The 3D view needs WebGL.
Drag the machine to turn it · slide to open its digital twin · scroll to go inside
Physics × data × AI
Combine physics, operational data and AI to understand industrial systems — and predict what happens next.
Nobody puts a thermometer on every tube. A physics model fills in what the sensors can't reach: hot water giving up its heat along the tubes, cooling water weaving around the baffles, and the thin layer at each tube wall where most of the resistance to heat flow lives.
Inlet and outlet temperature and cooling-water flow, read from the real pipework, feed the model continuously. From them it re-estimates how well the exchanger is transferring heat, so the physics follows the machine as it ages instead of drifting away from it.
Equations know how heat moves through a tube. They don't know that this exchanger fouls faster in summer, that one sensor reads a little high, or how the plant upstream behaves on a Monday. A machine-learning model trained on months of the plant's own data learns those things, and hands the physics better inputs than a fixed guess would.
The AI model forecasts what the plant will send in; the physics turns that into what the machine will do. In this demo the forecast is deliberately simple — a trend on the inlet temperature. Because its output goes through the physics, it can't predict a machine that breaks the energy balance.
The forecast is only useful once someone can act on it: the outlet will cross its limit within the half hour, this much cooling flow holds it, and this is how long until the tube bundle needs cleaning.
Demonstration asset with synthetic operating data. Physics: ε-NTU model of a one-shell, two-tube-pass exchanger, its conductance re-estimated from the measurements. The fields are simplified from that model, not CFD.
One high-fidelity simulation of flow, heat or structure can take hours to days on a cluster, and a real design needs hundreds of them. So most teams check a handful of options and hope. AI models trained on the physics change the arithmetic: they learn from simulations and measurements, then design, simulate and predict in minutes instead of days, with the physics still the judge.
A full simulation solves the equations from scratch for every design you ask about. An AI model trained on those simulations has already learned how the answer changes with the design, so asking it again costs almost nothing. The clock below is logarithmic: each step is roughly sixty times the last.
To make it concrete, follow one steel bracket through the three steps an engineering team goes through: design it, check it holds, and keep an eye on it in service. The stress analysis is real, solved for this page, and the same numbers carry through all three steps. With a fast model, thousands of variants like it can be checked instead of five. And once it is installed, the model keeps working: it turns one strain gauge into the stress where no gauge can go.
Drawing and 3D model: 176 × 120 mm, a 12 mm fillet where the arm meets the flange, a 14 mm pin hole for the load.
Finite-element solve: 5,636 elements, 3 kN on the pin. Stress peaks at the fillet, 138 MPa, 2.6× below yield. Bend drawn ×45.
SG-01 · strain 408 µε AI · fillet stress 138 MPa A strain gauge on the top face; the AI model, trained on the simulations, infers the fillet stress from it, live.
The bracket is simple on purpose. The same loop, a physics model, data from sensors and an AI model trained on both, works on far more complex assets: rotating machinery, ship hydrodynamics, aircraft structures. Every asset hides the part you most need to know. Each image below is cut the same way: the left half is what a twin computes, the right half is what you can see.
Turn vibration, pressure and power readings into early warning of wear and cavitation, before an inspection round would have caught it.
Replace hour-long hydrodynamic simulations with models that answer in seconds, so a design team can explore instead of only validating.
Air loads the wing, the wing bends, and the bend changes the air. AI models learn that two-way coupling, so a design can be checked in minutes instead of days of coupled simulation.
I built a local pipeline that reads part drawings, rebuilds them in 3D and prepares laser-cutting quotes. What works, and why fine-tuning on your own drawings pays off.
Construction drawings are precise and hard to read at a glance. Here's what it takes to turn a set of them into an interactive 3D warehouse you can rotate, cut away and inspect.
That is the problem I like: a physical system, an expensive model, and a question that needs an answer in minutes. Whether it's a design to explore or an asset to predict, glad to compare notes.
Get in touch