Arto Klami, Professor of Machine Learning, develops methods that help researchers choose their next experiment.
When baking sourdough bread, you may unexpectedly find yourself at the heart of the scientific process.
A sourdough starter grows in the fridge for years on end, and every starter behaves in its own way. Ordinary baking comes with ready-made recipes, but with sourdough a fixed recipe is not enough. A trial phase is needed, in which you test how the starter reacts. If the bread turns out rock-hard, it is worth adjusting the amount of flour next time. And so on.
This is how , Professor of Machine Learning, illustrates the scientific process, because the same cycle of experimentation repeats itself in the laboratory. An experiment is run, the result is observed, and the next attempt is based on what has been learned. This is exactly where Klami wants to bring a learning model to help: an AI that helps make sense of which experiments are worth running and which are better left undone.
Fewer experiments, more time freed up
The aim of a research-assisting AI model is to make the scientific process and companies' product development more efficient. Practical applications have already been found in drug development and in technology, for example in the manufacture of silicon wafers.
"A chemist usually knows roughly how a desired reaction should be made to work. There is a set of alternative solvents, and you start testing which combination produces the desired reaction. You draw up a passive experimental plan and try all the options one by one", Klami said.
Klami would like to know how this could be done more efficiently.
The benefit comes from reducing the number of experiments and thus saving time. When the model infers from the experiments already carried out what is worth trying next, unnecessary experiments can be skipped.
"If a hundred experiments were run in the old process and 75 are now enough, a quarter of the time and money is saved", Klami said.
The savings can be considerable. A single chemical experiment can take a chemist hours, and in some fields the cost of one experiment can run to tens of thousands of euros. In that case, cutting just a few experiments can save tens of thousands of euros.
Time and money are not the only benefits. As work becomes more efficient, researchers gain the chance to look into questions they previously could not afford to take on.
"I would not go so far as to say that AI would make scientific discoveries a human could not have made. The point is that we have not had the resources. As work is made more efficient, new kinds of things become possible", Klami said.
What happens to serendipity in science?
In 1928, one Alexander Fleming returned from his summer holiday to his laboratory at St Mary's Hospital. Over the summer, mould had grown in a Petri dish left uncovered. It was a mould called Penicillium, and the bacterial growth around it had been destroyed.
Many scientific breakthroughs, such as the discovery of penicillin, have happened by chance. Klami does not believe that systematizing research with the help of AI reduces the possibility of discoveries made by chance - quite the opposite.
"When a model is continuously trained to be better, an exceptional result does not go unnoticed. A good enough model brings a surprising observation to the surface, and a human gets to consider where it comes from", Klami said.
In the case of penicillin, it took a Scottish researcher who happened to look at his Petri dish at the right moment. AI could do the same by bringing into the attention when something seems exceptional.
The tendency of language models to produce hallucinations, fabricated answers, can also turn into an advantage.
"Proposing wild ideas is essential when innovating", Klami said.
AI is a tool
What will research work look like in 10 or 20 years? The core of research will not change: researchers still want to understand the world better. The tools, on the other hand, change, just as they have changed up to now - from typewriters to computers and computational methods.
Klami sees AI as one tool among others. AI agents can handle routine tasks, such as running comparison experiments and reporting results. This frees up the researcher's time for things that previously went undone.
Klami himself has moved across the boundaries between disciplines over the course of his career. The development of AI methods has always been at the core of his work, but already during his doctoral research he examined applications in systems biology. Since then, Klami has worked on areas including computational neuroscience, human behaviour and applications in physics.
The aim is not to focus on a single discipline. Klami hopes the methods he develops will benefit researchers regardless of their field, in both the natural sciences and the human sciences.
"It would be nice if I could help humanity in general in this way. If AI-assisted research helps find a cure for leukemia or solve problems related to climate change, it is hardly possible to pinpoint what my own contribution was. If I feel that I am doing something useful, that is enough for me", Klami said.
