility, where they cannot, and precisely what ground instrumentation closes each gap. Corner reflectors recover coherence on the beach; ascending and descending passes recover more of the movement direction.
Where a facility is raised with cycloned sand, the cyclone battery is not process equipment. It is the machine that manufactures the structural shell, and the fines content of its underflow sets the permeability and the drained strength of what gets placed. An apex wears open continuously, a roping cyclone is invisible from the control room, and a shift-based laboratory sand check is an order of magnitude slower than either. We instrument per cyclone: feed pressure on each inlet, feed density and flow on the manifold, acoustic roping and wear detection, and a fines content soft sensor calibrated against the routine laboratory result. The laboratory becomes the calibration, not the control signal.
Where the facility is spigot-deposited with no cycloned sand in the wall, we will tell you not to. The same capital placed into piezometer density, or into resolving the phreatic surface on more sections, does considerably more for your risk profile. Recommending instrumentation that does not change a decision is how monitoring programmes lose their budget, and we would rather keep yours.
Where the value lands
We do not publish a generic percentage return, because the honest answer depends on your facility, your raise method, your water price and what sits downstream. We publish the mechanism. During the audit we size each line against your own numbers and put them into the signed baseline.
We produce the evidence. Your Engineer of Record produces the judgement.
Tailings governance assigns named accountable people, and no software changes that. The Engineer of Record owns the design and the interpretation. The Responsible Tailings Facility Engineer owns day-to-day management. The Accountable Executive owns the risk at board level. Axiom is none of those roles and does not seek to be.
Reporting is structured against the frameworks you are actually held to, and against the part of them that can be tested. The Global Industry Standard on Tailings Management, published in 2020 and now under the custodianship of the Global Tailings Management Institute, comprises six topic areas, fifteen Principles and 77 auditable Requirements. The Standard's own Preamble states that the Principles "are intended to summarise the Requirements that follow and are not in themselves auditable". We map every detection, every threshold and every piece of evidence to the named, numbered Requirements, because a conformance claim made against a Principle is a claim against an object no auditor can test. Your Trigger Action Response Plan, which is the mechanism that turns a measurement into an action and which every detection in our stack maps onto. In South Africa, SANS 10286, the residue stockpile regulations under the National Environmental Management: Waste Act, and the dam safety provisions of the National Water Act where your facility falls within their scope.
Two commitments follow, and we hold both. Every automated output is reproducible from its raw inputs, because a governance record that cannot be reconstructed is not a governance record. And no automated system is positioned as a substitute for the statutory inspection regime, the independent technical review, or the professional judgement of the Engineer of Record.
A review of South African tailings monitoring practice in the SAIMM Journal found instrumentation "limited to standpipe piezometers", with drain flow measured manually on a monthly cycle. The Global Industry Standard on Tailings Management requires "electronic monitoring systems and…real-time monitoring", with "vibrating wire piezometers at critical areas".
That is not our characterisation of the market. It is a peer-reviewed statement, in a South African journal, of the exact distance between what most facilities in this country currently measure and what the Standard they are being audited against requires. The gap is the work.
Tailings Evidence Review
Ninety minutes, complimentary, with a technical principal. One question: what can this facility currently prove about itself between inspections? We walk your instrument set, your survey record and your TARP against what each trigger would actually need in order to fire, and you leave with a documented short-list of where the evidence is thin. No commitment, nothing invoiced.
A 90-minute review can establish what this facility can currently prove about itself.
Complimentary, no commitment. We walk your instrument set, your survey record and your Trigger Action Response Plan against what each trigger would actually need in order to fire, and surface where the evidence is thin.
Reliability Engineering
Probabilistic asset health, computer vision inspection, and lifecycle intelligence for the equipment your business actually depends on.
Predictive maintenance done properly is not a dashboard. It is a probability distribution, a financial impact, a lead-time-aware action trigger, and an audit trail. We build reliability platforms that translate raw maintenance records, sensor telemetry, and inspection imagery into the same currency the business actually speaks: forecast failure dates with calibrated confidence, the Rand exposure of each one, and the decision the engineer should take this week to defend the production plan.
A 90-minute review can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
Energy and Utilities
Energy is bought by one department and spent by another, and the number in between belongs to neither.
Procurement negotiates the supply agreement, manages the tariff and receives the invoice. Process operations decide how many kilowatt hours it takes to move a tonne through the mill or a charge through the furnace. Energy intensity is determined almost entirely by the second group, reported almost entirely by the first, and owned by neither. That gap is the work on this page.
One unmanaged half hour can set the price of the entire month.
Industrial supply agreements do not bill energy alone. They bill demand, and on most such tariffs the demand charge for a billing period is set by the single highest measured demand interval inside it, not by the average across the month. Two large drives that happen to start in the same interval do not cost a little more than two drives that start apart. They set a ceiling that every remaining hour of the month is then charged against.
The reason this survives is that a peak is an event, not a trend. It does not appear in a monthly energy report, it does not show up as a rising line on anybody's dashboard, and by the time the invoice arrives the interval that caused it is four weeks in the past and nobody can say which two drives it was.
Look at how the three billed quantities are owned and the gap becomes obvious. Energy rewards using less in total, and process operations can act on that. Time-of-use rewards moving load out of the expensive hours, and planning can act on that. Demand rewards never coinciding, and no single function on most sites is accountable for coincidence at all.
It is a scheduling problem with a control answer. Start sequencing, interlocks between the largest drives, and a demand forecast the control layer can see are ordinary engineering. They are simply nobody's job.
What the layer actually controls
Each item below is a constraint or an objective handed to the same optimisation layer that runs the process, not a separate energy system bolted alongside it. Energy competes with throughput and recovery inside one problem, because on a real plant that is what it does.
Demand is billed on an interval. If the plant historian stores that measurement at a coarser interval than the one the utility settles on, or a compression deadband removes the short excursions, then the peak that set your charge is not visible anywhere in your own data. You are then arguing about an invoice using evidence that cannot resolve the event it is about.
Before any load shaping is designed we check the metering interval against the settlement interval, the instrument transformer ratios actually installed against the ones in the drawings, the historian scan rate, compression and deadband, and whether the plant's own meters can be reconciled with the utility meter over a closed period. Where they cannot be reconciled, we report the residual and stop. An energy saving claimed against metering that does not close is not a saving, it is a rounding difference with a business case attached.
The asset side of the same practice
Energy infrastructure carries a high consequence of failure and a maintenance regime that is still frequently calendar-driven. The reliability methods set out under Reliability Engineering apply here without modification, on a different asset population.
A 90-minute review can establish whether your metering can settle an argument about your own invoice.
Complimentary, no commitment. We walk the supply agreement against the plant's metering and historian, and you leave with a documented short-list of where the energy evidence is thin and which constraint is worth writing into the control layer first.
Manufacturing
By the time the quality result comes back, the product already exists.
In most plants the authoritative quality measurement arrives after the material has been made: a laboratory result, an end-of-line tester, a coordinate measurement, an inspection panel. Everything upstream of it runs on the assumption that the last good result still holds. That assumption is usually correct. What it costs on the occasions it is not is carried as scrap, rework, concession and a downgraded batch nobody can explain.
A control loop closed by a laboratory is not a control loop.
Follow the sequence. A setpoint is held. Material is produced. A sample is drawn. It travels, it queues, it is prepared, it is measured. A result returns. Somebody decides. A correction reaches the line.
Every unit produced between the sample being drawn and the correction landing was made on an assumption, and it was already committed before anyone could know whether the assumption held. Widen the interval and you have not made a worse product, you have made an unmeasured one. The plant is not out of control in those hours. It is simply unobserved, and those are not the same condition even though they look identical on a trend.
This is structurally the same problem as a residue assay in a leach circuit, and it takes the same treatment. Build an estimator from measurements that already exist and already move at the speed of the process, hold it to a published accuracy floor before it is allowed to change anything, and then demote the laboratory from the thing that closes the loop to the thing that calibrates the estimator. The laboratory does not disappear. It stops being the bottleneck in a feedback path it was never designed to sit in.
The honest constraint is that this only works where the fast measurements genuinely carry the information. Where they do not, we will tell you the estimator does not clear its floor and scope a narrower objective instead of forcing one.
What the layer actually does on a production line
The architecture is the one described under Our Framework, applied to a line rather than a circuit. A healthy base control layer at the foundation, detection guarding it, estimators above that, control above those, and an optimisation layer at the top that recommends rather than imposes.
Overall Equipment Effectiveness is a product of availability, performance and quality. Multiplying them destroys the information an engineer needs, because two lines can report an identical figure for entirely unrelated reasons and the number gives no indication which. Worse, the choice of time basis, whether the denominator is calendar time, scheduled time or planned production time, moves the result substantially without a single thing changing on the line.
We therefore publish the three components separately, state the time basis inside the definition rather than in a footnote, and report which constraint bound in each period and how often. Where a line's figure improves because the denominator changed, we say that is what happened. This is the same standard applied everywhere else on this site: a number that cannot be tested against is not worth reporting.
A 90-minute review can establish whether a quality soft sensor is feasible on your line.
Complimentary, no commitment. We walk your process measurements against your laboratory record and the specification that defines saleable product, and you leave with a documented answer on whether the fast data carries the information, before anyone proposes building anything.
Live Demonstration
Advanced Process Control, demonstrably operational.
This is a real Axiom controller, running live in your browser. It manipulates air rate to hold froth velocity on setpoint, defending against ore-grade disturbance and sensor noise. Use the controls under the trends: toggle Ore change to inject a step disturbance, or Noise boost to inject high measurement noise, and watch it return the process to target without oscillation. This is the class of controller Axiom builds for flotation, comminution, and downstream metallurgical operations. The control code is real. The process it is controlling is a simulation running in this page. No plant data, no client data and no tag from any operation is used anywhere in this demonstration.
FrothVel_PV vs SP (cm/s)
AirRate_SP, manipulated variable (Nm³/h)
3×3 rule base, singletons {−3, −1.5, 0, +1.5, +3} Nm³/h
Modifier trims + gating
Let us build you a controller
that earns its place on the plant.
If your circuit is losing throughput to loop oscillation, sticky valves, drifting sensors, or a DCS that fights the plant, we can help. Every Axiom engagement begins with a Complimentary Circuit Data Audit. Once your foundation is stable, we build the robust multivariable controller that stabilises your circuits so you can optimise them for measurable performance gains.
Soft Sensor Architectures
for Mineral Processing Control.
A two-paper compendium on physics-informed soft sensors that close the latency between laboratory measurement and real-time process control. Designed as feedforward inputs to the Axiom hybrid Advanced Process Control layer, with operator-approved supervisory optimisation above. Authored by Axiom for public release; the methods are deployed under engagement on industrial sites.
A residual-based architecture combining the Morrell mill power model with Extra Trees regression, delivering a continuous, SCADA-only hardness signal suitable for Model Predictive Control feedforward. Extended to a geospatial mine-to-mill digital twin through Kalman fusion of blast block models, shovel GPS, and stockpile blending.
Download P-001A T+30 minute concentrate-grade prediction architecture stacking LSTM, 1D-CNN, and Random Forest models on physics-derived, reagent-intensity, and temporal-lag features. A ridge meta-learner produces both a point estimate and a 95% prediction interval, with SHAP attribution for metallurgist explainability.
Download P-002Measurement Validity
and Facility Evidence.
Two papers in preparation on the other half of the problem: not what a model can infer, but whether the measurements underneath it can carry the inference at all. Neither is published yet, and neither is available for download. They are listed here because we would rather state what is coming than let an empty page imply the library is finished.
The algebra set out on the Dewatering and Water page, developed in full: how a relative error on a density gauge propagates through solids fraction, dry tonnage and rate of rise, why the amplification worsens at the low densities most tailings streams run at, and what calibration interval that arithmetic actually implies.
A proposed reporting standard for what a deployed monitoring system achieves on a live facility, as distinct from what its instruments achieve on a datasheet: false-alarm rate per instrument-month, time to detect against a named rate of change, and minimum detectable change at the wall after transmission gaps, outages and drift.
Companion document. See also our Privacy Policy, POPIA-aligned, governing the operational and personal data we touch under engagement.
What We Decline
The work we turn down is the clearest statement of how we work.
A 90-minute review can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
What We Publish
Everyone publishes what a sensor can do in a laboratory. Nobody publishes what the system does on a live facility.
Instrument accuracy is well documented across this market. Millimetre figures for shape arrays, sub-millimetre for slope radar, stated ranges for piezometers. Every one of those numbers describes a device under controlled conditions. None of them describes what you actually get at the wall after transmission gaps, power outages, calibration drift and a wet season. That second number is the one that decides whether an alarm can be trusted, and it is the one we commit to publishing.
System-level performance floors
For any monitoring deployment we commit to publishing the following, per facility, on a stated cadence. These are the numbers a system can be tested against, which is the whole reason to publish them.
The estimator floor, restated
This standard is published in full on the Mineral Processing page, where it belongs alongside the control hierarchy it governs. It is restated here because it is the other half of the same commitment. No vendor surveyed in this market publishes any model accuracy metric at all, while all of them sell soft sensors.
An instrument specification is a claim about a device in a laboratory, and it is safe to publish because no operating facility can falsify it. A false-alarm rate per instrument-month on a named facility is a different kind of statement. It can be checked, argued with, and held against us at the next review.
We take the view that a monitoring system nobody can test is a monitoring system nobody should buy, and that applies to ours. Where a floor is missed, the number is published missed. That commitment is worth more than any figure we could put next to it.
The published record
Every figure on this site carries one of exactly two labels: Achieved, with the commodity, the year and the words measured against a signed baseline, or typical industry range, with the source cited. A number that can be given neither label is deleted rather than softened. This table is the record. It is short, and it will only lengthen when a closed engagement produces a figure that can carry one of those two labels.
A 90-minute review can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
Engagements
Five rungs, and every one of them ends somewhere you can walk away from.
Most engagements in this market open with a proposal for the whole programme, which asks a buyer to commit before anyone has established what the operation can currently prove about itself. Ours open with ninety minutes and a documented short-list. The ladder below is how we structure work, in the order we structure it. Every rung has a written scope agreed before it starts, a defined end point, and a deliverable that stands on its own even if you never take the next step.
- Instrument fitness register. Every instrument the evidence depends on, audited against calibration, installation, signal path, resolution and physical condition, with what fails and what it costs to fix.
- Solids and water balance closure attempt. Two independent measurement chains against one physical quantity. They close to a defensible tolerance or the residual is quantified and bounded in writing.
- Gap register against the 77 GISTM Requirements. Mapped to the numbered, auditable Requirements rather than to the Principles, with the evidence you currently hold for each and the evidence you do not.
- Every TARP trigger mapped to a detection layer. Which triggers are currently detectable, at what latency, by which instrument, and which are written into the plan with nothing behind them.
No two operations present the same job. A single instrumented section on a small facility and a programme across four facilities are not comparable pieces of work, and the instrument set we find on arrival changes the scope of everything built above it. A single published number would mean either quoting high enough to cover the worst case we might walk into, or quoting low and repricing once the work exposes what is actually there. We are not willing to do either.
What we will commit to before you have spent anything. The first rung costs nothing. The scope of every rung above it is written down and agreed before that rung starts. And nothing is invoiced against a number the signed baseline is incapable of measuring.
Ninety minutes, complimentary, with a technical principal.
A Tailings Evidence Review for a facility, a Circuit Data Audit for a plant. Both end with a documented short-list and neither creates an obligation.
How We Measure Value
Ore moves. A before-and-after comparison is not evidence.
Feed grade, mineralogy, hardness and market conditions all move underneath any comparison, and they move on the same timescale as the improvement being claimed. A difference measured before and after a controller was switched on is therefore not evidence that the controller caused it. We use three methods together, and we publish which one produced which number.
Alongside the metallurgical result we report adoption: acceptance rate by shift, implementation rate, the reasons recommendations were rejected, controller uptime, and which constraints bound and how often.
An optimisation layer running at a twenty percent acceptance rate is not an optimisation layer, and we would rather say so than report the recovery number on its own.
A 90-minute review can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
Security and Data Governance
The connection is one-way, and the models are yours.
Our platform reads from your historian and writes, when and only when your operating team has authorised it, to a bounded set of setpoints inside limits your engineers have set. There is no inbound path to your operational technology network. There is no mechanism by which an outage on our side can affect your plant, because there is no dependency in that direction. Everything below is written to be forwarded to your OT security lead without editing.
Reference architecture
A 90-minute review can locate the single highest-value loop in your circuit.
Complimentary, no commitment. We review the data landscape, walk the historian against the control narrative, and surface the most economically significant intervention available in the next 60 days.
Contact
Start with a structured Circuit Data Audit, the typical entry point to an Axiom engagement.
The Audit is complimentary, 90 minutes, and ends with a documented short-list of the highest-value control, sensing, or recovery interventions available to your operation. There is no commitment to proceed.