Research and Development

Research programmes with engineering outcomes.

Epcotec runs a research practice across three pillars: machine perception and remote data collection, the data and AI systems that turn measurements into answers, and optical and laser engineering. Applied research and development has run since the founding year 2011; peer-reviewed publication began in 2019. Every section is written for two readers at once. A business reader gets the value in the first paragraph. An engineer who reads on gets the physics, the failure modes and the discipline behind every claim.

Applied R&D service →Discuss a research engagement

Pillar I

Machine perception and remote data collection

Measurement-grade information collection in remote and hard-to-reach places, built on physics we can defend in front of an auditor.

Photogrammetry and structure from motion

We build measured models of ground, structures and assets from overlapping imagery, with the quality control that makes the numbers defensible.

Structure-from-motion turns overlapping photographs into dense, georeferenced point clouds and orthophotos. The processing itself is commodity; the discipline is the control. Ground control points surveyed to real tolerances, block adjustment checked against independent checkpoints, and uncertainty carried through to the delivered surface and stated with the export.

The same discipline applies to forensic and expert-witness work, where volumes and surfaces end up in commercial disputes and every figure must survive cross-examination. We have delivered auditable volume models where the answer a client could show an auditor mattered more than the answer that looked best.

RTK ground controldense point cloudsorthophoto productionauditable volumes

Lidar survey and point-cloud engineering

Laser scanning from tripods and vehicles, fused and classified into engineering-grade surfaces.

Lidar measures range by timing light, which gives surfaces that imagery alone cannot: vegetation-penetrated ground, structure edges at millimetre-class resolution, and geometry that holds under seasonal change. We fuse terrestrial, mobile and airborne captures into single coordinate frames, then run classification pipelines that separate ground, vegetation, structures and services into layers engineers can design against.

Point-cloud engineering is where most lidar programmes fail quietly: unclassified clouds look impressive and answer no question. Ours are classified, tolerance-checked and delivered with the documentation that lets a designer trust every surface.

terrestrial and mobile lidarvegetation penetrationclassification pipelinestolerance checks

Subsurface and penetration sensing

Ground-penetrating radar and controlled-source electromagnetic methods that see below the surface before anyone digs.

Ground-penetrating radar images the shallow subsurface by reflecting radio waves off changes in material properties: utilities, voids, pavement layers, foundations and moisture boundaries. Depth imaging is an inversion problem, and we treat it that way: velocity calibration against known targets, migration to collapse diffractions, and uncertainty stated in metres.

Electromagnetic conductivity methods complement radar where clay and salt water limit radar penetration. Used together, the two families cover most shallow investigation questions, from utility clearance before excavation to foundation and void searches under live infrastructure.

ground-penetrating radarutility and void detectiondepth imagingvelocity calibration

Electromagnetic survey methods

Transient electromagnetic and induced-polarisation surveys that map conductivity and chargeability at depth.

Transient electromagnetic methods drive current through a transmitter loop and listen to the decaying secondary field, mapping how conductivity changes with depth. The physics suits mineral exploration, groundwater mapping and geotechnical investigation where seismic access is impractical, and semi-airborne configurations extend reach across terrain that resists ground crews.

We treat electromagnetic data with the same scepticism as every other measurement: inversion outputs are models, models need constraints, and a conductivity section without an uncertainty statement is decoration. Our deliverables state what the data supports and what it does not.

transient electromagneticsinduced polarisationconductivity depth imagingmodel uncertainty

Magnetic and radiometric mapping

Magnetometry and gamma-ray spectrometry for geological mapping, infrastructure location and environmental baselines.

Magnetic field mapping locates ferrous structure: geological bodies, buried services, settlement of steel infrastructure, and boundaries between rock types. Gradiometer configurations sharpen anomaly geometry, and processed maps become direct inputs to geological interpretation.

Gamma-ray spectrometry maps the natural radioelements, potassium, thorium and uranium, which makes it a fast geological mapper and an environmental baseline tool. Combined with magnetic and spectral layers, it supports exploration targeting and land-management decisions with data that is repeatable and defensible.

magnetometrygradiometrygamma-ray spectrometrygeological mapping

Muon tomography

Cosmic-ray muons image the interior of structures too large or too sensitive for any conventional scan.

Muons from cosmic-ray showers penetrate hundreds of metres of rock, and their absorption pattern images density structure from inside: the interior of mountains, volcanoes, dams and large engineered structures. Detectors placed around a target record muon flux deficits, and tomographic reconstruction turns those deficits into density maps.

The method complements gravimetry and seismic imaging where access, safety or structural sensitivity rules out active sources. It is an emerging discipline with a small community of practitioners, and our interest is the engineering side: detector placement, exposure planning and the data pipelines that turn months of muon counts into interpretable models.

cosmic-ray muonsdensity imaginglarge-structure interiorsdetector placement

Gravimetry and quantum sensing

Gravity sensing, including cold-atom interferometry, for density mapping where other physics cannot reach.

Gravimetry measures density indirectly: voids, ore bodies, water tables and dense structures all bend the local gravity field. Relative and absolute instruments map those differences, and modern quantum gravimeters based on cold-atom interferometry bring drift-free absolute measurement into a field instrument.

Our work in this area is instrument-adjacent: understanding what each sensor class can resolve, designing surveys that separate signal from noise, and building the correction pipelines, tides, drift, terrain effects, that turn raw gravity readings into density interpretations.

gravity surveyscold-atom gravimetrysurvey designcorrection pipelines

Pillar II

Data, AI and telemetry systems

The software layer that makes remote measurement scale: edge computing, telemetry that survives the field, and AI that earns trust.

Onboard edge computing

Intelligence placed next to the sensor, so decisions do not wait for a data link.

Remote sensing platforms generate data faster than most links can carry it. Edge computing puts inference and triage on the platform: compress what is routine, flag what is interesting, and queue what needs the full pipeline. Power budgets, thermal envelopes and vibration tolerance decide what fits, and every design choice starts from those envelopes.

The result is a system that degrades gracefully: when connectivity drops, the platform keeps observing, keeps deciding, and ships the important things first when the link returns.

edge inferencepower budgetsstore and forwardgraceful degradation

Hybrid telemetry and link engineering

Link budgets across cellular, satellite and mesh paths, engineered so the data always has a way home.

Every remote programme lives or dies on telemetry. We design hybrid links, LTE and 5G where coverage exists, low-earth-orbit satellite where it does not, mesh backhaul where terrain allows, with link budgets calculated before hardware is bought. Bandwidth envelopes, latency tolerances and cost per megabyte are engineering constraints with numbers attached.

The same discipline covers store-and-forward policy, compression strategy and the failure logic that decides what to send when only half a link exists.

link budgetsLEO satellite backhaulbandwidth envelopesfailure logic

Event-driven cloud processing

Pipelines that turn raw captures into deliverables in hours, each stage triggered by the event before it.

Cloud processing at scale is an event-driven factory: capture completes, ingestion validates, processing fans out across workers, quality gates check products against tolerances, and deliverables land with provenance attached. The architecture matters because the alternative, manual processing queues, does not scale past a handful of projects.

We build these pipelines on mainstream cloud foundations, containerised workers, queue-driven orchestration, cost telemetry on every stage, with the same engineering rigour we apply to client platforms: reviewed, tested, observable and reversible.

event-driven pipelinesquality gatescontainer orchestrationcost telemetry

Machine-learning inversion and analysis

Learning methods applied where physics gives priors: geophysical inversion, spectral classification and change detection.

Geophysical inversion infers subsurface properties from surface measurements, and it is ill-posed: many models fit the data. Machine learning helps by learning priors from archives, guiding inversions toward geologically plausible solutions, and quantifying where ambiguity remains. We publish the uncertainty with the answer or we do not make the claim.

Active learning closes the loop: models flag the observations that would most reduce uncertainty, and the next collection targets those. Archives grow more valuable with every campaign because each labelled result retrains the base.

geophysical inversionlearned priorsuncertainty quantificationactive learning

Vision AI for inspection

Defect detection on imagery and lidar, with humans adjudicating what the model finds.

Vision models detect defects: cracks, spalling, corrosion, vegetation encroachment, deformation. The engineering is in the regime, because false-negative costs differ from false-positive costs, and a model tuned for recall floods inspectors with noise while one tuned for precision misses the crack that matters. We tune to the client risk, then put every finding in front of a reviewer who confirms, corrects and feeds the correction back.

The review queue is the product: model output becomes a shortlist, human adjudication becomes labelled training data, and inspection quality compounds with every adjudication.

defect detectionrisk-tuned thresholdsreview queueslabelled feedback loops

Spectral AI and living ground models

Signature libraries, foundation models for spectroscopy, and digital twins that stay current with every new measurement.

Spectral signatures identify materials the way fingerprints identify people, and the classifier is only as good as its library and its data hygiene: labelled spectra with honest provenance, wavelength-calibrated instruments, and cross-validation that catches leakage before a client does. Spectral foundation models trained across hyperspectral archives extend the approach to new instruments and new geographies.

The long game is the living model: archives catalogued to open standards, epochs comparable, and every new measurement fused into a ground model clients subscribe to, one that stays live between campaigns. That is the difference between delivering reports and maintaining knowledge.

spectral librariesfoundation modelsopen cataloguesdigital twins

Pillar III

Optical and laser engineering

Linear optics and laser systems engineered from published physics to measurable, transferable hardware.

Beam shaping and structured light

Vortex, Airy and Bessel beams, spatial light modulators and wavefront metrology under one discipline.

Beam shaping rewrites what light does: vortex beams carry orbital angular momentum and trap particles in rings, Airy beams bend around obstacles without diffracting away, Bessel beams regenerate after obstruction. We shape beams with spatial light modulators and micro-electromechanical mirrors, then verify the result with wavefront and polarisation metrology.

Optical trapping extends shaping into manipulation: particles held and steered in hollow beams, particle chains formed along beam axes, and traps operated at milliwatt powers that are safe to work around. The physics was published; our contribution is making it repeatable in an engineering environment.

vortex and Airy beamsSLM and MEMS steeringwavefront metrologyoptical trapping

Standoff elemental analysis

Laser-induced breakdown spectroscopy at a distance: elemental composition read without touching the target.

A focused laser pulse forms a microplasma on a distant surface, and the light it emits names the elements present. Standoff LIBS pushes the geometry from bench centimetres to field metres with pulsed lasers, gated detectors and collection optics, and it answers questions contact methods cannot: what is that rock face, that coating, that residue, without putting anyone or anything at the target.

Quantification at standoff is signal engineering: calibration curves, plasma temperature control and spectrum-normalisation strategies decide whether numbers or only identifications come out. We publish the error bars with the concentrations.

LIBS at distancegated detectioncalibration curveselemental mapping

Standoff molecular analysis

Raman spectroscopy at range, with fluorescence defeated by shifted-excitation and deep-ultraviolet methods.

Raman scattering carries a molecular fingerprint: minerals, salts, plastics and residues each with a spectrum as distinctive as a barcode. Standoff Raman pairs a pulsed laser with a gated, telescope-collected spectrograph, and time gating plus wavelength selection suppresses the fluorescence that would otherwise bury the signal.

Where fluorescence persists, shifted-excitation Raman difference methods and deep-ultraviolet resonance excitation separate the fingerprint from the background. The paired use of standoff elemental and molecular analysis on one target, elements and molecules together, is the capability frontier of this line.

standoff RamanSERDSdeep-UV resonancetime-gated collection

Spectral intelligence for materials

The software layer under the optics: models that name materials from spectra, trained on libraries with honest provenance.

Turning a spectrum into a material name is a classification problem with a physics twist: spectra shift with instrument, illumination geometry and temperature. Models trained across instruments and archives, with wavelength alignment and preprocessing treated as first-class engineering, hold accuracy where laboratory-only models fail.

We build the libraries as carefully as the models: documented provenance, measured uncertainty, and public reference data where it exists. A named material carries a confidence; an unnamed one stays unnamed.

material classificationcross-instrument trainingreference librariesconfidence scoring

Autonomous observation systems

Scheduling, safety interlocks and constrained onboard inference for systems that decide what to observe next.

Autonomy in measurement is not autonomy for its own sake. It is the engineering that lets a sensor system choose its next observation under constraints: observation value, safety envelopes, power, and the rules that bound every decision. Onboard inference is constrained by design, the safe action set is explicit, and hard interlocks sit below every clever layer.

The reference point is planetary science, where spacecraft have scheduled their own observations for years under the same discipline: bounded autonomy, explicit safety envelopes, and humans reviewing every exception. We apply the same pattern to terrestrial observation systems.

bounded autonomysafety interlocksobservation schedulingconstrained inference

Discharge control and filament research

Laser filaments and particle-seeded channels that steer electrical discharges, toward lightning protection and static-dissipation hardware.

An intense ultrashort laser pulse self-focuses in air and forms a filament: a thin, conductive channel that a discharge prefers to follow. Field experiments at a European mountain tower have guided real lightning along laser filaments, and the civilian prize is protection: lightning threats to wind farms, airports and launch pads cost their sectors billions each year.

Our counter-path is gentler and cheaper in power: a shaped beam floats a chain of light-trapped microparticles that stages the breakdown locally. Published work demonstrated discharge guiding at hundreds of milliwatts. The engineering targets are aviation static dissipation and pre-ionised channels for accelerator research, pursued as bench-scale, instrumented, safety-reviewed programmes.

laser filamentsparticle-seeded channelslightning protectionstatic dissipation

Depth discipline

Three levels, one standard

Every discipline on this page is held to the same three-level structure, so a business reader and a specialist reader get what each came for.

Positioning claim

One sentence per discipline, on this page and in the service tiles: what the capability is and the standard it is held to.

Structure

The section leads: what each method does, what problem it answers, and what the deliverable looks like.

Deep detail

The technical bodies: physics, quality control, failure modes and the engineering discipline behind each claim, for readers who check.

Have a measurement or research problem?

Bring the physics question, the terrain or the dataset. We answer with a feasibility plan, not a brochure.

Start the conversation →