Can you design super-smart materials with long-term memory that are capable of learning?
That was the question physics and mathematics bachelor’s student Daan de Bos set out to explore during his thesis research. Applying machine learning theories to materials presented several challenges, but his efforts led to a working theory that can now be tested on real materials in laboratories.
Self-learning materials must do two things. They need to perform calculations while simultaneously learning from those calculations. A material can be seen as a system with many parameters, like a mixing panel with numerous knobs, each of which can represent different values. This ‘information’ — or the configuration of all these knobs — must also be stored; otherwise, the material ‘forgets’ what it has learned and no longer exhibits the desired properties. Ideally, during the learning process, the material would autonomously adjust these parameters to optimise the desired characteristics.
How do you combine two conflicting processes?
When simulating a self-learning system with computer software, the two activities of calculation and learning can coexist easily. However, in materials, it quickly became evident that these two processes could interfere with one another.
‘We solved this by allowing the two processes to occur on different timescales. The processes that handle data processing are much faster than those that adjust the parameters,’ explains Daan.
Using mathematical techniques and computer simulations, he extensively analysed the model and discovered that this approach works. ‘As far as we know, this is the first example of a model using this principle,’ says Daan. He and his supervisor, Marc Serra Garcia, are currently writing a paper about their findings. Their theoretical model could potentially be applied to other materials to describe similar ‘learning’ behaviour.
Presenting to experts at the AMOLF Physics Institute
Halfway through his project, Daan was invited by Serra Garcia to give a presentation at AMOLF’s biweekly physics seminars. A thrilling experience for De Bos: ‘Luckily, I love presenting, so I was happy to do it.’
These seminars are usually reserved for PhD candidates and postdocs, but Serra Garcia had so much confidence in Daan’s research that he wanted to share it with his colleagues at AMOLF. ‘It was an incredible experience to stand before a group of scientists and explain my research. It’s very different from presenting to fellow students — you can’t just brush over details or make things up. It forces you to be thoroughly prepared and truly understand your topic,’ says Daan.
Wanting to calculate everything with physics
Daan has been fascinated by physics from a young age. ‘I still remember the first time I heard the word ‘physics’ and how I immediately knew I wanted to learn everything about it. That must have been during the last years of primary school. My father told me how he calculated the trajectories of tank shells during military service. I couldn’t wrap my head around the idea — how can you calculate that? How can you predict where a bullet would land? I just had to figure it out,’ says Daan enthusiastically.
That fascination with predicting the world around us has never left him.
A glimpse into the academic world
Throughout his bachelor’s project, Daan learned many things. ‘I learned how a research group functions, how to write a grant proposal, how to search and read scientific literature, how to give a scientific presentation, and how to write an article and design figures professionally,’ he shares. Additionally, it was his first time attending a scientific conference.
Daan has become very enthusiastic about a career in academia. ‘I will probably complete my master’s degree in Leiden. I’m already looking for an exciting project to begin with. After that, I’ll likely pursue a PhD. My ultimate goal is to make a living through research and work at a university.’