THE QUESTION EVERYONE ASKS

ARTIFICIAL INTELLIGENCE

Everything else on this site describes engineering: research and development, software, the tools in between. So what about AI?

It is the question behind every conversation at the moment, and it deserves a straight answer rather than a paragraph of the word “innovation”.

We were among the first companies in Italy, and among the first anywhere to build an AI tool that accelerates numerical analysis — in late 2022, before the LLM age made the subject unavoidable. Neural networks standing in for the expensive part of a simulation, so a result arrives while the design decision is still open.

That work grew past the point where it could be one capability among several. It needed its own team, its own roadmap and its own customers — so we spun it off into a company of its own.

THE SPIN-OFF: SPAIDER

Spaider provides AI-driven solutions for engineering companies. In the design phase those solutions accelerate numerical simulations, cutting both time to market and design cost.

Specifically: neural network architectures trained to accelerate numerical simulations far enough to predict results in real time, during the geometric modelling itself — so a designer sees the consequence of a change while making it, rather than the next morning.

The approach has a name: physics-AI. A network trained only on data learns to imitate results, and fails the moment it is asked about a case outside what it has seen. Here the governing equations are built into the training itself, so the model is penalised for disagreeing with the physics as well as with the data.

WHAT IT DOES

Trains neural networks that speed up and improve numerical simulations, and enable real-time performance prediction.

WHO IT IS FOR

Engineering and manufacturing companies — anyone whose design process depends on running a significant number of simulations.

WHERE IT CAME FROM

A team of mechanical and aerospace engineers, most holding PhDs, who worked together at NablaWave before founding it.

SPAIDER TECHNOLOGIES

Different data suits different architectures, and the model is chosen case by case rather than by fashion. Two families are in use:

PHYSICS-AI

Convolutional neural networks — Leveraging spatial hierarchies to extract patterns efficiently and optimise learning.

Graph neural networks — Capturing the relationships between data to unlock complex optimisation in structured data.

Transformers — Attending to what matters across the whole domain, predicting fields without using a mesh.

Physics-informed neural networks — Embedding the governing equations into the network, learning more physics from less data.

Generative geometry — Generating new shapes from a learned design space, exploring configurations nobody drew.

OPTIMISATION

Local methods — Optimising through the power of gradients, finding the steepest path to minimal cost.

Global search methods — Exploring the entire solution space, so no optimal point is left behind.

Surrogate modelling — Approximating complex functions to speed up optimisation without giving up accuracy.

A physics-informed network: the governing equations folded into the training loss

A predicted flow field through a compressor cascade, and the design space it was sampled from

APPLICATIONS OF SPAIDER-AI

The technology is not one model applied everywhere. Spaider builds verticals: edge AI trained for a specific class of problem, in sectors where a simulation is the bottleneck and a designer is waiting on it.

Four verticals are already available:

the thermal behaviour of a power transformer enclosure predicted in seconds;
the external aerodynamics of a MotoGP motorbike, from a shape a designer has just drawn;
the VOF hydrodynamics of a ship hull without the multiphase run;
transonic compressor blades explored across a 32-mode shape space.

More follow the same pattern: take a costly analysis, train on it, give the answer back while the design is still open.

The two companies still work together: where a problem needs physics, instrumentation and a bench, it stays here; where it needs a surrogate model trained on thousands of runs, it goes there.