New Science Lab develops physics-aware AI solutions for the generative design and optimization of engineering systems whose performance depends on geometry, flow, and physical phenomena.

Engineering Intelligence: the market we operate in

Software for engineering simulation and analysis, CAE for short, is a mature, multi-billion-dollar market. For years it developed in one direction: increasingly accurate computation that costs increasingly more time and computing power. Artificial intelligence combined with the rules of physics changes that equation. Instead of computing every variant from scratch, a model can be taught how a given class of objects behaves, and hundreds of ideas can be checked in the time it used to take to check one.

The category attracts significant capital: the funding rounds of PhysicsX and Neural Concept confirm that AI for engineering is treated as a market in its own right, not a feature of existing tools.

We solve a specific class of engineering problems

Where flow decides the outcome

Flow-driven engineering covers systems whose performance depends on geometry, flow, and physical phenomena: hulls, impellers, propellers, turbines, wings, aerodynamic components, heat exchangers, and cooling systems.

Why these problems in particular:

• They share the same range of physics — flow and its influence on the behavior of the object. Experience from one area carries over to the next.

• Evaluating a single variant requires a costly simulation or experiment. In practice, a designer therefore compares a few versions, not a few hundred.

• The criteria are numerous and often conflicting, and they have to be optimized together — drag, stability, range, energy efficiency, payload, manufacturing requirements.

• Wherever geometry can be parameterized and physics computed or measured — that is the condition for a surrogate model to make sense at all.

Seven steps from idea to shape

For each class of objects we build separate geometry, physics, and models, because a hull and an impeller are entirely different worlds. The path from idea to finished shape, however, always looks similar.

• 1

  Geometry

  we parameterize the shape and check which variants are physically and manufacturably admissible for a given class of object.

• 2

  Physics

  we encode the laws and constraints of the flow environment, along with standards and manufacturing requirements, in a machine-executable form.

• 3

  Synthetic data

  from the encoded physics we generate training sets, controlling the coverage of the design space.

• 4

  Machine learning

  we build surrogate models that evaluate a variant in seconds instead of hours of simulation; active learning indicates which further simulations will contribute the most information.

• 5

  Verification

  an experiment and/or CFD is a mandatory checkpoint, with a record of the model version, iteration history, and the level of confidence in the result.

• 6

  Multi-criteria optimization

  we look for geometry that satisfies conflicting technical criteria simultaneously, instead of improving one parameter at the expense of the others.

• 7

  Export

  we hand over geometry and parameters in formats for further engineering work (including STEP), into the existing CAD/CAE stack.

Our solutions do not replace CAD programs or CFD simulation. They work earlier: they help decide which variants are worth computing at all.

HullGen, our first product

Early-stage ship hull design is too slow and too costly today. The conventional path requires multiple iterations and expensive CFD analyses, so teams compare too few variants, and poor decisions only surface at later, costlier stages of the project.

HullGen is meant to change that. Based on the given constraints and objectives, it generates successive hull variants and points to those that best meet the assumptions.

HullGen 1.0 works on the basis of 45 shape parameters, controls 49 geometric constraints, and evaluates hull behavior in water with 138 indicators. A single variant is produced in a few seconds.

HullGen 1.0 is at TRL 4. That means a working integrated technology demonstrator in a laboratory environment. It is not yet a deployment-ready product. An example result: a hull variant with approximately 27% lower wave-making resistance at the analyzed operating point (11 km/h); this is an internal result, prior to independent CFD and experimental validation.

We are currently building HullGen 2.0 on a new architecture, drawing on the knowledge and experience we have gained. The new version is being developed with small unmanned planing craft (USV) in mind. As part of this work we plan to carry out independent validation in cooperation with the Ship Design and Research Centre: CFD simulations and towing tank tests on physical models.

We show the existing demonstrator during a call. It is not a deployment-ready product, but it clearly shows how the tool works, the optimization logic, and the form of the output.

Next directions

The next direction in which we are beginning work is centrifugal fan impellers: a different object and a different environment than a hull, but the same type of problem. Further applications — propellers, turbines, and aerodynamic components among them — we consider selectively, based on market insight and a validation partner.

Where flow problems occur

The list below describes the market, not our product plans. It shows how broad the ground is on which a single specialization operates.

The common denominator: geometry affects flow, flow affects operating cost, and checking every idea requires a costly simulation or test.

An advantage that grows with every application

Every class of flow problem has its own geometry, physics, and models. With each application other resources accumulate and carry over, and that is precisely what builds the advantage.

• The knowledge of how to parameterize geometry and encode flow physics so that artificial intelligence can learn from it.

• Data: large synthetic datasets we have generated, and data obtained from simulations and experiments.

• Rules and verification methods which, once developed, serve in subsequent projects.

• Validation partners: research centers, universities, companies from the industry.

Let's talk