# Free Lunch Project URL: https://goudbeek.com/en/free-lunch-project Language: en Last updated: 2026-10-15 Provider: Goudbeek — Almere, the Netherlands — samgoudbeek2007@gmail.com > The Free Lunch Project builds specialised AI models for physical and chemical simulation: the same answer as a classical simulation, in seconds instead of hours. AI that computes physics in seconds instead of hours. ## Why In 1997, two scientists, David Wolpert and William Macready, proved that no algorithm is the best at everything. They called it the No Free Lunch theorem. Averaged over every conceivable problem, every algorithm performs equally well: whatever a method gains on one kind of problem, it loses on another. The one that wins is the one that best fits its problem. You see the same in AI. Language models like ChatGPT can do a lot: write text, explain code, answer questions. But they do not compute physics exactly. They predict words, and a number that looks right is not yet a calculation. They estimate. For a conversation that is fine; for a design a machine has to run on, it is not. So the great breakthroughs in scientific AI did not come from models that can do a little of everything. AlphaFold predicts how proteins fold, a question researchers had been stuck on for fifty years; that work was awarded the 2024 Nobel Prize in Chemistry. Weather models such as GraphCast make a ten-day forecast in about a minute, where classical weather models need a supercomputer for it. Each time, it was a model built for a single task. Meanwhile, engineers wait hours to days for simulations: how heat flows through a machine, how air moves past a part, how a material behaves under load. Every change to a design means computing again. So they test ten designs where they would like to test a thousand. The best design often stays unseen, and that holds back innovation in chips, energy and industry. There is no free lunch. There is a specialist that fits the problem exactly. ## What The Free Lunch Project builds specialised AI models for physical and chemical simulation. They give the same answer as a classical simulation, only many times faster. Specialised means one model per kind of physics. Not a model that knows a little about everything, but one that knows a single kind of problem inside out, the way a hospital specialist knows one organ better than anyone. Precisely because it does not have to do everything, it can do its one task quickly and precisely. For an engineer, that changes the way of working. Where a simulation now takes hours, the answer comes in seconds. Suddenly it becomes possible to compute a thousand variants of a design instead of ten, and to choose on the strength of what works rather than what happened to get tested. We start with heat, a core problem in precision machines and energy systems. In a machine that has to be accurate to the nanometre, metal already expands at a fraction of a degree: steel grows by about twelve micrometres per metre for every degree it warms. Predict that heat poorly and precision is lost. In batteries, heat exchangers and installations, heat decides how safe and efficient something is, and how long it lasts. Flow, materials and chemistry follow: how air and liquid move around and through a design, how a material holds up under force and temperature, and how molecules behave and react with one another. Each time with the same idea: a model that fits the problem exactly. ## How Physics is the teacher. We do not train on text from the internet, but on exact simulations we generate ourselves. That way we know the right answer for every example, and the model learns from physics itself rather than from what has been written about it. And the laws of nature are built into the model itself. Energy does not simply disappear, and heat flows from hot to cold; the model does not have to infer that from thousands of examples, it is in the way it is built. So it learns faster and makes fewer nonsensical mistakes. AI and simulation work together in this. The AI gives a first answer in milliseconds, and a classical solver finishes it exactly. Because the solver does not start from zero but from an answer that is already close, it is done much sooner. That gives the speed of AI with the reliability of a simulation: the answer is not estimated, but checked. And everything is measured. We test every step for accuracy, for speed and for how well the model works on designs it has never seen, because that is what counts in practice. We compare it with the classical simulation and with general language models, so that it shows what a specialist delivers. ## Europe Europe wants to make its own chips again. With the Chips Act, in force since September 2023, the EU aims to double its share of global chip production to twenty percent by 2030. In chips and in the machines that make them, heat is one of the hardest design limits: a chip that cannot shed its heat has to run slower, and a machine that expands by a fraction through heat misses its mark. Computing heat in seconds instead of hours means testing far more designs in the same time. Europe also pays more for energy. In September 2024, the Draghi report on European competitiveness showed that European companies pay two to three times as much for electricity as companies in the United States. Computing power costs more here, which argues for small models that do one thing well: a specialised model needs far less computing power than a large language model. It also helps design things that use less themselves, from batteries and heat pumps to whole installations. And Europe wants AI of its own. Specialised models are small enough to run in Europe, on European servers and under European rules, so designs need not leave the continent. That fits the European strategy for AI in science of October 2025, with RAISE as a European institute bringing together computing power, data and researchers. The Free Lunch Project is built in Europe, for European manufacturing and for everyone who computes with physics. Want to know more about the Free Lunch Project? [Get in touch.](https://goudbeek.com/en/contact)