Built by Quanterall · EU project 101103966 · TechUPGRADE
TechUPGRADE Digital Twin Platform
A living digital copy of an industrial plant: its drawing, its measurements, its physics and its operation in one place, its state continuously compared against live measurements from the real plant, and looked after by AI agents that propose changes for people to approve.
Built by Quanterall for the TechUPGRADE project (EU project number 101103966) around a thermochemical heat-storage reactor, and designed as a product that serves many organisations at once.
Watch the platform in 16 minutes
All films
Technical reports
Why a digital twin
A thermochemical storage reactor is a dynamic system whose most important state cannot be measured while it runs: how far the salt has reacted, the temperature field inside the beds, where the reaction front is. Operators see a handful of temperatures, pressures and flows on the outside of the equipment.
The TechUPGRADE Digital Twin closes that gap. It keeps a live digital counterpart of the plant, synchronised with its measurements, and combines four kinds of knowledge about it:
| Knowledge | How the platform holds it |
|---|
| Its drawing | the P&ID, recognised from the engineering drawing and bound to the live devices |
| Its measurements | every sensor and actuator, streamed over MQTT and kept as time series |
| Its physics | component and system models, 3-D CFD of the reactor, and a physics-informed neural network trained on the CFD that predicts a whole charge in seconds |
| Its operation | scenarios and state machines that put the plant through situations before they happen, and rules that turn live values into alarms and actions |
What it gives you
- Observe: the complete plant, as drawn in its P&ID and in 3-D, bound to live measurements, with a control board per machine.
- Understand: what cannot be measured (conversion, internal temperatures, heat taken up or released) from physical models and the trained network running alongside the plant.
- Anticipate: the plant run through scenarios before they happen, from one charge cycle to a year of hourly district-heating dispatch, with optimisation and uncertainty studies.
- Verify: real against simulated behaviour, compared continuously, with deviations, limit violations and anomalies caught as they happen.
- Act safely: automation and AI are kept apart. Explicitly configured rules can trigger their pre-authorised automatic actions. AI agents read, investigate and propose; any new or agent-proposed command to the plant, or any change to the software, requires approval by an authorised person. Proposed changes can first be tried on a test copy of the machine.
- Scale: new plants and machines connected by their owners, through a device profile, a connector, signed firmware and staged rollouts.
- No third-party AI services: the reference platform and its application data run on Quanterall's own NVIDIA DGX Spark. Its AI models are served through FVLLMP: most run on that machine, while the vision model for drawing recognition and the AR view runs on the second node of FVLLMP's two-owner pilot. When that model is used, the image or other input it needs is sent to that node, under the agreement between the two owners; the node's owner can technically inspect the work it receives. Nothing is sent to a public cloud AI service or a third-party AI provider. A customer's own installation would run on the customer's hardware; none exists yet.
Features
Features below are live on Quanterall's running reference platform unless marked otherwise. The platform is built for many organisations; installation on another server has been tested end to end, but it has not yet been deployed on an outside customer's server.
What a feature is
- Live
- running on the reference platform
- Tested end to end
- built and verified in a defined test, not in daily use
- Partial
- works, but incomplete
- Experimental use
- shows how far the platform carries; not a certified engineering result
Where it is deployed
- Deployment: not yet at a customer site
- implemented, but not yet installed on an outside customer's server; this says nothing against the feature itself
Live plant: IoT and Control Board
Live
- A map of every plant and machine of an organisation. Each opens its live diagram, wired to that machine only.
- Control Board: the plant's P&ID as a live operational view (the drawing itself is edited in the 2D Editor). Every sensor shows its value, a 6-hour trend and a prediction ahead. Authorised users can command supported actuators, according to their roles.
- Threshold alarms per sensor, with alert channels (in-app, MQTT, HTTP, Telegram, e-mail) and automatic actions.
- State rules: a known state of the plant captured from live readings or a simulation, then watched on the server on every reading, whether or not anyone has the page open.
- Fleet view: every machine in one table, worst first, with its open incidents and pending actions. A fleet digest can be exported as Markdown.
- A built-in self-test of 22 cases against any machine, and test machines: copies of a real machine whose commands never reach hardware.
- A 14-step setup wizard for a new plant.
Films Theta 5 kW reactor room in 2-D: the room's P&ID as a live control board, 42 devices bound to the reactor cycle.
2D Editor: from an engineering drawing to a live board
Live
- Upload a P&ID as PDF, SVG, PNG or JPEG.
- A vector PDF is read directly from its text and strokes. A scan or photo is recognised by our own vision model, served by FVLLMP, and nothing goes to an outside service.
- Several independent recognition passes are merged by position and label voting for higher accuracy.
- Review and edit: a palette of sensors, valves, pumps, tanks and lines, auto-arrange, and an overlay of the original drawing to check against.
- Publish to IoT shows the device contract first, then creates the devices and topics and puts the diagram on the Control Board.
Films 2D Editor
3D view, CFD and augmented reality
Live AR recognition: partial
- The plant in 3-D, rendered on the server's GPU and streamed to the browser: the reactor alone or the whole 5 kW room, with live valve positions and flows.
- CAD models are loaded from STEP, STL, OBJ or VRML.
- CFD inside the reactor: OpenFOAM runs a charge or discharge on the reactor's real geometry in about 2.5 minutes. The result is shown as temperature, conversion, pressure or velocity fields, with cut planes, isosurfaces, vapour flow and a playable timeline.
- Every function is a tool a script or an AI assistant can drive, including through a plain-language panel.
- AR view on a tablet: the camera recognises the reactor and the room's equipment, and shows the reaction inside the vessel. (Recognition is partial.)
Films Theta 5 kW reactor room in 3-D · 3D Visualization
Simulation Manager
Live
- Build a system on a canvas from a library of blocks (with their equations, parameters, ports and units), from templates, from a saved system, or directly from a published P&ID.
- Fixed-step co-simulation: any topology runs, recycle loops included. Runs are queued, tracked and fully reproducible, with the solver version and a hash of each model's code kept with every run.
- Three fidelities for the reactor: a fast lumped model, the trained PINN (a whole run in seconds), or full CFD. Results are cached by their inputs.
- Studies: parameter sweeps, design of experiments (Latin hypercube), sensitivity analysis (Morris, Sobol), optimisation with constraints, and Monte Carlo uncertainty, each estimated before it starts.
- Your own components: researchers write components as equations or as restricted Python. Each one is unit-checked and test-run, and admin-approved before use.
- Connected to the plant: start a simulation from the plant's current values, replay a recorded window, and validate a run against the recording (RMSE, MAE and more).
Films Simulation Manager · Salt alternatives: the reactor with 15 other salts from the NBS tables, checked against DLR's measurements.
PINN: physics-informed neural networks
Live
- The reactor's bed modelled in 16 shells, each with a temperature and a conversion: the physics, calibrated on the 3-D CFD runs, plus a learned port-Hamiltonian correction.
- In an internal comparison, about 800 times faster than the CFD it was trained on, for a whole charge or discharge (OpenFOAM takes about 2.5 minutes for one on the reactor's geometry). Held-out test error is 0.766 K in temperature and 0.0128 in conversion, on CFD runs the served model (trained 18 September 2026) was not trained on. A technical report with the hardware, the number of held-out runs and the range of conditions is not yet published.
- Soft sensing (the bed's inner state from what is measured), energy audits and diagnostics.
- Training runs from the browser with live charts, a held-out test, and a guarded promote step that puts a new model into service.
Films The PINN module
Scenario Manager
Live
- District heating: the reactor placed in a district-heating plant (heat pump, electric heater, thermal storage, waste heat) and dispatched hour by hour over a full year, to compare the cost of heat.
- Heating studies: high and low demand, high and low waste heat, system failure, emergency shutdown.
- Process scenarios that act on a plant through its controllers and check what follows.
- Custom scenarios built from a state-machine graph, validated before they run.
Films Scenario Manager
Comparison Manager
Live
- The real plant against the simulated one: time series, histograms, box plots, CDFs and cumulative sums.
- Statistics between series (RMSE, MAE, correlation), transformations (moving average, FFT, autocorrelation, lagged cross-correlation), heat maps, threshold exceedance and KPIs.
- Annotations, reference curves, parameter differences between runs, and forecasts.
- Live mode: a real machine's sensor beside its test copy's, with injected faults, live thresholds and anomaly marks.
Films Comparison Manager
AI agents: owners for every component and machine
Live
- Component owners: one AI agent per platform component. It watches its metrics, learns their normal ranges, keeps a knowledge base of confirmed fault cases, and writes proposals for improvements.
- Gates: every proposal passes human approval, then a staging release, then the production deployment. Agents only read. Anything that writes waits for a person.
- Device owners: every machine has a technical owner agent. It triages alarms into incidents, writes work orders for technicians, proposes actuator commands for an operator to approve, and keeps the service history and reliability view.
- Experiments on test machines: the agent investigates a fault on a copy of the machine, screens candidate causes, and records what it learns as a fault case that is then watched for. Owners even sit exams on known faults.
- Owners ask each other: a machine's part owners and related components are consulted, and every answer can be rated helpful or not, so they learn.
- Runs on our own language models on our own GPUs. Nothing goes to an outside AI service.
Films Component owners · Short version · TechUPGRADE AI agents · AI infrastructure
Research: from a paper to a paper
Live
- A research agent reads arXiv every day for new physics-informed methods and scores each paper for the platform.
- Anyone can search in plain words with a purpose, or add a paper by its DOI.
- An approved paper is reproduced on the platform beside the paper's own numbers. The agent then writes the comparison as an arXiv-style report, with the results exactly as computed and the time from search to report.
Films Research: from a paper to a paper
Connectors, firmware and fleet
Live Deployment: not yet at a customer site
- Device profiles: a client describes its machine's devices, protocols (Modbus TCP, REST, MQTT and more) and state machine in a six-step wizard. Devices can be imported from a CSV register map, and the connection tested before saving.
- Published profile versions never change. Connectors can be upgraded or rolled back between versions, with the difference shown.
- Connectors are deployed over SSH to the client's gateway from the platform, with host keys pinned, and connect back over TLS (MQTT 8883). Heartbeats, telemetry, live settings and logs are visible from the platform.
- Firmware library with signed firmware (Ed25519). The platform never holds the maker's private key, and every install is verified and logged.
- Fleet and staged rollouts: a new profile version or firmware image rolled out to chosen gateways in order, stoppable at any step.
Films Connectors · The reactor's controller, connected
Deployment Manager
Tested end to end Deployment: not yet at a customer site
- Installs TechUPGRADE on another server (a test copy, staging, or a client's own installation) from the platform itself.
- The exact commit is built on the target over one SSH session, with its own fresh secrets and its own TLS certificate, and with stages and a live log. (End-to-end tested. No deployment to an outside client server has been made yet.)
Films Deployment Manager
One platform, many organisations
Live
- Multi-tenant by design: every machine, drawing, connector, scenario and record belongs to one organisation. Another organisation's resource answers "not found".
- Roles: superadmin, admin, operator, researcher and viewer. Organisations have invitations and teams.
- The navigation shows each organisation only the features it has.
- Open to other systems: an external system can hand work to the platform through a signed hand-off (Ed25519, short-lived, single-use), and its users open the result with no second sign-in. GenHub2 is the first to use it.
Security
Live
- Passwords hashed with PBKDF2 (600,000 iterations). HttpOnly session cookies with CSRF protection, a strict content security policy, rate-limited sign-in, and every session ended on sign-out or a change of password or role.
- Organisations are kept apart at every service, not only in the user interface.
- A security event log kept for a year, readable by the platform's administrators.
- A security audit with 19 findings, published with their fixes and what is still open. Nightly self-tests and backups, and a penetration test prepared.
Films Security: why trust the platform
Use cases
The SrBr₂ thermochemical reactor (TechUPGRADE). The project's test rig: two strontium-bromide reactors in alternating charge and discharge, an evaporator, a condenser, a heat-transfer-fluid loop and their instruments. Its CFD runs, built from the reactor's CAD model, train the physics-informed network.
The Theta 5 kW reactor room. The 5 kW installation as a plant: 42 instruments on its P&ID, its 3-D room and its recorded history.
District heating. The reactor dispatched hour by hour over a year inside a district-heating system, to compare operating strategies by the cost of heat.
Beyond the reactor: experimental uses
Experimental use Not a certified engineering result
The same platform has been tested far outside its original reactor. These are experiments that show how far it carries, not certified results.
- Papers reproduced as their own organisations: a Hall thruster PINN (film), plume reconstruction with PINN–PONN, a distillation column, a composite autoclave cure and a solar refrigerator. Each comes with a technical report on the reports page.
- Engineering projects for a self-sufficient home: a 5 kW vertical-axis wind turbine taken from the site's wind to a production package (film), a cluster of five small vertical-axis turbines, and a pumped-hydro store for two weeks off-grid in winter (film). They are not certified designs. The results must be verified and approved by qualified professional engineers before any construction or production.
- Antarctica: a hut for two researchers at St. Kliment Ohridski Base, heated by wind, sun and the SrBr₂ reactor, and the whole platform with its AI running offline on one box at the station (film).
- Virus structures for GenHub2: genetic laboratories open the platform straight from their own results and see a virus's changes along the genome and on 213 3-D protein structures across 29 viruses, two of them AlphaFold predictions labelled as such; the joint test of the hand-off passed all 14 checks (film). The biology has not yet been reviewed by an independent virologist.
- Smart homes: a home's devices and cables planned in 3-D, wall by wall, with an AR view of what is inside the walls.
Technology
Erlang/OTP core (Nova, Cowboy) · React and TypeScript front end · EMQX (MQTT) · PostgreSQL, InfluxDB, Redis · Python services (FastAPI, NumPy) · OpenFOAM v2412 for CFD · ParaView and trame for server-side 3-D · OpenCV and a local vision model for drawing recognition · our own language models on an NVIDIA DGX Spark (ARM64, GB10 GPU) · Docker Compose.